WorksComputer Science
Revision-Bound Evidence Cuts: Authenticated Queries and Contextual Witness Capacity
Abstract
Authentic records can support unauthorized narratives when their qualifications, query scope, or counterevidence are silently dropped. We present revision-bound evidence cuts, a framework for authenticating both supporting and opposing inference obligations across changing source versions. Signed complete-update frontiers permit transport of unaffected certificates, alternative-cut recovery, and an exact certified-reuse radius given by a frontier-hypergraph transversal number. Their source obligations induce an integral model of context-bound authorization. To secure database-derived claims, we introduce an authenticated relational query contract that binds original source records, complete selected bags, counting units, parameters, and aggregates independently of execution strategy. A full-source verifier checks 16,824 Chogha Mish records across SQLite scans, indexes, and materialized views; three valid strategies agree and eight contract/view mutations are rejected. A separate 340-row NISP cohort shows that changing the admitted weighting rule can reverse a comparison while preserving every source record. A frozen English text-to-frame compiler produces qualified evidence receipts, with linguistic accuracy evaluated separately from cryptographic integrity. We also separate logical-source from transport commitments so lossless re-encoding need not invalidate unchanged scientific evidence. The SCENE prototype evaluates signed revisions, authenticated queries, scoped disclosures, and adversarial mutations at scales reaching one million source records. These guarantees rely on declared policies and independently pinned trust roots; they authenticate specified computations and obligations, not historical truth, complete counterevidence, or institutional endorsement.
Problem: authentic evidence in an unauthorized scene
An authenticated archaeological measurement can support a geometric approximation. It cannot by itself establish the truth of a caption, the intentions of a past society, or the source institution’s approval of a contemporary composition. Likewise, displaying several recognizable authorities together does not establish that their authority is interchangeable or jointly available without restriction. The engineering problem is to preserve the missing relations explicitly.
We distinguish three predicates: Attributed records the claimed origin of an evidence item. The predicate Supports records a qualified support relation under a declared context. Finally, Endorses records a signed, policy-limited authorization for a particular scene. Neither of the first two predicates entails the third. The signature authenticates an assertion and its issuer, conditional on trusted keys; it does not make the assertion historically true.
The motivating media hypothesis is contextual authority transfer: viewers may interpret the presence of an authentic source or familiar symbol as approval of a newly assembled narrative. Bourdieu supplies a conceptual account of recognition and symbolic power [12]; Hall’s encoding/decoding model allows negotiated and divergent reception [38]. Framing [27], source credibility [40], source monitoring [42], and repetition effects [28] motivate variables for future experiments. These sources do not establish that symbolic power is literally conserved, that media producers share a coordinated intention, or that repeated reuse must end in collapse. We replace those unsupported conclusions with a precise, falsifiable computational question: which authenticated obligations disappear when a presentation is reduced to a list of source names?
The contribution begins with the missing distinction between support, contradiction, uncertainty and authorization. Revision-bound polarity cuts retain both decision polarities and mandatory qualifications; a verified change frontier determines whether an old cut can be transported or replaced by an avoiding cut. A derived reuse threshold then connects the certificate family to a transversal problem. Its source footprints enter contextual witness capacity, which counts complete presentations under conjunctive obligations and consumable grants. The SCENE artifact implements the component mechanisms over original source bytes, recorded transformations, typed interpretations and explicit local issuers. The method is investigator-designed and inspectable, with no learned component.
The relational contribution is a reconstruction-aware authenticated query contract. It distinguishes complete selected bags from grouped answers, binds counting units and original-column mappings, and verifies equivalent scan, index and materialized-view executions against the same full-source scope verifier. Thus a genuine aggregate cannot answer a different parameter, unit or source-version contract merely because the displayed ratio agrees. The actual SQLite workload retains all 59 source fields and 16,824 records; source verification, query execution and materialization costs are measured separately.
A language-layer experiment compiles actual English source descriptions and catalogue-date phrases into explicitly qualified span frames. Frozen rules, source-group holdout, exact component scores and controlled text revisions measure extraction separately from receipt validity. External human-annotated historical corpora test transfer without changing those rules. The resulting receipts bind text, extraction policy and semantic slots; successful verification does not make a misparsed qualification linguistically correct.
Relation to earlier work. Our previous studies distinguish provenance from authorization in autobiographical models [4] and bind approval to published content and recovery state [3]. The present construction adds authenticated revision transport for polarity-complete inference certificates, a cut-frontier reuse threshold, and a link from surviving certificate alternatives to integral endorsement capacity. We derive these objects from established logic, provenance, authentication and combinatorial optimization; their exact scope and nearest technical precedents are given with the construction below.
Mathematical lineage and the point of failure
The relevant history comprises several questions that should not be collapsed into one mathematical tool. Hall’s 1935 representative theorem [37] addresses choosing one permissible representative per demand. Gale, Kuhn and Tucker’s 1951 treatment of linear-programming duality [31] supplies the price certificates used below. Ford and Fulkerson’s 1956 maximum-flow/minimum-cut result [29] turns suitable capacity constraints into integral allocations. Karp’s 1972 reductions [45] locate the obstruction when joint choices become combinatorial packing.
Ralph Merkle’s authenticated-tree construction, represented here by his 1980 paper [58], addresses efficient integrity authentication. It does not solve witness admissibility or scarce-resource allocation. Girard’s 1987 linear logic [34] distinguishes reusable truth from consumable premises; Bowers et al. implemented consumable credentials in authorization systems in 2007 [13]. Green, Karvounarakis and Tannen’s 2007 provenance semirings [36] preserve derivation structure and multiplicity. The failure below is therefore a lossy projection of provenance, not a limitation of every provenance algebra.
A source set records which names occur but loses occurrence counts and the distinction between prerequisites and alternatives. A provenance polynomial retains derivations and multiplicity; an admissibility policy must still specify which witnesses are permitted and which signed rights remain available. Hall conditions and network cuts solve allocation among singleton substitutes. Joint prerequisites instead require a conjunctive allocation model, whose fractional price bounds do not guarantee integral feasibility.
Authentication and resource state address different remaining questions. A committed root binds records against alteration; truth and admissibility depend on explicitly admitted evidence and rules. Serialized debits account for spent capacity, while freshness requires an accepted current state. Neither layer by itself determines audience interpretation. These distinct information requirements motivate the following occurrence-preserving construction.
The idempotence obstruction
Let be the support set of one displayed object. The support set of two copies is still . Suppose a positive additive exposure tariff factors through source sets as , taking values in a cancellative commutative monoid. Then
Thus a nonzero additive charge cannot be recovered from an idempotent union summary. This elementary argument identifies the required boundary: deduplicate a source inside one object’s evidence closure, but retain multiplicity across distinct rendered occurrences. A second loss is connective structure: the union cannot distinguish “both and ” from “either or $b.” An untyped source list loses both pieces of information before any cryptography is applied.
Evidence closures and context-admissible witnesses
Let and be finite nonempty sets of source principals and object types, and let be a finite directed acyclic evidence graph. Each object has a nonempty set of root evidence records. Each record names a source, a statement, a claim class, qualifications and its complete list of mandatory supporting records. For root records , let be the reflexive transitive support closure. In a fixed reconstruction,
where counts rendered occurrences of type . A source contributes once to an occurrence even if several support paths reach it; a second displayed occurrence contributes another unit. A unit is a protocol-defined endorsement authorization, not a viewer, impression, copyright permission or measure of prestige.
Definition 3.1 (Admissible witness family). A context specifies a purpose, audience and explicit evidence-admissibility policy. For each type , the finite nonempty family contains acceptable witness bundles. Every source in a selected bundle is required (AND); selecting a member of the family is a choice among acceptable realizations (OR).
For a fixed manifest, . Alternatives require separately declared and reviewed realizations. An optimizer does not infer that two historical interpretations are equivalent. Budget is the remaining capacity of a selected grant for source ; additional grant issuance is outside that bound. Every possible rendition must retain its chosen witness and obtain fresh authorization for its exact resulting root. Here a rendition means one new authorization event; rereading an already accepted artifact is not another rendition or debit. Actual display frequency is outside this accounting model.
Compatibility assumption. Witness choices for different object occurrences are independently compatible within a rendition. This excludes, unless explicitly encoded, constraints such as “all objects must follow the same reconstruction school.” Such applications need complete-scene configuration variables or additional coupling constraints. This assumption is necessary for the counting equivalence in the next section.
Deriving contextual witness capacity
Definition 4.1 (Contextual witness capacity). is the largest integer number of complete scene renditions whose occurrences can be assigned admissible witnesses without exceeding .
Put . Counting how many occurrences use each witness gives the configuration integer program
We include zero in . An actual collection of renditions yields these counts. Conversely, integral counts provide occurrences of each type; partitioning each type into groups of size gives complete renditions under the compatibility assumption. Hence the formulation is exact, rather than merely a necessary aggregate budget condition.
The provenance-polynomial form
For formal variables , define
A bundle is a squarefree monomial: mandatory sources multiply. Alternative bundles add. Raising to the number of occurrences retains repeated exposure in exponents. Polynomial provenance is established machinery [36]; its application here yields the equivalent capacity expression
Here is the set of exponent vectors of nonzero monomials, and , so is always feasible. The coefficients count derivations producing an exponent vector; feasibility depends on exponent support. For one fixed witness per type, the entire problem reduces to
This is a capacity of explicitly limited rights, not a predicted cultural life span.
Where matching remains exact
Assume every witness is a singleton. For nonempty let and . The capacitated form of Hall’s condition gives
Necessity follows because all occurrences must use . For sufficiency, a network sends capacity from a super-source to each type, then to its admissible source nodes, then through capacities to the sink. The displayed conditions ensure every cut can cover total demand; integral maximum flow supplies witness counts [29, 37]. This is a specialization of known matching theory, not a new Hall theorem.
Deriving the source-price equation
Relax and to nonnegative reals and denote the optimum by . For a nonnegative price vector , define
Every feasible allocation obeys
The inequality produces a checkable bottleneck certificate. It becomes an exact fractional equation by duality:
Theorem 4.2 (Contextual source-price representation). For the finite nonempty witness families above,
Proof. Write the relaxed primal as max subject to , , , where identifies each variable’s object type. Its dual is
For fixed , object potentials may be chosen as their largest admissible values . Feasibility becomes . Every occurrence uses at least one source, so the zero-feasible primal is bounded by finite total capacity. Strong linear-programming duality [31] therefore applies. Homogeneity normalizes to one and gives the ratio. Equation (9) and integrality give the last inequality.
Zero source budgets are permitted, including , for which both capacities are zero. The normalization never divides by a budget: a strictly positive price on every source gives because objects and witness bundles are nonempty. The source prices identify scarcity in the declared policy. They do not score historical truth. A rational provides an exact upper bound even if a numerical optimizer found it; optimality additionally needs a matching primal certificate or an independent argument.
The AND bottleneck: complexity and integrality
The singleton flow formula cannot simply be reused for joint prerequisites.
Proposition 4.3 (Width-two feasibility obstruction). Given explicitly listed witness families, deciding is NP-complete even with unit source budgets, unit object multiplicities and witnesses of size exactly two.
Proof. Take a perfect three-dimensional matching instance with nonempty equal-sized, pairwise disjoint sets . If some belongs to no triple, map the instance to the fixed valid NO instance consisting of two object types, both requiring the sole witness with . Otherwise, for each create one object type; a triple supplies its admissible witness . Give every source in the disjoint union capacity one. Choosing one nonconflicting witness for every is exactly a perfect three-dimensional matching. This is a polynomial reduction from the classical problem [45]. A selected witness per object is a polynomial certificate, establishing membership in NP. Empty three-dimensional matching inputs can separately be mapped to one object requiring with unit budgets, a fixed YES instance. Thus every output respects the model’s nonempty witness-family requirement.
This classification is a consequence of known combinatorial hardness. It is not an average-case cryptographic hardness assumption, and we do not use it to derive encryption keys or a trapdoor. For a concrete integrality failure, use seven unit-budget sources and one type whose witness triples are
Each pair of triples intersects; each source occurs three times. Thus . Assigning to each triple is fractionally feasible. Summing source constraints gives , so . Consequently . This explicitly constructed Fano-plane incidence example separates integral authorization from its fractional certificate. A verifier that accepts a fractional plan as a complete rendition would be unsound.
A derived measure: obligation-erasure amplification
Consider the particular defective operation
It replaces every conjunction by singleton alternatives. The ratio measures excess capacity caused by this specific loss of obligation structure. At zero true capacity we report infeasibility explicitly rather than divide by zero. The notation means the same instance with and all other inputs unchanged. Any true feasible plan remains feasible after replacing each bundle by one of its members, so .
Theorem 4.4 (Unbounded erasure amplification at width two). For every integer , there is a unit-budget instance with maximum witness size two, and .
Proof. Use sources , each with budget one, and one object type of multiplicity one. Let
Every true rendition consumes the shared prerequisite , so at most one is possible and any pair realizes one. After erasure each of the singleton sources is an alternative and all can be used once. Thus .
There is no bound on amplification depending only on witness width. The lost shared prerequisite, rather than a broken hash, causes the apparent increase. In (12), the three capacities are 1 (integral), (fractional) and 7 (erased); they answer different questions.
Corollary 4.5 (Source-list insufficiency). A summary retaining only source unions, multiplicities and budgets cannot give a sound constant-factor capacity approximation for all witness policies: no fixed finite and estimate based only on that summary can satisfy for every instance.
Proof. The star policy and its erased policy have identical such summaries and capacities 1 and . A common estimate sound for the former is at most one; an approximation within factor for the latter must be at least . Choose .
Context restrictions. If with unchanged budgets and multiplicities, then and by feasible-set inclusion. Merely changing a context string does not change these quantities unless admissibility changes. The cryptographic layer separately prevents transferring a permit to that new string. No general monotonicity of the ratio follows.
Revision-bound polarity cuts
The witness families in § 3 specify what a rendition must retain. We now derive such families from executable inference obligations and authenticated evidence revisions. The problem is finer than preserving a source hash: a qualification or registered counterexample can change a presentation’s eligibility even when every underlying source file remains unchanged.
Figure 1. Constructed mathematical experiments. (a) Keeping a shared prerequisite fixes capacity at one; replacing AND by OR gives . (b) The Fano fixture distinguishes an integral plan, a fractional certificate and erased obligations. These are not measurements of cultural influence.
Typed evidence states and closed-catalogue guards
Fix a finite nonempty leaf registry , a scope and an acyclic policy DAG . The scope identifies the intended question, purpose and audience. Each leaf has a complete canonical record
The source hash identifies the underlying scientific record; the qualifier and two bits record its admission for this particular inference. The four states are unknown , supported-only , opposed-only and contested . They describe admitted evidence, without assigning probabilities or asserting objective truth. Four-valued evidence semantics are established [10]. In particular, an unknown premise does not become supported by negation, and inconsistency does not entail an unrelated proposition.
Each gate authenticates its ID, operation, complete ordered argument list, qualifier and type, either interpretation or eligibility. There is no authorization gate or inference rule producing issuer consent. For interpretation gates we use the usual truth-order operations
Logical disjunction must be distinguished from a choice among individually admissible witnesses. For example, , although neither input is supported-only. If an application requires each chosen alternative to be individually uncontested, that requirement belongs in its reviewed policy; it does not follow from the FDE disjunction in (15).
An eligibility gate can additionally use an explicit guard , , whose registered defeaters are part of :
Here means no admitted support for this defeater in the committed catalogue. It is not a proof that the defeater is false in the world. A contested defeater still has support and blocks eligibility. The guard is a declared nonmonotone admission rule; it is not identified with FDE negation. An eligibility decision passes only at . The completeness of the catalogue, scientific suitability of its rules and reliability of admission remain explicit assumptions.
Obligation-preserving local cuts
For either bit of a gate, expand its defining equation into an AND or OR of signed child-bit literals; NOT changes the channel, while the guard introduces the literals . A local certificate applies the following rules:
| output 1 | output 0 | |
| AND | all literals equal 1 | one literal equals 0 |
| OR | one literal equals 1 | all literals equal 0 |
Table 1.
Both target bits must be proved. A decision cut contains the resulting reachable proof DAG, complete records for its opened leaves and gates, and their authenticated openings. Its frontier is the set of opened full leaf records. The verifier rejects duplicate or extraneous steps, cycles, omitted required children, altered argument lists, missing qualifiers and unsupported status/type changes. A leaf opened for one bit nevertheless discloses its whole record. This is selective disclosure of unused records, not zero knowledge about the disclosed frontier.
Lemma 5.1 (Local cut soundness). Fix and the records on . If is valid for target with status , then every assignment of records to the other leaves, using the same registry and well-formed bit domain, evaluates to under .
Proof. An opened leaf fixes each asserted bit. Induct through the reachable proof steps in topological order. All-one premises determine a true conjunction; a single zero determines a false one. Dually, one true premise determines a true disjunction, while all-zero premises determine a false one. Literal polarity and channel changes are fixed by the authenticated gate. Applying this induction to both root channels proves , independently of unopened records.
This is soundness of a particular local calculus, not completeness for every semantic certificate of a reconvergent circuit. Algebraic identities may admit smaller certificates in another proof system. Our avoiding-cut procedure is complete for the rules (17): at a forbidden leaf it returns unavailable; an all-required rule requires all child certificates; a decisive rule requires one available child with the decisive value. Memoization over node/channel pairs decides feasibility and selects the first feasible branch. It does not globally minimize shared source footprint, bytes or cost. These distinctions are consistent with the established tradeoffs in knowledge compilation and certificate complexity [16, 23, 61].
Signed revisions and complete update cuts
The reference implementation commits the ordered leaf vector using a cached perfect binary tree padded to the next power of two. A record leaf hashes as , an internal node as , and padding uses a separate domain and its eight-byte position. This fixed perfect-tree encoding differs from the RFC 9162 tree in the later scene commitments. Gate records have a separately committed tree. The policy descriptor commits the registry digest, leaf and gate counts, gate-tree root and scope hash. A checkpoint signs
with included in the signed body. Here denotes the canonical named-field object, not an alternative array serialization, , is the leaf-tree root, the registry digest, the policy digest, an epoch and a nonce. The Ed25519 signature covers the canonical body after the distinct SCENE/revision-checkpoint-signature/1 domain and a NUL separator. The consumer pins the public key, current , and outside the packet, compares the packet target with its requested , and requires the intended decision type. A correctly signed alternative checkpoint does not replace these pins. A trusted current root or external history mechanism supplies freshness and fork resolution; signatures and increasing sequence numbers alone supply neither.
For consecutive admitted snapshots, define the full-record change set
A complete update cut partitions the entire padded tree into disjoint shared subtrees and opened changed leaves. A shared subtree supplies one hash used in both reconstructions. A changed leaf supplies its position and complete old/new records, with equal IDs and unequal records. The verifier reconstructs both complete roots, checks that every position is covered exactly once and that no supplied node is unused, and verifies both checkpoints. It requires the same registry, policy and scope; sequence ; parent ; nondecreasing epoch; and a different nonce. Padding cannot change. A change to leaf identity, policy or scope requires a new policy outside this interface.
Lemma 5.2 (Complete-delta soundness). Under authentic admitted checkpoints and collision-resistant hashing, an accepted update cut returns exactly .
Proof. Coverage and old/new root reconstruction authenticate the two complete vectors. Every explicit changed record is unequal. If an actual unequal record were omitted, it would lie in a claimed shared subtree with equal old/new hash. The two distinct canonical subtree contents would then yield a hash collision. The same binding excludes inventing an unequal record outside the admitted vectors. Coverage rejects overlap and unexplained gaps.
The initial producer stores leaf-tree hashes. A persistent revision rehashes the union of affected paths, bounded by for changed records, and stores sparse overlays. Cached gate aggregates and reverse edges update affected gates without repeatedly scanning every argument of a large fanin gate. Enumerating the update cut skips equal cached subtrees. Initial ingestion, affected-gate work, proof output and lookup overhead across long persistent histories remain real costs; the path bound is not a bound on all producer work.
Certificate transport and its security game
Theorem 5.3 (Cut-stable revision transport). Let be a valid cut at and let a verified complete update cut connect admitted checkpoints with fixed registry, policy and scope. If
the same local inference steps certify the same target pair at . Its leaf authentication paths can be repaired using the update cut, yielding an ordinary independently verifiable decision cut at the new root.
Proof. Complete-delta soundness and (19) imply equality of every opened full leaf record. Gate records are unchanged because policy is fixed, so local cut soundness preserves both target bits. For each opening, a sibling subtree exposed in the update reconstruction supplies its new hash; a sibling inside a shared subtree retains its old hash. Replacing the former entries reconstructs the new leaf-tree root and checkpoint wrapper. The ordinary new-root cut verifier checks the repaired result.
Disjointness is necessary for reusing the same full-record frontier by this rule, not for semantic output stability. A qualifier-only revision belongs to even if both bits and the source-file hash are unchanged. A touched certificate may also have an unaffected alternative: generating an old cut with and then transporting it recovers that case. If no such cut exists, current reevaluation can still find a new certificate. The batch API verifies one complete delta once, then validates and repairs each old cut; its verified map is local to that invocation. This amortizes delta checking without making individual proof checking free.
For precision, consider an adversary that receives admitted canonical states, their signed checkpoints and legitimate openings, and outputs a decision cut, complete delta or transported cut under externally pinned roots, policy, scope and query. It wins if acceptance certifies a different evaluated pair, an incorrect changed-record set, or an unchanged frontier containing an unequal record. The admission authority is honest about the fixed registry/DAG schema; source truth, policy adequacy and a compromised signing key are outside this game. With strict schema checking and injective canonical descriptors, there are polynomial-overhead reductions satisfying
Here the multi-user term accommodates the admitted key set; for one checkpoint key the ordinary EUF-CMA game suffices. An unissued accepted checkpoint gives a signature forgery. Otherwise, a different committed record, gate, descriptor or hidden change gives a collision by the tree-binding argument. With neither event, the two preceding lemmas and transport induction rule out winning. This is a conditional authenticity reduction for a construction from standard primitives, not a new hardness assumption or a bound estimated by rejection counts. Public benchmark seeds instantiate functional tests only.
An exact certified-reuse threshold
Fix an admitted old snapshot and a nonempty family of all valid local cuts for its exact target pair. Its frontier hypergraph is . Our syntax has no constants or empty-argument gates, so all frontiers are nonempty. Define
This is the classical transversal number applied to the concrete old-certificate family. Boolean certificate complexity supplies the underlying mathematical language [16, 61]; our functional measures authenticated old-cut reuse rather than the number of queries needed to evaluate a function.
Theorem 5.4 (Certified old-cut reuse radius). The largest integer for which every changed-record set of size at most leaves some transportable old cut is exactly
Proof. A set of fewer than records is not a transversal, so it misses a frontier and the transport theorem applies. A minimum transversal of size intersects every old frontier, ending that universal guarantee. Every record set can be realized as a well-formed qualifier-only update, so the argument does not require that those updates flip the target’s bits.
For positive record costs , replace by to obtain threshold . Every update with cost strictly less than this threshold leaves an old cut, and a minimum cost transversal attains the obstruction. Only for integer costs is the corresponding integer budget radius . An empty eligible-cut family has no zero-update guarantee; a proof system admitting empty-frontier constant certificates would instead need an infinite-threshold convention. Neither case is silently assigned an ordinary finite radius.
Proposition 5.5 (Guard-erasure inflation). For each , deleting one registered guard can inflate the certified reuse threshold by a factor , while all source hashes and cryptographic primitives remain fixed.
Proof. Let be supported-only leaves and . Its inclusion-minimal local frontiers are : one branch proves support and zero opposition of the OR. Every valid frontier contains such a singleton, so . Add an inactive catalogue defeater and define . Proving requires and a positive branch; proving also requires . The minimal frontiers are therefore and hits all of them, giving . The thresholds have ratio ; the radii are and zero, for which a ratio would be undefined.
The mechanism parallels obligation-erasure amplification in § 4.5: dropping a shared prerequisite before authentication changes the resource being certified. In a guard star with all defeaters inactive, every positive eligibility cut must expose and every . Omitting any leaves two completions with identical disclosed records but different eligibility after activating that defeater. Thus records are necessary, and the local cut attains this bound. Authentication of alone cannot replace these negative obligations.
From surviving cuts to witness capacity
For a supported-only target , let collect the source principals of its opened leaves, deduplicating within that occurrence. Define
The family is generated by positive, opposing and qualifier obligations in the actual policy, rather than by an unauthenticated source union. It can be exponentially large. An empty family for any required target means zero complete-rendition reuse capacity; this case is handled before invoking the nonempty-family planner of § 4.
Theorem 5.6 (Transport-certified source capacity). Under the independent-compatibility assumption of § 3, the maximum number of complete renditions certifiable by unaffected old cuts with source-obligation capacities equals the configuration optimum (3) with families (23) and the specified target multiplicities.
Proof. Group each actual target occurrence by its transported cut’s source footprint. The resulting integer counts satisfy the configuration equalities and source capacities. Conversely, each selected footprint in an integral feasible plan has a corresponding old cut by definition. Transport supplies a current cut; the compatibility assumption permits grouping occurrences into complete renditions. Deduplicating within an occurrence and separate charging of repeated occurrences give exactly the coefficients in (3).
For fixed old state, only removes witnesses, hence this certified-reuse capacity is nonincreasing. Recomputing a changed state’s inferences can instead increase or decrease its current witness family. This theorem concerns a plan under available source capacities: it does not transfer a previously issued permit. A transported interpretation or eligibility cut is never issuer consent. The later protocol still requires every necessary issuer’s exact-scene permit and atomic budget admission. Checkpoint signatures use a different message domain from grants and permits; promotion is rejected even when a laboratory fixture uses the same key for two roles.
Relation to established methods. Neither negative provenance nor tracking updated data is new: -relations with difference, logical provenance and ProvSQL already cover substantial parts of that problem [32, 35, 73]. Authenticated answer completeness [63], persistent authenticated dictionaries [5], dynamic hash accumulators and their proof updates [25], and incrementally verifiable computation such as Nova [48] establish that state authentication, incremental verification and proof reuse have broad prior art. Revocable and use-once proof-carrying authorization also predates this work [59]. Our proposed contribution is the specific composition of typed polarity/guard obligations, complete update cuts, repairable local-cut transport and cut-generated source allocation, together with its certified-reuse functional and guard-erasure analysis. It supplies inspectable conditional claims about admitted scientific records. It cannot turn the authenticity of a record into truth of an interpretation, completeness of all possible counterevidence, or authorization by the record’s source.
SCENE: authenticating the counted object
The mathematical planner is not an authorization oracle. SCENE binds a chosen realization to an exact structured manifest, then atomically consumes its declared source obligations. It combines established components. Macaroons already support contextual caveats [11]; sanitizable signatures [8] and policy-based redaction [46] address authorized editing and extraction. C2PA authenticates content-associated assertions and ingredients and leaves trust interpretation to consumers [22]. SCENE neither implements those standards nor demonstrates a vulnerability in them. It specializes provenance and consumable credentials to the capacity semantics above.
Authenticated manifest and disclosure closure
A manifest contains a header (schema, renderer, cost rule and context), source snapshot records, evidence records and explicitly identified object occurrences. The reference scene supports square-pyramid geometry with integer millimetres, positions, labels and colors. Source records hash packaged numeric source cards; their URLs cite the publications, and separate metadata records the inspected publication copies. These local hashes must not be mistaken for publisher signatures.
Canonical encoding accepts NFC strings, ASCII dictionary keys, safe integers, booleans, null, lists and dictionaries. Floats, duplicate JSON keys, unpaired surrogates and integers outside the exact JavaScript range are rejected; booleans are rejected in integer fields. Enc emits compact UTF-8 JSON with ASCII keys in lexicographic order, no insignificant whitespace, and integers in . Strings must already be NFC; the encoder rejects rather than silently normalizes them. Object, source and evidence records are ordered by stable IDs. The implementation’s restricted encoding is documented as such, not advertised as general JCS compliance [70]. Rendering resolves ties by depth, object ID and face index, so record-list permutations cannot change the committed geometry’s deterministic rendering order.
Let be that ordered record list. With RFC 9162 leaf/internal domains and tree shape [50], set
Here the domain string is ASCII, 0x00 denotes one NUL byte, and hex is lowercase hexadecimal without a prefix. Thus both the inner tree root and the returned scene root are serialized exactly as in the code. Signing the count is necessary here: an RFC-shaped proof of leaf zero in a three-leaf tree can verify under a claimed four-leaf size if an opaque sibling digest is reinterpreted as a larger subtree. No collision is needed for that proof-shape ambiguity. Our size-bound wrapper rejects it.
A disclosure for selected object IDs opens the header, each selected object, its full transitive evidence closure, and every source record reached. The verifier checks all inclusion proofs against the pinned wrapped root and count, the exact external selection, record uniqueness and closure. Membership does not imply that hidden real-world evidence is absent. Completeness is relative to the authenticated adjacency lists and policy, which a trusted reviewer must have admitted. A disclosure is not a new endorsed composition: producing a strictly smaller projected manifest changes its root, except with negligible collision probability, and requires its own authorization. Opening the entire manifest or merely permuting its record lists does not create a distinct committed scene.
Grant, permit and atomic state transition
The lab uses maintained Ed25519 [43] and SHA-256. Pinned source keys live outside untrusted packets. A source grant signs its ID, source ID, ledger ID, evidence-subject digest, purpose, audience, renderer, cost rule, budget and validity interval. The subject digest includes the complete source record and that source’s evidence snapshot, including support edges and qualifications. For an instantiated scene the same source signs a permit containing grant and ledger IDs, , context hash, recomputed cost and a request nonce.
The single authoritative SQLite ledger checks all required source permits in a BEGIN IMMEDIATE transaction. It validates pinned keys, grant signatures, scope, evidence snapshots, time, costs, nonces and balances. It debits all sources and records all spent nonces, or rolls back every change. For each grant the enforced invariant is
This is stipulated authorization accounting, with established consumable-credential precedents [13]. Copying evidence or citing public heritage data is not forbidden; the policy governs claims of the issuer’s endorsement.
Theorem 6.1 (Conditional scene and resource integrity). Assume collision resistance of , unforgeability of the signature scheme, the specified injective encoding of record lists, honest admitted source/permit issuance, externally pinned issuer keys, a trusted clock, correct schema/closure/cost enforcement, and one intact serialized ledger. For selective disclosure, also assume that the externally supplied expected root was produced from a valid admitted manifest. In a committed transaction, the ledger recomputes the supplied scene’s root and context digest and compares them with each signed permit; it does not trust a root claimed by the scene submitter. Then a permit issued for one committed scene or context cannot authorize a distinct one except by signature forgery or hash collision. Accepted transactions satisfy (25), including object cloning costs.
Proof. If an accepting permit body was not signed under a required pinned key, it is a forgery. Otherwise the issued permit fixes the root, context digest, grant and cost. Acceptance of a different committed record list under that root yields a collision in the wrapper or domain-separated tree; acceptance of a different encoded context under its digest yields a collision. Different source evidence snapshots are similarly excluded by their grant-subject digests. The resource claim follows by induction over committed transactions: initial balances equal signed budgets; each atomic step checks nonnegative residual capacity, records previously unused nonces, and subtracts the recomputed vector. Object duplication adds occurrence terms in (2). A failing source causes no state transition, so interleavings cannot create a partial multi-source debit.
This is a standard conditional reduction and invariant proof; it does not establish the safety of arbitrary renderer implementations, input policies or maliciously issued endorsements.
State and display limits. Ledger events form a domain-separated hash chain. Receipts are not ledger-signed checkpoints: public verification of committed spending requires an independently trusted pinned head or trusted ledger access. Auditing replays signed grants and permits and checks balances, nonces and the chain; recomputing scene costs also needs archived manifests. Database integrity and a single authoritative state are explicit assumptions; hashes alone provide no consensus, freshness or split-view prevention. Two disconnected offline verifiers starting with the same unspent state can each accept a first spend, so universal offline prevention is absent. The browser workbench marks edits as drafts and never claims to verify endorsement. Camera controls are exploratory projections of committed geometry; exact pixels, user attention and display frequency are not authenticated.
From symbolic envelopes to actual digital evidence
The empirical inputs include original museum photographs, an externally acquired photogrammetric surface, earlier Anatolian source models and tables, and archaeological-genetic supplementary records. Simple monument envelopes remain only a small, fully transparent regression fixture. No image or surface in the corpus is AI-generated, and no learned depth estimator is used.
Corpus, archaeological reuse and source qualifications
The image corpus contains 48 distinct museum objects: 33 from Cleveland and 15 from the Metropolitan Museum, with 48 different original-image hashes. Museum catalogue attribution groups 32 as Egyptian and 16 as Sudan/Nubia. The selection covers 11 investigator-defined material/form domains, including inscribed stelae, reliefs, stone sculpture and tools, wooden figures and coffins, pottery, glass, faience, textile, beadwork and metalwork. These are heterogeneous source objects, not 48 independent excavations or a representative cultural sample. Every bundled web JPEG has explicit CC0/public-domain support in its museum record under the museums’ open-access programs [20, 82]; the 48 original files total 11,151,616 bytes. Separate metadata snapshots retain accession, medium, date, provenance, dimensions, bibliography, rights and exact file hashes.
Twelve Sudan/Nubia objects have Faras as their catalogued site. Four others retain qualified or broad attribution: the Napatan pendant’s findspot is unspecified, the Taharqo shabti is probably from Nuri, a ram’s-head amulet is possibly Nubian, and the kneeling king has a regional Egypt/Sudan attribution. No cultural label is promoted into a certain excavation location.
Two actual objects provide the relevant historical depth. Cleveland’s pendant 1987.1 is described as a lion-head component, probably Egyptian and most likely an earlier gaming piece, adapted as a Napatan amulet [21]. Its page and metadata retain the museum’s differing chronological descriptions rather than resolve them by guesswork. The Met’s kneeling Kushite king, 2002.8, is catalogued with later reduction of the double uraei and erasure of neck amulets and the royal name [81]. These attributed interpretations document material reuse and selective alteration. They motivate source, transformation and qualification records; they do not establish the motives of every ancient viewer or a universal causal law connecting ancient and modern media.
Actual photogrammetry and computed surface depth
The SEAMS dataset for World Museum Liverpool 56.22.225, a wrapped votive cat mummy from Thebes, provides the original OBJ, material, texture and README under CC BY 4.0 [68]. Our parser finds 133,414 vertices and 266,844 triangles. The source reports 181 aligned cameras. The actual four-file deposit does not include the raw camera images or project despite its generic README listing them, so we reproduce downstream analysis, not original SfM. The source units are reported as metres, but focal-length descriptions disagree (50 versus 55 mm), the README reports GCP RMSE 0.185771 m with no check points, and we do not independently validate physical scale.
The original coordinates and connectivity are retained. An edge-incidence analysis finds no boundary or nonmanifold edges and five connected components; these are computational properties of this mesh, not evidence that the physical object has five independent parts. PCA supplies a recorded exploratory frame. For each of its three axes and both viewing directions we project actual vertices into a orthographic grid and retain the nearest sampled depth, with a square splat and an explicit occupancy mask. Let be the rounded projected pixel coordinate of vertex , and let be the signed unit viewing axis. In symbols,
The depth expression applies only when ; otherwise the stored depth is NaN. The global-minus-local maximum gives the nearest sampled distance from the view-relative origin in either viewing direction. Empty pixels include both background and sampling gaps; they are not automatically scan holes. These are computed views of a reconstructed surface, not sensor depth, dense triangle rasterization, CT, internal anatomy or calibrated metric measurements.
A deterministic 2,000-vertex sample supports covariance analysis over 32-point neighborhoods, including the query vertex. For covariance eigenvalues , we report the dimensionless local roughness proxy
Here is the normalized Laplacian of a symmetrized eight-nearest-neighbor sample graph with Gaussian weights; its first 32 computed eigenvalues are archived. The roughness, unsigned neighboring-normal variation and spectrum depend on vertex density, sampling, bandwidth and reconstruction. They are not calibrated curvature, dating evidence or cryptographically unique shape identifiers.
Figure 2. Actual source objects and computed surface views. Upper row: Napatan pendant, kneeling Kushite king, Egyptian stela and Faras jug (Cleveland and Met, CC0; catalogue IDs identify full source records). Lower row: two orthographic point-depth projections of SEAMS 56.22.225 (Pubblico / National Museums Liverpool, CC BY 4.0), derived with new cameras and masked gaps. These are surface-model projections, not measured depth photographs or validated physical scale.
Image analysis without invented depth
Each photograph is analyzed at long-side limits 128, 256 and 512 pixels using grayscale entropy, Sobel-gradient summaries, smoothed structure-tensor coherence and high-frequency Fourier energy. All 144 rows retain parent object IDs and analysis scales. They describe image texture and edge structure; an RGB image alone supplies no metric depth in this method. As a concrete information-loss control, reversing the decoded pixel order preserves every channel histogram for all 48 objects while changing all 48 exact RGB hashes. This demonstrates why visual descriptors cannot replace the authenticated spatial representation. The observation is algebraic and corpus-checked, not a trained authenticity classifier.
Cryptographic region extraction over real pixels and geometry
The fixed-scene protocol is extended by two typed evidence layers. Whole-file hashes identify original inputs; structured commitments verify the actual disclosed content, its selection coordinates and mandatory qualifications. No institutional signature is fabricated: every experiment uses an explicitly public local test key. Its private seed is also published for reproducibility: these fixtures test implementation, not unforgeability against a reader who knows that seed. Real issuance requires fresh protected private keys.
Image tiles, scale and off-crop context
The image manifest signs original-file and decoded-RGB hashes, dimensions, decoder and resampling versions, exact rational scale maps, source identity, scene/context bindings, calibration status, evidence graph and a count-bound tile commitment. The reference pipeline uses EXIF orientation, RGB conversion without an ICC color transform, and sequential BOX half-sampling with signed level dimensions. Tiles are except at boundaries. A leaf binds its level, row, column, bounding box, source and RGB-buffer digest.
An independently fixed policy specifies a half-open crop, level, admitted manifest, required evidence and mandatory tile IDs. Supplemental tiles may lie outside the visible crop: for example, context needed to interpret a local detail can remain mandatory. A proof supplies actual RGB tile buffers and paths. The verifier checks the signature, expected policy, count-bound tree, coordinates, pixel hashes and complete evidence closure, reconstructs the crop from verified tiles and requires equality to the supplied crop pixels. An optional original-file check also verifies the encoded file, decoded base image and actual derived levels. Rebuilding expected policy from the attacker’s packet is forbidden.
The scene first commits original asset hashes, then its final root is bound by the signed asset manifest. Committing that whole asset manifest back into the same scene would create a dependency cycle and is not the construction. Byte integrity, decoded-pixel equivalence and perceptual similarity are separate policies; this prototype enforces exact equality within the declared pipeline.
Triangle blocks and vertex dependency closure
The mesh layer commits little-endian float64 vertex buffers and uint32 triangle buffers in blocks of 1,024 vertices and 2,048 triangles. A signed, count-bound tree authenticates 131 vertex blocks, 131 triangle blocks and one metadata header. That header binds the original OBJ hash, units, calibration caveats, analysis frame and context. A disclosure opens exactly the independently selected triangle blocks plus all vertex blocks referenced by those triangles. Missing or extraneous blocks, changed coordinates, rewired indices and altered metadata fail. This is reference closure, not a claim that all geometrically adjacent faces or all historical evidence have been disclosed.
Proposition 8.1 (Conditional regional extraction integrity). Fix the trusted issuer key, expected admitted root/manifest, selection and qualification policy before receiving a packet. Under signature unforgeability, hash collision resistance, injective descriptors and correct deterministic decoding, a verifier cannot accept changed disclosed pixels or mesh buffers, different selection coordinates, or omitted mandatory dependencies as that authorized extraction except by forgery or hash collision.
Proof. An accepted signed body outside the honest issuer’s issuance record gives a forgery. An honestly issued alternative with a different expected manifest digest or root fails the fixed policy unless a hash collision occurs; an identical canonical binding is the same admitted commitment. Under that commitment, changing an opened descriptor without changing its root yields a tree/wrapper collision. Changing buffer bytes while retaining the authenticated digest yields a buffer-hash collision. Exact selection comparison and deterministic grid/block indexing exclude positional or scale relabeling. Finally, required tile IDs or vertex-block IDs are recomputed from the fixed selection and authenticated dependency data; the verifier demands set equality, so omission is rejected rather than inferred from absent unauthenticated records. The image verifier additionally reconstructs the crop from checked buffers and compares the supplied pixel output. The mesh checker authenticates disclosed buffers; extending that guarantee to a visible surface requires a renderer to consume those checked buffers.
These proofs concern exact evidence extraction. They provide no confidentiality, capture authenticity, robustness to benign edits, historical truth, institutional consent or stateful spending guarantee on their own. Spending remains the separate serialized authorization layer; an archived membership proof is reusable.
Real-corpus authorization and attack workload
Across the 48 image objects, three resolution levels produce 9,103 authenticated tiles. Four region selections per level give 576 actual-image controls, all accepted with original-file checking. Nine mutation families generate 5,184 unauthorized packets, all rejected: crop and tile pixel edits, coordinate and source relabeling, crop-policy transplantation, evidence/qualifier removal, mandatory tile omission, context editing and wrong-key checks. Their case-level records identify the parent object, crop, level, proof size and outcome; they are constructed tests of exact rules, not independent samples from real attackers. Additional valid-proof transplantation experiments isolate selection binding from mere rejection of malformed signatures. Across three real region selections per image, 144 original controls and 144 honestly signed alternate-context controls are accepted under their own policies. Under the original fixed policy and key, 864 substitutions are rejected: valid other-asset or other-level proofs, the wrong actual source file, foreign-key signatures, honestly signed different-scene claims, and histogram-preserving swaps of distinct image patches. All 144 patch pairs differ spatially while preserving their joint RGB histogram. These cases specifically show that a valid signature or preserved photometric summary does not authorize a different extraction.
On the real mesh, 16 actual triangle-block selections are accepted; 160 buffer, topology, dependency, context and calibration mutations plus three external selection/metadata transplants are rejected. Block closure reveals a practical cost: original indexing scatters selected triangles’ vertices across many blocks. This motivates the explicit layout experiment below.
The capacity equations also apply to the 49-object image/SEAMS corpus without pretending that distinct artifacts are historical substitutes. Each occurrence requires both a snapshot-admission role and an interpretation-review role. One declared review round uses counts (33, 15, 1, 49) for Cleveland-snapshot, Met-snapshot, SEAMS-snapshot and interpretation review. All are local study resource counters, not museum-issued endorsement rights. Twelve budget cases agree with the fixed-witness capacity equation and integer optimization. At one round per role, keeping the conjunction gives capacity one; erasing it gives two. The corpus demonstrates the implemented accounting semantics; the unbounded-family theorem remains a mathematical construction.
Engineering the closure cost without changing geometry
For an original spatial target face set , a layout induces its covering triangle blocks and the vertex-block dependency closure . The base64 buffer contribution to proof size, before JSON framing and other fields, is exactly
with descriptors, metadata, authentication paths and JSON framing counted separately. Twelve bytes encode one uint32-indexed triangle and 24 encode one float64 vertex; both raw lengths are multiples of three, so these base64 buffers need no padding. The security rule fixes complete reference closure; ordering controls how much unrelated material the blocks additionally reveal. Reducing that cost does not justify deleting required dependencies.
Figure 3. Same-region extraction on the actual photogrammetric mesh before and after spatial reindexing. All coordinate bits and triangle incidence remain identical under archived ID maps. Reduction concerns complete serialized proof packets; one-time maps are an additional shared cost.
We compare original indexing with deterministic Morton ordering [60] of vertices and triangle centroids, using 21 interleaved bits per coordinate for ordering only. Original coordinate bits, connectivity and winding are unchanged. Signed layout metadata commits archived bijective original-ID maps. Sixteen nonempty cells of a original-coordinate centroid grid supply identical target face-ID sets for both layouts. The complete-map experiment verifies the target predicate and exact geometry equivalence; the lightweight disclosure verifier checks pinned block covers and vertex closure, not a succinct proof of the geometric centroid predicate.
Median packet size falls from 7,417,431 to 842,247 bytes; the median of paired reduction factors is 9.13. Median referenced vertex-block count falls from 131 to 20.5, and median triangle overfetch from 36.37 to 1.55 times the requested face count. Original-ID maps and ROI target IDs are shared one-time data excluded from these per-packet sizes and retained in the archive. This is a measured layout improvement on one real mesh, not a universal compression bound, runtime speedup or new cryptographic primitive.
Earlier archaeology, structured models and qualified social claims
The expanded inputs distinguish geometry, excavation-related tables, a formal ontology and genomic/osteological classifications. Each supports a different kind of reconstruction. Their record counts cannot be summed into a population sample, and physical/model resolution does not establish interpretive certainty.
Göbekli Tepe beyond monument outlines
Dietrich et al.’s early-Neolithic study concerns cereal processing at Göbekli Tepe in the tenth/ninth millennia BCE [24]. We retain all twelve original supplementary workbooks unchanged, with source URLs, file hashes and CC BY 4.0 attribution. Read-only XML extraction preserves worksheet/cell coordinates, original numeric lexemes, formula caches and explicit missing categories. The workbooks contain 3,374 nonempty source cells. The S1 aggregate is 7,268 tools: 3,351 handstones, 1,945 grinding bowls/slabs and 1,972 pestles. This aggregate differs from the 821 distinct find numbers in detailed inventory subsets; neither is a count of raw tool scans. Overlapping subsets are not added.
The phytolith table supplies 48 morphotypes across nine samples, yielding 432 actual count cells. Independent column sums reproduce all nine published totals, from 290 to 1,210. Figure 5 exposes the full matrix, including recorded zeroes, instead of substituting a schematic monument. These observations support a source-specific analysis of plant processing; the tables do not record the operators’ sex, gender or authority. Any gendered labor claim would need an additional, admitted inference and evidence.
Figure 5. Source-conditioned quantitative evidence. (a) All 432 Göbekli phytolith counts; logarithmic color separates positive counts from recorded zero. All nine column totals match the source. (b) Original molecular/osteological categories for 57 records in the separate British dataset, with missingness retained. (c) An unresolved cross-table identifier difference. Authentication establishes the recorded values, while a join or social interpretation needs its own rule.
The sources also expose an unresolved join: S10 worksheet 1, K3 records M12–133, whereas S12 A1212 records M13–133. A second comparison uses M11–133 in S8 A8 and M13–133 in S6 A7. The strings remain unchanged. A valid digest on both records cannot establish that either pair refers to the same specimen. An alias relation must be a separate qualified interpretation record; strict equality cannot be replaced by an unaudited normalization rule. This is a real example of the boundary between source integrity and admissible inference.
Lourentzaki, Papadopoulou and Roche’s Göbekli ontoterminology supplies a machine-readable model of pillar finds and their named relations [53]. The original 274,754-byte RDF is retained without alteration. Parsing its RDF/XML syntax yields 3,732 unique triples, 575 subjects and 162 explicitly declared named individuals. These syntax counts are different from the repository’s Protégé axiom/individual metrics; anonymous nodes and reasoning conventions must not be conflated. The ontology is an authored conceptual model, not raw excavation geometry or complete world knowledge. Its DOI registry specifies CC BY-NC-ND 4.0; the ancillary license explicitly isolates this unchanged source from the other data licenses. We distribute syntax statistics and inspection code rather than a modified complete ontology.
Original Çatalhöyük artifact models
The Perry–Lara–Palma deposit provides thirteen artifact groups and fourteen original mesh files, including two variants of one skull figurine [64]. Its twelve OBJ members and two binary FBX geometries retain exact publisher bytes, with reversible storage recipes for OBJ and gzip wrappers for FBX: decompression recovers exact original archive members. The complete collection has 495,152 stored vertices and 628,600 original polygons, which yield 990,248 explicit fan-triangles. Excluding the additional skull variant gives thirteen groups, 490,985 vertices and 981,918 triangles. These are source-published printing-ready processed models; acquisition uncertainty, physical units and site georeferencing are unreported. Their dense meshes are not evidence of raw-scan fidelity or metric accuracy.
The analysis preserves the original polygons, coordinate bytes and source hashes, records the fan triangulation, and checks boundary/nonmanifold edges, components and zero-area faces. Flint and horse each contain one exactly degenerate fan-triangle and respectively one and two nonmanifold edges; the stamp has three boundary edges and two nonmanifold edges. These defects are reported without repair. An analytical triangulation does not certify convexity, planarity or absence of self-intersections in every source polygon. FBX local geometry is kept distinct from scene-node transforms.
Vertex-weighted PCA supplies documented exploratory cameras and independent per-object display scales. Three 256-pixel point-splat projections per model produce 42 computed views with explicit occupancy masks. We also retain scale-free edge-length, area and radial quantiles, PCA shape fractions and component-level summaries. Model-level aggregation gives each of thirteen artifact groups equal weight; it does not treat vertices as independent samples. The median first-PC variance fraction is 0.6791 (range 0.3486–0.9936). These describe the supplied representations rather than historical function. Figure 4 shows the complete collection. The source filename mother-goddess remains a quoted cataloguing interpretation, not a finding about the depicted person’s identity or social system.
Figure 4. All fourteen original Çatalhöyük model files, including two skull variants (Perry, Lara and Palma, CC BY 4.0). Computed PCA point-splat views use independent display scales and explicit sampling gaps. Labels report source model complexity; neither color nor mesh density encodes acquisition confidence. No physical calibration or gender-role inference follows from this display.
Social interpretation as an evidence obligation
Archaeology can test simple accounts of past social organization, but the tested variables matter. Yüncü et al. analyze 131 East-Mound Çatalhöyük paleogenomes and report changing kinship patterns, maternal connections and sex-associated differences in subadult gifts [87]. Their analyses provide no indication of sex-biased mobility. Co-burial, maternal relatedness, mobility, gender and political authority remain separate propositions. The 2026 v4 data record warns that earlier relatedness-simulation files were incorrect [88], making version-pinned interpretation especially relevant. The large chromosome-list archives are not burial/gender tables. We did not acquire the paper’s original S1–S28 individual tables and therefore report this as literature evidence, without claiming a reproduced 131-person analysis or bundling unlicensed sequence data.
For an actual larger social-data fixture, the artifact retains Cassidy et al.’s original CC BY 4.0 supplementary workbook [17]. It covers Iron Age Britain and European comparison samples; it is not an Anatolian substitute. Six source-coordinate-preserving extracts contain 2,405 comparative sample rows with 2,404 distinct source IDs, 1,485 complete WBK pair classifications (), 386 selected within/between-site relationships, 156 site/group aggregates and two 57-individual tables. The duplicated source ID StPet2_MIX is retained at both original rows. The selected 386 relationships are not a complete population-wide matrix; an unreported pair is not a measured unrelated pair.
The 57-individual summary records 24 XX and 33 XY molecular classifications. Osteological source categories are six Female, five Male, one Male? and 45 -. All 45 unreported labels remain visible. One XX record carries an osteological Male label and another Male?; these are source classification differences requiring contextual review, not discoveries of gender identity. Neither matrilocality nor a genetic classification directly establishes legal office, political control or universal matriarchy. The method allows such an interpretation to be represented and challenged by explicit premises, while preventing its presentation as an already measured source field.
A 2026-informed, offline research interface
The accompanying atlas links exact source cells, the phytolith matrix, model projections, relationship classifications and a provenance/detail pane. Selecting an item reveals its original coordinate or file hash and the applicable qualifications. Separate acquisition, transformation and interpretation records keep a computed projection distinct from a measurement. Detail is progressively revealed without hiding the immediate numeric view or requiring motion.
This design draws on actual 2026 work: Wiebel, Brender and Eckardt’s linked annotation/graph/attribute views for Roman millefiori artifacts [86], and LLUVIA’s staged uncertainty and level-of-detail workflow [54]. Linked views and uncertainty layers are established visual-analysis techniques; the atlas adds source-coordinate bindings to this paper’s verification workflow. It uses local HTML/DOM/SVG and original data, without a learned embedding or external runtime. These references guide interface construction, not a claim that their user-study effects transfer to archaeology or that this atlas has been validated with participants. All published figures are reproducible static scientific plots; interactive views are separate ancillary files.
Regional archaeology as a source of comparison constraints
The next research object is a qualified archaeological contrast. A comparison can change even when every disclosed number is genuine: the selected deposits, counting unit, taxon mapping, denominator or weighting may have changed. We therefore extend the source investigation to Iran, Transjordan, Oman/UAE, Lebanon and China, using original specimen exports and supplementary files. Persian, Arabic and Chinese queries and their actual outcomes are recorded in the artifact. Local-language institutional records supply useful excavation and terminology checks; a search result is not represented as an acquired database.
The historical question is deliberately narrower than a universal explanation of pig avoidance. Absence from routine human dairying does not imply absence of economic value or identify a prohibition’s cause. Price and Corcoran-Tadd contrast the reproductive, feeding and mobility properties that make livestock useful to households and institutions [67]; Price’s Mesopotamian synthesis also distinguishes dietary importance from administrative visibility [66]. At Xipo, isotope evidence supports intensive pig husbandry within a cereal economy [91]. These are substantial antecedents, not new discoveries of this paper. Ecological costs, household production, institutional procurement and culturally specified avoidance can coexist. Their relative causal importance requires dated, context-compatible evidence; a modern country’s predominant religion is not a treatment variable for a prehistoric assemblage.
A source-specific archaeological hypothesis at Chogha Mish
The original Open Context export for Chogha Mish, southwestern Iran, contains 16,824 distinct specimen URIs and 59 flat fields [9]. The project’s description of more than 30,000 historical specimens does not become the size of this acquired table. Three source predicates share the label Number of fragments; the flat columns have 16,681, 114 and 29 populated rows respectively. Each row has exactly one such value. Source-specific coalescing gives a recorded-field sum of 16,825; using only the first column omits 143 records. The separately published 4,705-row Protoliterate table is wholly included in the full export, which has 4,819 Protoliterate rows. They are overlapping releases, not two independent samples.
For an exploratory comparison, fix the source phase Protoliterate, source contexts Room floor and Occupational Debris, and the five exact taxon labels Sus sp., Bos taurus, Ovis/Capra, Ovis sp. and Capra sp.. The numerator is the recorded-field total of Sus sp.; the denominator counts these five selected labels. It does not count all excavated material, unique physical fragments, independent animals or meat weight. The primary chapter states that jaws were recorded repeatedly, one line per tooth, before merging in its published analysis (p.451, footnote4); no validated merge/refit crosswalk is present here. The 12 floor and three debris Sus URIs are fifteen source records, not fifteen independent bones or animals. For example, the maxilla record CGM12288 and upper-right tooth records CGM12289–12293 cannot establish five additional jaws. The fragment-field sum above must therefore not be promoted to a unique-piece count [51]. Sus sp. does not distinguish wild boar from domestic pig. The resulting source-coded contrast is
The preferred analysis excludes the source body-part group Head, leaving versus , with difference . Thus the direction is not confined to repeated jaw/tooth records. The seven postcranial Sus URIs have no validated individual/refit links and do not establish seven animals. Postcranial restriction alone does not remove recovery, transport or depositional effects. Figure 6 displays both the source contrast and the exact anatomical strata.
Figure 6. The left panel inventories the original regional workbook’s source rows by country label. The right panel shows the exploratory source-coded Chogha Mish comparison. Exact taxon, phase, context and fragment-field rules define the denominator. Anatomical strata expose the support of the floor/debris contrast; the values count recorded-field weights; head records can repeat a jaw once per tooth. The postcranial comparison is primary, and neither comparison counts independent animals. The proposed archaeological test requires matched dating, depositional review and controlled recovery.
The prospective hypothesis is that Sus enrichment of floor-context assemblages persists within independently dated coeval contexts after matched recovery/sieving and restriction to comparable postcranial material. It would be weakened or refuted if matched recovery or revised context assignment removes or reverses the contrast. Persistence would challenge a purely context-mixture explanation but would still leave animal transport, disposal and household management as alternatives. Useful new observations include screened sediment volumes, recovery mesh sizes, refitting fragments, element density, cut marks, direct phase associations and osteometric wild/domestic evidence.
This is hypothesis generation after inspection of the existing source. The independent source-transfer audit below tests its scope of transport; no independent confirmation of the exact floor-use hypothesis is established. The source phase has an editor-assigned range of 3600–3200 BCE; pooled legacy context types are not demonstrated individual households or exactly simultaneous deposits. All fifteen selected Sus records share a phase-level context URI; it is not a unique room/deposit identifier. Their editor-assigned dates establish phase compatibility, not coeval deposition. The primary chapter reports no sieving on p.446; mesh size, screened sediment volume and individual deposit associations cannot be reconstructed from this export. Fragments can derive from the same bone or animal, so an independence-based significance test would overstate the evidence. The original 2017 study already reports broad changes in pig representation and explicitly sets aside more specific spatial analysis [51]. The focused context/body-part contrast was not located in the checked chapter, but that check is not an exhaustive priority proof. Its warning that Sus may include wild boar remains applicable.
The shared weighting constraint exposed by anatomy
Let and be the two recorded Sus shares in a common anatomical stratum , and write . A comparison standardized to one declared anatomical distribution uses the same weights on both contexts:
These weights define an estimand. Equal stratum weights do not estimate an unknown ancient carcass distribution: the preferred four-stratum analysis assigns 25% to an axial stratum supported by only one versus two recorded fields.
Proposition 10.1 (Common-weight comparison bounds). For fixed nonempty matched strata,
If the shared-weight restriction is erased and independent simplex weights are permitted, the lower bound becomes
Proof. A convex combination of the lies between their minimum and maximum; placing all mass on an extremizing stratum attains either endpoint. For independent weights, minimize the first combination and maximize the second separately. Both choices are simultaneously feasible only after their coupling is removed.
For the preferred postcranial analysis, in source order Axial, Distal extremity, Forelimb and Hindlimb, the observed recorded-field shares are
Every shared nonnegative weighting preserves a weakly nonnegative contrast; strict positivity requires positive weight on Forelimb or Hindlimb. The exact interval is , and common quarters give . The raw head-including audit is separately available, with interval ; repeated tooth records prevent interpreting it as a unique-piece comparison. In the preferred analysis, coupling erasure lowers the endpoint to , obtained by contrasting a zero-Sus floor stratum with debris Hindlimb. The lower-bound loss remains exactly four percentage points. That endpoint answers a different anatomical question although no source value has changed.
More generally, let be a nonempty set of admitted joint assignments of mappings, weights or shared measurements, let be a relaxation formed by deleting specified consistency constraints, and suppose both infima below are finite. Define the constraint-erasure lower-bound gap
The inequality follows directly from feasible-set inclusion. This application-level functional records a particular lost coupling; it is not a new theorem of relaxation or interval arithmetic. Its value is that an archaeological qualification becomes an explicit mathematical constraint whose removal has a calculable effect.
A larger regional compilation: the estimand changes the sign
Gaastra, Lawrence and Tumolo’s comparative study supplies a second original workbook [30]. Its acquired worksheet has 577 rows and 216 distinct literal site strings. The article reports 245 sites and 786,353 remains; these are not silently identified with the worksheet’s row/string inventory. Repeated site phases or source assemblages are not independent settlements. The workbook spans eight country labels and contains 507 NISP and 70 TNF quantification labels. Number of identified specimens and total fragment counts are distinct units.
The strict computational cohort admits NISP, exact urban/rural labels after trimming whitespace, nonnegative integer cattle/Sus/caprine counts, and exact equality of their sum with the source domesticate denominator. First-failure exclusions are 70 non-NISP rows, 166 rows with other or unresolved settlement labels, and one invalid count/denominator row. The remaining 340 rows comprise 172 urban and 168 rural source assemblage rows. These exclusions are scientific admission choices retained with the original source coordinates; they are not claims that excluded observations are historically false.
For cohort , let count admitted Sus specimens, the selected-taxon NISP denominator, and . Two legitimate summaries are
weights source assemblage rows equally; weights their identified fragments. Neither directly estimates an average animal’s behavior, and neither is a causal effect of urbanism. The urban counts are 15,867/246,979 and the rural counts 11,070/153,688. The exact-source calculations give
The direction reverses without editing any row. This does not refute the original study’s region-, land-use- and time-sensitive interpretation; it identifies the question encoded by a particular aggregation rule.
Using population covariance over the admitted rows, with , straightforward expansion gives
Indeed, . Division by proves the identity. No model fitting or sampling assumption is needed. Under the explicitly defined interpolation , the zero is
This sensitivity path mixes two estimands; it is not a historical probability. Figure 7 makes the dependence visible instead of assigning a single geographic explanation to the sign.
Figure 7. Same admitted source rows, different comparison rules. The regional urban/rural sign differs between equal-source-row and pooled-NISP summaries. The displayed interpolation is a declared sensitivity path. The right panel decomposes that reversal into the equal-row contrast and the two cohort covariance corrections; no confidence interval or causal effect is encoded.
Exact deletion certificates for a declared source-row contrast
Sensitivity to omitted data has substantial prior art, including the approximate maximum influence perturbation work of Giordano, Meager and Broderick [33]. For the restricted equal-source-row contrast above, exact deletion has a simple finite solution. This specialization does not solve general case-removal sensitivity for fitted models or arbitrary aggregate ratios.
Assume initially . Sort the rates decreasingly as and the rates increasingly as . Let and , with . Retain at least one row in each cohort.
Theorem 10.2 (Exact sorted-prefix deletion threshold). For unit-cost deletion of source rows, the minimum number that loses the strict positive direction is
with if the displayed feasible set is empty. Removing the first sorted rows and first sorted rows supplies an attaining witness.
Proof. For fixed deletion counts , minimizing the retained mean requires maximizing the removed sum; its largest rates attain that maximum. Maximizing the retained mean requires minimizing the removed sum; its smallest rates attain that minimum. Exchange of any nonextremal selected rate for an available more extremal rate cannot increase the retained difference. Thus if any pair of subsets of these sizes loses the direction, the indicated prefixes do so. Conversely, each displayed prefix pair is an actual allowed deletion. Taking the minimum over the finite grid proves exactness.
Prefix sums allow evaluation in arithmetic operations after sorting, using exact rational rates and deterministic ID tie-breaking. Arithmetic bit cost is additional. The result applies to equal-cost whole-source-row deletion, not individual fragments, unequal deletion costs or pooled-NISP means. A negative baseline swaps cohort roles; a zero baseline has no strict direction to preserve. A finite threshold certifies safety against every permitted selection deletion of at most rows. It does not make such a deletion acceptable under an unchanged externally fixed cohort.
On the admitted 340-row compilation, the archived exact computation gives : every permitted deletion of at most two source rows retains the strict equal-row direction. One attaining witness removes gaastra-row-0003, gaastra-row-0200 and gaastra-row-0214. These are source-row coordinates, not three animals or a measure of historical sampling uncertainty. The certificate concerns a hypothetically changed selection; those omissions still fail the original fixed-mask verification request.
Binding the scientific question, not only its answer
Authenticated publication and aggregate verification are established techniques [52, 56]. The source-specific extension here makes the scientific query part of the externally checked statement. A scope query fixes the source snapshot root, exact ordered eligible-ID mask, unit, question, taxon/phase/recovery policies, estimand, weighting and a canonical rational interpolation weight. Admitted rows bind original source hashes and coordinates, cohort, numerator/denominator, qualifications and auxiliary metadata. Version2 admits both the NISP registry and a distinct Chogha recorded-fragment-field-total profile, with inference unit source-record-weight. The same certify and verify_certificate entry points dispatch on the typed source profile; NISP and recorded-field weights cannot be interchanged.
For Chogha Mish the verifier reads the complete original CSV, supplied inline in a bounded lossless witness or as an independently pinned local source file. It verifies its byte count/hash, all 16,824 unique URIs, exact ordered headers, project identifier, exactly-one fragment alias rule, count lexemes and source phase/date fields. It recomputes all ten context/anatomy aggregate rows before checking the signed count-bound registry root and eight postcranial openings. The unresolved predicate-URI correspondence of the three flat count aliases is retained as a declared source-specific coalescing assumption. A fabricated aggregate signed with the experiment key fails even if that false aggregate’s root is newly pinned: it is inconsistent with the independently fixed source.
The query fixes both cohorts’ complete anatomical support and one of three modes: pooled recorded fields, fixed shared rational weights, or the whole shared-simplex envelope. Empty/missing-stratum and cohort-specific vectors fail. The coded qualification profile fixes known no-sieving, unmerged jaw/tooth repetition, unestablished physical/animal independence, coevality and domestic status. Promoting these outputs to MNI, independent animals, unique physical fragments or same-year deposition fails, including when an attacker pins a query containing that unsupported promotion. The primary-source interpretation of these qualifications is reviewed scientific input; cryptography checks its preservation rather than discovering it from bone rows.
A packet pins the external query and current root, verifies the dedicated-domain checkpoint, checks unique row openings against the complete ordered mask and recomputes exact arithmetic. Chogha Mish accepts six valid controls and rejects 27 constructed attacks; its inline-source packet is 1,242,100 bytes. One local M4 Pro observation measured 0.612s for snapshot construction, 0.624s for certification and 0.605s for verification. These include full-source parsing. The separate Gaastra experiment accepts four controls and rejects sixteen mutations. Proof construction there uses a clarity-oriented path helper; neither experiment establishes a new performance bound. The Gaastra deletion radius is not applied to derived anatomical strata, because deleting a stratum and deleting original specimen records are different interventions.
The typed claim and exact result are authenticated. A caption or plot is covered only when a trusted rendering path consumes that verified claim and binds the presentation. Freely edited pixels or captions are outside the row verifier’s guarantee. Neither profile is succinct or zero-knowledge.
Conditional authenticity follows from the earlier tree-binding argument. An accepted unissued checkpoint gives a signature forgery; altered opened rows under the admitted root give a hash collision. With neither event, exact mask comparison excludes omission or transplantation, and deterministic arithmetic excludes a changed result for the pinned query. For the NISP adapter, admitted-row extraction and unit/mask appropriateness remain trusted scientific inputs. The Chogha adapter additionally recomputes extraction from the full pinned CSV, while its source-specific field and archaeological interpretation profile remains reviewed scientific input. A textual policy label does not independently prove that its admission rule is archaeologically appropriate or that all eligible material in the world was excavated. Current-root freshness needs an external pin or history mechanism. The public experimental key represents local custody of a calculation, not a museum’s endorsement; this signature domain cannot replace SCENE grants, permits or atomic spending. The construction’s specific contribution is to preserve source-induced comparison constraints and their sensitivity witnesses across presentation and verification.
Further regions supply different constraints
The remaining acquired sources enlarge the research programme while retaining different populations and measurement types. China’s 2026 compilation has 3,492 human and 2,919 animal isotope rows [90]; these are selected measurements from literature, not NISP abundance. Its human and animal Xingyi strings carry incompatible coordinates, and some broad phase fields need an explicit sign/unit convention. An audit produces 88 site-coordinate-phase join candidates, only 24 with identical coordinates and exact date bounds; even those remain candidates pending context and specimen-selection review. The original Xipo workbooks provide 87 selected faunal metadata rows, including 70 Sus, with Yangshao and later Zhou contexts kept separate [91]. A selected isotope-sample fraction cannot be substituted for the site’s reported pig-NISP proportion. These sources motivate authenticated join and chronology obligations, not a merged cross-country consumption estimate.
The Oman/UAE caprine source has 354 rows in S1, including five without LSI, and 71 S2 osteometric rows including six modern references [55]. A modern Iranian reference is not an archaeological Iranian animal; Daba funerary and Salut settlement/ceremonial selection are distinct. The two source indices and share one lateral depth . For positive admitted measurement and ratio intervals, joint feasibility requires exactly
Each interval is the projection of one ratio constraint onto ; a common member permits independent choices of and satisfying both. Checking the ratios separately can instead use incompatible depth witnesses. The exact checker finds all 71 rows feasible under hypothetical closed mm input boxes and ratio-point boxes. Two rows fail if input measurements are instead treated as exact. This is a precision-policy sensitivity result, not a source-error finding: stated caliper accuracy alone does not prove the hypothetical rounding model. Six focused tests cover these source and constructed consistency cases.
Lebanon’s 2026 author files concern wild Pleistocene fauna roughly 400–100 ka, not domesticated livestock or later food prohibitions [71]. Named-taxon NISP totals 77, 61 and 65 differ from broader totals 559, 63 and 99 that include indeterminate material. A blank row for the as-Kelb Sus count in the main table and a supplementary 0.0% entry are retained as distinct source statements. Tepe Yahya’s original isotope workbook has 254 successful caprine records, 30 failed records and 76 analytical-repeat rows from 36 IDs [72]. Literal quality filtering and source phase-label disagreements change eligible cohorts; no value is silently repaired. Transjordan’s small 2026 comparison is represented by a clearly labelled eight-column transcription of the original main table, not an acquired specimen-level supplement [85].
The transferable research requirement is therefore concrete: preserve the eligible cohort, units, mapping decisions, recovery qualifications and shared variables needed for the particular question. A future corroborating excavation can strengthen or falsify the narrow Chogha Mish hypothesis, while new source releases can revise the admitted mappings. The formal checks preserve those decisions and expose their consequences; the archaeological interpretation remains answerable to new material evidence.
Independent transfer and archaeological admission
The prospective archaeological hypothesis is that postcranial Sus representation among recorded cattle, caprine and Sus remains is higher in floor-use deposits than in occupational refuse, within independently dated coeval contexts and comparable recovery conditions. A country label supplies none of these conditions. Candidate collections and exclusions were recorded before calculating the independent comparison; after the Chogha tooth-row issue was identified, the postcranial specification and final source-label mapping were frozen separately. These are locally timestamped decisions, not an externally registered preregistration. An exact transfer requires a source-supported deposition crosswalk, phase and recovery match, and common anatomical support; an approximate occupation/dump comparison retains a different declared role.
The primary Chogha study materially limits what can enter that hypothesis. Its p. 446 reports Wheeler’s recollection that sieving was not practiced and explains the omission of detailed spatial analysis. Page 451, footnote 4, documents repeated spreadsheet entries for teeth belonging to a mandible or maxilla, which the authors’ cleaning merged [51]. In the acquired export, consecutive records CGM12289–12293 specify upper-right P3, P4, M1, M2 and M3 beside maxilla record CGM12288. Their generic completeness labels do not establish five independent maxillae. Unique row identifiers cannot establish fifteen animals, and the seven selected postcranial Sus records have no supplied individual/refit links. All selected rows share a phase-level context URI and the editor-assigned 3600–3200 BCE interval: neither identifies coeval rooms or deaths. The chapter already published the broad temporal pig trend; our focused inspection is hypothesis generation without a worldwide-priority claim.
An independent English catalogue, OXSTAD16, provides a more limited test. Broderick’s CC BY 4.0 source has 605 unique catalogue identifiers and 88 fields; its metadata defines Quantity as the specimens represented within a record, not individuals [14, 15]. All 606 CSV lines were retrieved through ordinary public web access after direct local requests returned HTTP 403. Restoring the original CRLF serialization reproduced the publisher’s legacy MD5 checksum 779fab3221ec3063b2d620ab22506e8f; the resulting original byte sequence contains 150,707 bytes. This checksum match supports reconstruction, not a modern collision-resistance assumption or an issuer signature.
Exact floor/refuse replication remains unidentified. The excavation report distinguishes a real floor sequence from dismantling and cellar backfill, while candidate floor-sequence contexts 5027/5031 are absent from the bone catalogue. Context 4014, labelled occupation layer in that catalogue, belongs to a later latrine; occupation cannot therefore be translated automatically into floor use [79]. The frozen partial test instead admits literal Phase 2, source-unissued records, certain selected taxon labels and explicit postcranial anatomy, comparing declared occupation-layer labels with dump labels. Question-mark taxa, head/dental material, unknown anatomy and the original anomalous Phase values 3059/4004 remain excluded without repair. The 29 admitted records produce
There are two versus eight source context IDs. Forelimb, hindlimb and distal contrasts are respectively , and 0; axial support is absent, rather than an observed zero. Common simplex weights on the three supported strata give the exact interval , with equal-weight index . The positive direction does not transfer to this declared partial comparison. It neither confirms nor refutes the stricter floor-use hypothesis, and no independent-animal significance test or causal interpretation follows.
Larger acquisitions preserve the same admission discipline. Forcello’s published specimen examples are usable, but the required original phase/context join was described as an Access table on the dissertation CD and was not acquired [83, 84]. The Russian collection supplies 123,937 original source-value records from its primary worksheet, with coordinates, taxonomy, anatomy and contextual qualifications [69]. This is not a harmonized NISP cohort or an individual-animal count. The bundled compressed exports retain source rows and fields; they are not original XLSX bytes, all worksheets, or an offline reconstruction of the workbook. Zvyagino-4 includes floor material, but its yellow-ash comparison has a roof-material interpretation, and cultural agreement does not establish simultaneous deposition [2]. Roof fill is not admitted as a refuse proxy merely to obtain a larger comparison.
The receipt therefore includes the complete source and interpretation-policy SHA-256 pins,2 original coordinates, a full admission ledger, counting unit, deposition crosswalk, recovery/phase predicates, supported weights and the exact/partial transfer role. The executable replication_analysis.py recomputes the comparison; twelve tests cover source/policy mutation, catalogue and admission coverage, unchanged source anomalies, context/recovery/taxon/anatomy exclusions, exact arithmetic, attained weighting bounds and missing or empty support. These are concrete SCENE admission inputs. The current standalone OXSTAD16 audit verifies local pins and calculations; it does not issue a SCENE scope certificate. Authentic bytes and reproducible arithmetic leave the archaeological premises independently open to falsification.
Source-exact reconstruction as an authenticated obligation
A submission budget creates a second opportunity for scope loss: a smaller representation may preserve a rendered shape while changing the actual evidence file. We therefore implement source-exact reconstruction profiles within SCENE. Each admitted target has a retained role: original publisher member or derived source-cell extract. Its reconstruction preserves the complete byte sequence, including attributes, order, comments, decimal spelling and whitespace. The profile separately binds that scientific target, its physical payload and the packaged decoder files. All manuscript figures remain ordinary PDF inputs; the TeX compiler need not execute a reconstruction program.
What the 2026 methods contribute
Chen et al.’s Dolphin, published in February 2026, uses the first two XOR bits and a following seven-zero interval to skip a redundant prefix in oscillating binary64 data [19]. We implement and measure that prefix-skip submechanism in our own explicitly framed stream, with unchanged-word, leading/trailing-zero-window and raw-XOR alternatives. This is not a benchmark of Dolphin’s full adaptive framework. Binary64 recovery alone does not recover original decimal text, so this candidate is admitted only when formatting every recovered value at the declared precision exactly restores its original token. Unsupported spellings retain literal bytes.
The May 2026 triangle-removal study distinguishes geometry-equivalent recovery from original triangle serialization and omits attributes from its experiments [89]. That distinction changes our reconstruction target: topology or image equivalence cannot justify dropping serialization residuals. The September 2026 Parquet integration of ALP explicitly derives from its 2024 algorithm [1, 6]. We likewise distinguish current experiments from older component ideas: fixed-point scaling, predictive residuals, byte/bit planes, dictionary sharing and LZMA are established techniques. Lossless predictive geometry compression also predates these works [41]. The contribution here is their bounded, source-committed composition with an executable scientific scope, tested on the actual artifact.
An exact lexical profile and a constraint used only as prediction
The implemented SCENE-LEX3 profile stores a literal/template skeleton for OBJ records and whitespace, then separate token channels by record kind, arity and column. Unrecognized records remain literal. Fixed-decimal channels may use scaled integers, signed deltas, or byte/bit planes; face channels may predict from earlier components or columns. A finite measured choice includes each codec’s metadata and compressed body. Negative-zero flags, missing slash components, exceptional spellings and the skeleton are charged payloads. No vertex, face, material line or texture coordinate is deleted or reordered.
One useful constraint has an entirely different role from archaeological admission. For shared fixed precision , let and let be the exact scaled integer lexemes of a supplied normal. With its recorded sign , define
Integer square root makes the predictor deterministic. The residual and lexical sign, including negative zero, are stored exactly. Therefore nonunit, rounded or exceptional source normals remain unchanged: unit length is an opportunity for a shorter residual, not an imposed correction or a claim about measurement accuracy. The decoder needs earlier channels, the sign bits and every residual. Erasing any of them destroys the reconstruction obligation even if a viewer still displays a plausible object. Unsupported precision/range profiles fall back to an admitted literal representation.
Two commitments and a typed revision frontier
Fix mandatory target files and their original lengths and digests before compression. An admitted recipe produces payload , metadata and named dependencies , subject to
Experiments compare the actual byte arrays as well as the pinned SHA-256 values. A logical-source root binds ordered target paths, original/derived roles, lengths and digests; a domain-separated transport root binds payloads, codec identifiers and packaged decoder-file digests. A dedicated-domain checkpoint signs both, with a fixed decoding policy and bounds. The issuer key and expected roots must be obtained independently to resist an active substitution attacker. Reading all pins from the same archive checks internal consistency only. The public research key does not represent a source institution.
Proposition 12.1 (Source-preserving composition). Assume a trusted deterministic decoder/extractor/verifier runtime, externally pinned issuer key, source/transport roots, scientific policy and current scope root. An accepted reconstruction followed by accepted full-source scope verification yields the same qualified result as direct verification on the pinned original source bytes, except for signature forgery or a source/payload hash collision. If and the logical source members are unchanged between two accepted storage profiles, the scientific-record frontier is empty; the transport frontier may still be nonempty.
Proof. Transport checking fixes each complete payload and declared decoder dependency. Bounded deterministic decoding followed by the original length/hash check fixes its source output, unless the pinned digest is collided. Full-source extraction therefore receives the same bytes and policy as the direct calculation and recomputes the same registry, selection and exact result. A checkpoint outside the admitted issuance history would instead forge its signature. Applying this argument to both profiles gives equal scientific records. Their payload and decoder-file records need not agree, so their transport frontiers need not be empty. Changing invalidates the stated equality premise even if every source byte is unchanged.
This connects compression to revision-cut reuse without confusing two events. An unrelated storage change can preserve a scientific cut after reconstruction, while a changed taxon mapping or weighting is a scientific-policy revision. Transport freshness must still be verified against its own current pin. The implementation fixes decoder identifiers and never executes downloaded recipes, shell instructions or serialized objects. It limits individual source outputs to 128MB, the total declared reconstruction to 256MB and each XZ decoder’s working memory to 64MiB. The last bound is not a total process-memory bound. File checks cover the packaged Python dependencies under a trusted bootstrap, interpreter, liblzma and cryptography runtime; code already executing cannot authenticate its own replacement. Path traversal, symlinks, malformed frames, trailing compressed data and changed existing source destinations are rejected.
Storage cost is a union, authorization cost counts occurrences
BagIt and OCFL already separate inventory/fixity from logical organization, and conditional reconstruction is longstanding delta-compression practice [47, 49, 62]. Let a finite acyclic reconstruction graph have nonnegative stored-node costs , mandatory dependency closure , and selected recipes . Its additive physical cost is
Shared dictionaries and decoders are counted once in this union. A fixed closure model is ordinary monotone submodular weighted coverage: the additional nodes introduced by one recipe can only decrease as grows. For binary recipe selectors and stored-node indicators , a finite joint model is
Here includes selector/recipe serialization. The model does not include layout-dependent solid compression unless those costs are explicitly added. The implemented profile selects measured pertarget candidates with one common fixed decoder, rather than solving this general joint optimization. The complete outer ZIP byte count is the authoritative size test.
SCENE endorsement budgets instead charge each disclosed occurrence. Reusing one compressed source in many displays can save storage while consuming many authorizations. Identifying the two costs would allow compression or deduplication to manufacture endorsement capacity. Keeping their domains separate is an operative security condition, not a cryptographic property of the compressor.
Measured source recovery and archive budget
The benchmark includes all thirteen OBJ members: the SEAMS mummy and twelve Çatalhöyük models. It compares raw-member DEFLATE9, XZ6 and complete SCENE-LEX3 frames; the frame includes its skeleton, sign/exception streams and codec metadata. Every selected payload restores its exact original source bytes. Across 129,303,415 original bytes, raw DEFLATE9 uses 32,988,098 bytes, XZ6 uses 21,795,388 and the complete thirteen SCENE-LEX3 frames use 16,402,504: 24.7% less than XZ6. The release selects XZ6 for one model where it is smaller; the thirteen selected model payloads total 16,362,028 bytes. A single local run measured total encoder times 11.908s, 52.446s and 191.651s, respectively, and decoder times 0.094s, 1.001s and 20.708s. Encoder selection effort and lexical checking are included for SCENE-LEX3; these are not isolated throughput or optimality claims. Cumulative process peak RSS was 760,004,608 bytes. Figure 8 shows each actual member rather than only the aggregate saving.
Figure 8. Complete exact-byte representations of all thirteen source OBJ members. The left axis is logarithmic; the right panel gives additive frame totals. DEFLATE9 counts a raw RFC1951 payload, XZ6 its full frame, and SCENE-LEX3 its header, skeleton and compressed channels. Every method restores all original bytes. Decoder/code/recipe costs, other source assets and outer-ZIP structure are measured separately at package level.
The 2026-inspired candidate was evaluated on 80 numeric channels; 69 exercised its prefix-skip branch, for 395,697 such transitions. It won only one constant channel, where the skip branch was unused. Thus this corpus supplies a measured negative transfer result for that mechanism: the observed gain comes from other exact lexical/integer/residual choices. A comparison that called the selected compressor “Dolphin” would misidentify the implemented method.
The original photographs, textures, workbooks and unchanged licensed ontology remain available. Selected large CSVs use exact-byte XZ storage, with their original line endings and count lexemes restored before source parsing.
The release must satisfy two physical budgets. Let be the sum of the stored files’ lengths after the outer ZIP is opened, the complete ZIP length, and a declared allowance for the processed PDF and build products. We require
The earlier 49,778,451-byte ZIP had 56,697,985 bytes of opened members: checking alone was insufficient. The corrected packager measures both actual container properties and fails at either boundary. The allowance is a conservative local engineering margin, not an assertion of arXiv’s exact server accounting. The much larger reader-restored logical source volume is a third quantity and is reported separately; TeX processing does not run that restoration.
The two costs can also be attached componentwise to a fixed dependency closure: , with recipe and container overheads additionally counted. Shared dependencies remain unions in each coordinate. A plan feasible for one resource can fail the other; direct complete package measurement remains necessary for layout-dependent terms. This refines Equation (45) to a constrained multi-resource problem without changing any scientific source value or comparison policy.
To reduce as well as , the new profile uses exact-byte XZ for eligible data records and libjxl 0.11.2 JPEG reconstruction for all 77 source JPEGs [44]. They total 22,725,571 original bytes and 18,753,153 stored JPEG XL bytes; all 77 original JPEG byte arrays and SHA-256 values are recovered. This is JPEG-bitstream reconstruction, not pixel decoding followed by a new JPEG encode. The wrapper requires reconstruction metadata, fixed decoder version/options, a 60-second timeout, an output-file limit and the original source digest. Native decoder memory is not bounded by this wrapper. The trusted installed libjxl runtime is an additional dependency, while no native binary is hidden in the archive. These established codec capabilities are attributed, not claimed as a 2026 invention.
The outer ZIP retains all manuscript figure PDFs. Original photographs, textures, workbooks and source metadata restore to their original paths with unchanged bytes and roles. Generated result arrays, duplicate plots, browser captures and HTML are rebuilt by the offline runner. Earlier investigator OBJ gzip wrappers were replaced while publisher OBJ bytes/digests were preserved. The Russian exports retain their source-value-extract role; compression does not make them an original-workbook archive. The artifact records the final , , values and verifies every restored target before scientific processing.
Authenticated query contracts across execution and materialization
The database problem is to preserve a qualified relational question across storage reconstruction, alternative execution plans and cached results. Authenticating aggregate answers has substantial prior art [52, 56]. Our database contribution is an executable source-specific query contract that connects complete source extraction, relational bag semantics, anatomical support and the existing SCENE scope verifier. The independent SQL execution is checked against that verifier, rather than treating a database’s declared source hash as proof of its contents.
A contract binds the source-byte and complete parsed-record digests, current scope root, original-column mapping, registered parameterized statement hashes, parameters, count unit, result schema/order and scientific qualifications. Its eligible source bag retains each record’s URI, ordinal, source-value hash and recorded weight. Unique URIs prevent an accidental duplicate import; they do not deduplicate animals, teeth or physical fragments. For the complete Chogha source, SQL COUNT(*) = 16,824 whereas SUM(recorded_weight) = 16,825. Substituting one for the other changes the counted unit. Blank original fields remain empty strings; the importer validates integer-valued decimal lexemes and exactly one active count alias before SQLite casts the coalesced value. SQLite’s own count/SUM/NULL semantics are not a substitute for this source admission check [75].
For a fixed grouping operator , an answer carries both grouped counts and complete selected-row coverage. The acceptance condition is
Here the existing full-source scope certificate establishes the selected bag and expected aggregates from the independently pinned original CSV. A received coverage hash alone does not establish completeness: the ordered URI/ordinal, record-hash, cohort, anatomy, taxon and weight tuples must equal that verified bag. Missing groups are rejected rather than interpreted as observed zeros. Only fixed, registered statement templates execute; packet-supplied SQL text cannot enter the database engine.
This condition separates result equivalence from contract equivalence. Alternative scan, index and materialized-view plans may answer the same contract. Changing numerator parameters, anatomical support, unit or source version cannot answer it merely because a scalar ratio happens to agree. Even multiplying a numerator and denominator by the same factor fails the exact-group check. If two storage recipes restore identical source bytes, the source and query roots, bag and qualified result are unchanged. A different transport payload alone therefore does not stale this query contract. A changed source root or parameter contract invalidates the materialized-view binding; these are different revision domains. The argument composes deterministic relational evaluation with SCENE’s existing source-binding result, under its stated runtime/key/root assumptions. It does not claim a new SQLite index or a succinct proof that a physical plan executed.
Actual relational workload. The artifact imports all 16,824 records and all 59 original field lexemes into a fresh SQLite database, with an explicit source-column dictionary, URI uniqueness and count-alias consistency constraints. Complete base-value verification recomputes every parsed record hash against the original source, including fields unused by the current aggregate. The declared postcrial query fixes Protoliterate, both context labels, the five exact taxon labels and four anatomical strata. It returns eight cohort/stratum groups covering 225 source records, 5/61 versus 2/164, and the same exact contrast 349/5002 through the scope_contrasts.certify and verify_certificate interface.
Three measured physical strategies use a table scan, a covering index, and a covered materialized bag. Their EXPLAIN QUERY PLAN outputs are retained as diagnostics, never parsed into an authentication decision; SQLite warns that this format can change between versions [76]. The experiment separately reports import/index/materialization cost, database bytes, aggregate SQL time, full source-value checking and full-source answer verification. In one SQLite 3.53.3 run, aggregate SQL took 4.215, 0.127 and 0.079 ms, respectively; database files occupied 24,399,872, 27,901,952 and 27,975,680 bytes. Import and complete base-value verification took 0.753 and 0.910 s, while full-source answer verification added 0.570, 0.572 and 0.575 s. These costs are reported separately so the small aggregate time cannot conceal verification. The materialized view binds its coverage and scope to the complete contract before its answer can be accepted. Eight constructed attacks cover changed aggregates, omitted groups, rehashed incomplete coverage, duplicate coverage, changed parameters, unregistered statements, animal-unit promotion and stale view bindings; all are rejected. Twenty tests additionally exercise relational keys, count-alias constraints, unchanged-ratio tampering, source/column mutation and storage-wrapper invariance. Timings are one local sequential run, not a claim of optimizer novelty or end-to-end speedup: full-source verification is additional to the aggregate SQL operation. The generated database is an inspectable reproduction output; it does not replace the pinned original CSV.
A frozen documentary-language extraction and revision pilot
The language question is whether a source-bound parser preserves the distinction between a representation claim, a qualification and a calendar statement before they become scene/database records. We implement a restricted English finite-state/rule parser, with no trained model or language-model inference. Finite-state information extraction, lexical negation handling and rule-based temporal normalization have substantial precedents [18, 39, 77]. We evaluate a particular documentary-frame contract and its revision boundary; we do not identify ordinary regular expressions as a new parsing algorithm.
Natural source records and development separation. The corpus includes every description sentence from the 33 Cleveland Museum of Art records in the visual-source inventory, all 15 Metropolitan Museum titles, and all 48 literal catalogue-date strings: 243 records comprising 180 prose sentences, 15 title fragments and 48 date phrases. Text remains an exact substring of a licensed CC0 JSON field, with its original source digest, Unicode-codepoint offsets and text digest. No sentence was selected because a trigger or parser succeeded. Two near-duplicate object pairs were merged before splitting, using word-trigram Jaccard similarity of at least 0.65 on segments of at least 12 words. The 48 objects therefore form 46 source groups. A fixed SHA-256 group rule assigns 188 records in 38 groups to development and 55 in 8 groups to test. Whole objects and duplicate clusters cannot cross the split; common short date/title strings are not evidence of independent linguistic constructions.
Reference frames were prepared in an investigator’s AI-assisted workflow, blind to the evaluated parser’s predictions. They are provisional reference annotations, not independently human-adjudicated gold or an inter-annotator-agreement study. Structured catalogue dates remain separate weak metadata and are never substituted for text labels. After three development inspections, the parser’s code/lexicon hash was frozen before test or external-corpus evaluation. A blind source audit then corrected two development-only cue labels; test bytes did not change. Current development scores use that corrected reference, rather than claiming perfect fit to the earlier version. The artifact retains both audit receipts.
Frame grammar and measured accuracy. Admitted relations have an inflected depict, show or represent trigger, a bounded concrete argument head and source-scoped polarity/modality. Entity classes include documentary king/queen/man/woman/deity/animal/other labels; they do not assert biological sex, ancient authority or a real person’s identity. Unsupported paraphrases abstain. Dates require an explicit textual era and normalize inclusive BCE-negative/CE-positive intervals without year zero. For example, an explicit third century AD becomes 201–300; an unmarked century, relative age or dynasty number is not silently converted using catalogue metadata. Source spans, date alternatives and record genre remain explicit.
Exact-frame micro scores use , and ; undefined empty-target ratios are reported as null. An abstention on an admitted reference frame counts as a false negative. On the 55 held-out records, the frozen parser matches the single positive relation frame and all ten admitted date frames. These tiny positive denominators cannot establish broad relation/date generalization. The natural relation positives do not test negative or uncertain relation scoping. Negation has one matched cue. Uncertainty extraction instead gives 9 true positives, 4 false positives and 2 false negatives: precision 9/13, recall 9/11 and . Whole-record outputs match 51/55 records; grouped results are also retained. Two near-duplicate records falsely treat about in brought about as epistemic uncertainty, and two additional errors concern cue boundaries. No rule was changed after observing these failures.
Two declared ablations distinguish the target from word presence. A deliberately weak, unlinked entity-mention heuristic produces 49 spurious held-out relation frames under the strict trigger/argument contract; it is not a competitive information-extraction baseline. Removing qualifications from the same grammar loses all admitted negation/uncertainty cues. Its positive relation score remains unchanged in this sparse test because the positive relation has no such qualification. The artifact reports all slot scores, genres, empty records, coverage and failures instead of equating document-level empty answers with correct extracted entities.
Fresh source receipts and semantic-cut reuse. Let be the extracted span frames and the declared slot compiler. A dedicated-domain receipt binds
Verification pins the current text, record, genre and rule identity, recomputes the extraction and checks the separately scoped slot contract. It authenticates this parser’s documentary output, not linguistic correctness or historical truth. A semantic decision cut may avoid a changed raw-text leaf when its admitted slot record is unchanged, but a fresh source-to-frame receipt remains mandatory. An old receipt cannot answer a new text merely because its normalized slots happen to agree. Source/slot-policy changes remain distinct revision events.
The predeclared fourteen development-derived edit pairs are explicitly investigator constructions, not publisher revisions. All fourteen produce the intended extraction distinction; four unchanged-slot cases transport their cuts, and every old source receipt is rejected against the revised text. An expanded systematic audit has 358 constructed variants: 243 whitespace changes, 57 cue removals, 44 date substitutions, 7 negation insertions and 7 entity substitutions. It matches 353 intended slot outcomes and rejects all stale source receipts. The five failures are conservative false invalidations: internal whitespace in multiword cue lexemes remains in the frozen slot representation. Those benign edits unnecessarily prevent reuse; they do not admit stale text. The fixed pipeline and these failures are retained for inspection. Twenty-eight software tests check source grounding, scoping, chronology, abstention accounting, external receipt conditions and actual revision-cut verification. The frozen lexicon’s age-role collapse is another disclosed out-of-corpus counterexample; it is not validated as an inference from a child’s representation to an adult social role. This pilot motivates broader independent language evaluation, without returning to a learned model or inferring source consent.
External human-annotated temporal evaluation
To distinguish investigator-curated museum references from an external human benchmark, we applied the unchanged date grammar to every English document in the acquired AncientTimes and WikiWars releases. AncientTimes was manually annotated for historical temporal expressions; WikiWars provides human TIMEX2 annotations in historical narratives [57, 78]. The parser rule hash was frozen at 20:38:15 UTC on 27 October 2026, before external annotation inspection; neither corpus was used to adjust the rules. The evaluation adapter was written subsequently. These are transfer scores for the complete acquired English editions, not a reproduction of either paper’s experimental protocol. Two AncientTimes articles derive from WikiWars, so the corpora do not constitute independent replications.
First, we supplied each human TIMEX span to the grammar and measured year projection. A complete four-digit calendar year in a human DATE, TIME, or untyped TIMEX2 value, optionally prefixed by BC and followed by calendar/time refinements, projects to ; BCE is negative, CE positive, and year zero is rejected. Durations, sets, references, incomplete years, centuries and decades remain counted but unsupported. Month, day, time and season correctness are outside this projection. Exactly one returned frame constitutes an attempt, even when it covers a proper substring of the supplied span; zero or multiple frames abstain, without selecting a candidate using the gold value. Covered precision is correct/attempted and coverage is attempted/eligible. Human values may incorporate surrounding discourse, whereas this task supplies only the span text. Consequently these scores cannot be interpreted as complete TIMEX normalization or span-extraction accuracy.
AncientTimes contains five documents and 306 body TIMEX annotations in the pinned release. Of 210 complete-year values, the grammar attempts 124 and matches 123: covered precision 99.19%, coverage 59.05%, and correct year-pair recall 58.57%. The remaining 96 annotations have unsupported values/types. WikiWars contains 22 documents and 2,659 body annotations; of 2,088 complete-year values, it attempts and matches 83: covered precision 100% and coverage/recall 3.98%. Its 571 other annotations are unsupported. Relative/context-resolved dates and ordinary years without the explicit era required by the museum grammar account for substantial abstention. The one AncientTimes mismatch is a season encoded under one gold year whose surface expression spans two BCE years; it remains an error under this scoring rule. Correct year pairs constitute only 40.20% and 3.12% of all human spans, respectively. The high covered precision therefore does not establish broad temporal understanding.
Second, we ran the same grammar on each complete document and scored exact source spans against all human body temporal annotations, including unsupported types. AncientTimes yields 124 predicted spans, 78 exact matches, precision 62.90%, recall 25.49%, and F1 0.36279. WikiWars yields 103 predictions and 49 matches, precision 47.57%, recall 1.84%, and F1 0.03548. These separate extraction scores expose the restriction’s cost; they include neither normalization credit nor an asserted comparison with HeidelTime. The cryptographic contract binds the source and declared parser output even when linguistic recall is poor; authentication supplies no missing linguistic inference.
All spans address XML-decoded TEXT Unicode codepoints; nested elements and tails retain their order, with no whitespace normalization. DCT/DATETIME outside TEXT is excluded. Reconstructed offsets agree with all 306 supplied AncientTimes offsets and all 2,640 supplied WikiWars offsets; 19 WikiWars annotations lack that attribute. The ancillary evaluator rejects changed archive or parser hashes and records original member and decoded-text hashes, aggregate denominators, and unsupported types. The source ZIPs contain no corpus license and explicit original redistribution permission was unresolved; we therefore distribute code, acquisition checksums and aggregate receipts, with separate acquisition required to rerun this external study. The museum experiment remains available offline in the package.
Dimensional regression fixture
Petrie’s digitized 1883 publication reports the Khufu casing-base mean as 9068.8 inches and reconstructed original height as 5776 inches (chapter VI, sections 21 and 25) [65]. Reproducible conversion using 25.4 mm per inch gives 230.34752 m and 146.7104 m. The signed fixture rounds these to 230348 and 146710 mm. This convention does not calibrate an 1883 survey to modern SI. Petrie’s probable errors, 0.5 and 7 inches, are preserved as historical descriptions, not recast as standard deviations or guaranteed bounds.
The El-Kurru field report gives Ku. 1’s plinth side as 26.65 m and reconstructs original height at approximately 34.5 m [26], p. 60. No numerical uncertainty is supplied, so none is invented. Reported upper-course block-cut angles of 70–75 degrees describe a different scope from the ideal-envelope angle degrees. Courses step back. The artifact retains that qualifier; transferring the block-angle claim to the whole envelope is a benchmarked scope-substitution attack.
The two newly generated meshes each have five vertices and six triangular faces. They are purposively selected fixtures, not a representative cultural sample. A supplemental Nuri base measurement has no reported height in the selected source and produces no mesh. We inspected digitized 3D resources from Digital Giza and the Zamani Project; the examined Meroe model was not offered as downloadable with a redistribution licence. No viewer geometry was extracted. This small pyramid fixture therefore uses numerical transcriptions from digitized publications rather than those scan datasets. The principal actual-image and photogrammetric inputs are described in § 7. Source locators, rights notes, file hashes and retrieval limitations accompany the cards. No copyrighted publication pages or photographs are redistributed. This fixture’s newly transcribed facts and generated envelopes are CC0; software is MIT licensed. The SEAMS input and its derivatives retain their separate CC BY 4.0 attribution.
This distinction follows the London Charter’s emphasis on sources, documentation and sustainable visualization [80]. Geometry makes evidence obligations concrete and reconstructible; it supplies neither a universal theory of ideology nor secret key entropy. For a deterministic key derived solely from public geometry , in Shannon’s sense [74]. This observation does not rule out computational hardness of other public-instance problems; it rules out treating this publicly specified deterministic key as a secret.
Reproducible computational evaluation
All reported runs used an Apple M4 Pro, 48 GiB memory, macOS 26.3 arm64, Python 3.14.2, cryptography 46.0.4, NumPy 2.4.1, SciPy 1.17.0, Matplotlib 3.10.8 and Pillow 12.1.0. The fixed protocol clock is a deterministic test input, not an observation of the execution date. No model was trained or queried by the method.
Capacity instances. The artifact evaluates 64 fixed-closure budget pairs for the two-monument scene, eight shared-prerequisite stars, the Fano obstruction, and 60 seeded small hypergraphs. A general-purpose integer solver and independent memoized exhaustive enumeration agree on all 133 instances. Fractional capacities are computed by linear programming with independently solved price duals; recovered rational prices certify their reported upper bounds. The Fano instance has integer capacity one and fractional bound , the only strict floor gap in this suite. The eight-alternative star has true capacity one and erased capacity nine. These small examples verify implementation and deductions; they do not establish large-instance solver complexity or historical causal regularities.
Adversarial protocol suite. Ten mutation families have 40 variants each: purpose, audience, narrative, arrangement description, height, spatial position, uncertainty, dependency edges, block-/envelope scope and cloning. Replay, fresh-permit overdraft and transitive source omission add three cases, for 403 unauthorized cases and one positive control. The source-omission case is a deliberately synthetic cross-source dependency, not an archaeological hypothesis. The verifier accepting only authenticated input cards admits all 403 attacks; exact scene binding with closure but without spending admits two; using only direct sources admits three. Full closure and atomic spending admit none. All four accept the single positive control. The ablations are explicitly implemented toy verifiers, not evaluations of C2PA or other deployed products. Variant counts indicate constructed coverage, not independent population samples or empirical attack probabilities.
Independent tests and interface checks. The automated tests include every Merkle index for sizes 1–65, 960 selected-object views over all 64 four-node DAGs, mandatory-opening omissions, concurrent replay and overspending, atomic rollback, malformed encodings, key/ledger substitution, scope, expiry, source-byte modification and corrupted audit state, plus real-source image proofs and mesh-buffer extraction. Capacity tests independently check the dual, Hall boundary, integrality and comparative statics. The final verification report records the complete test count. In actual Chromium, Python and browser fingerprints agreed; camera changes preserved the manifest, cloning increased occurrence cost and marked the scene as an unendorsed draft, and restoration returned the original fingerprint. The viewer checks no institutional endorsement. Its JavaScript fingerprint is only an optional content-consistency aid.
Local timings. Eleven within-process repeats per size, in seeded shuffled order after warmup, measure roots and single-object disclosures for 2, 8, 32, 128 and 512 occurrences. Median root time increases from 0.483 to 14.065 ms; disclosure verification from 0.284 to 0.366 ms. The single-object proof grows from 4649 to 7129 bytes, including the mandatory source/evidence closure. A two-source committed debit takes a median 2.409 ms. Figure 9 shows observed ranges; CSV rows retain every measurement. Proof construction recomputes trees in this clarity-oriented implementation and is not an optimized production benchmark. The timings include validation and serialization, are susceptible to shared-host load, and should not be interpreted as calibrated cross-machine performance claims.
Figure 9. Local artifact timings on the stated M4 Pro. Shading shows observed minima and maxima, not confidence intervals. Proof size includes semantic closure, so a constant number of selected objects need not imply constant-size disclosures.
Revision transport at source-record scale
The revision experiment constructs a universe of 1,168,790 uniquely addressed source records: 3,374 original Göbekli cells, 191,154 nonempty cells in the original Cassidy workbook, and 974,262 fan-triangles from the twelve original Çatalhöyük OBJ files. A triangle digest binds the original file hash, fan index and three original vertex indices. The file digest transitively binds the referenced coordinate bytes; this benchmark does not open each triangle’s coordinates or replace the mesh extraction verifier. We deterministically sample 10,000, 100,000 and 1,000,000 distinct records from this universe, without replicating rows to inflate scale. These are storage/proof records, not independent people, artifacts or archaeological observations.
Each size adds 64 explicit local review records and a fixed 256-gate policy: two four-record conjunctions, their disjunction, and one registered-defeater eligibility guard per query. The query meaning is presence and digest of the admitted archived records under declared qualifications. Distinct objects are not thereby asserted to be historically interchangeable. Query count and rule size remain fixed while the authenticated registry grows, so this isolates registry-scale update and proof costs; it does not benchmark inference over one million independent historical claims.
Three separately signed revision branches change local qualifications or review inputs while preserving every original source file. First, 100/1,000/1,000 records outside the selected frontiers receive a qualification edit. Second, one selected-branch qualification changes for each of 16 queries, without changing the truth bits. Third, eight registered review defeaters become active. At all three sizes, the first case transports 64 original cuts; the second transports 48 original cuts and recovers the other 16 through old alternative cuts; the third transports 56 and has no avoiding positive cut for the remaining eight. Fresh current cuts independently verify the corresponding status for all 64 queries. Deleting a real changed opening from each of the nine deltas is rejected. These are controlled mutation experiments, not observations of source authors revising their publications.
At one million records, a full initial registry takes 20.618 s in the archived run. The three persistent updates take 19.603, 0.722 and 0.428 ms, rehashing 2,012, 256 and 32 nodes respectively. Their complete serialized deltas contain 632,290, 31,503 and 7,696 bytes; batched verification and transport of the 64 cuts takes 112.156, 127.708 and 127.534 ms. Creation of 64 fresh cuts from the already materialized current state takes 2.740, 2.728 and 2.401 ms. Thus transport is valuable as a verifiable relation between checkpoints; these observations do not show it to be faster than asking a trusted current-state holder for fresh cuts. Update, delta generation, public verification and fresh-cut construction are separately timed operations. Changed IDs in the first case are selected lexicographically outside the frontiers and can share tree paths. A single measurement per size/branch supports functional scaling inspection, not a confidence interval, random-update average or universal speedup.
A continuing research program and its limits
The immediate archaeological priority is an independent test of the Chogha Mish context hypothesis with matched recovery, dating and deposition. A second prospective hypothesis asks whether local pig provisioning can lag a crop/diet transition; unresolved Chinese site/phase matches and local ecological baselines currently prevent its test. These questions are not inferred from a verifier’s rejection rate. Qualified new observations can revise source admission and the comparison policy without overwriting the earlier evidence.
The mathematical model opens several concrete questions: minimum-cost avoiding cuts under shared subgraphs, exact or approximate transversal computation for cut families, compact representation of exponentially many witness alternatives; exact algorithms for restricted incidence families; stronger integral certificates; and scene configurations with reconstruction-school compatibility. Context can alter admissibility and budgets, but a learned or discretionary context classifier would introduce a new trusted boundary requiring separate evaluation. Revocation, key rotation, multi-ledger consistency and independent checkpoint witnesses are also future work rather than properties of the local prototype.
Figure 10. Actual source-record registry scaling and three controlled revision branches. Full build and incremental update are different operations. Delta bytes include all changed leaves and unchanged subtree frontier hashes; query count stays 64. At one million records, alternative cuts recover 16 touched certificates, while eight active defeaters prevent positive eligibility reuse.
The human hypothesis remains untested. The accompanying proposed protocol uses a design varying contextual framing, qualifications and an enforced export checker. Its primary belief outcome is erroneous attribution of institutional endorsement; successful export is a separate computational endpoint. Neither is inferred from rejecting a forged packet. The plan includes source recall, randomized stimuli, mixed-effects logistic analysis and pilot-based power simulation, without fabricating a sample size or participant results. Follow-ups at days 1, 7, 30 and 90 occur after immediate debriefing and therefore measure retention after correction, not uncorrected natural persistence. No participants have been recruited and no ethics approval is claimed.
Source cards, witness policies, renderer versions, signing identities and preregistered analysis decisions should form separate versioned registries. Archaeological and community expertise is needed to review qualifications and interpretive equivalence. A stricter verifier can preserve declared distinctions; it cannot determine ancient meaning or force a human to believe a caption. The derived results locate two distinct losses: erasing conjunctions can create unbounded apparent endorsement capacity, and deleting a shared guard can unboundedly inflate the certified-reuse threshold. Retaining occurrence, polarity and full-record frontier structure permits explicit feasibility and revision checks under the declared policy.
Artifact and AI assistance. The anc/ directory provides original licensed photographs, photogrammetry, losslessly wrapped artifact models and unchanged published workbooks, source cards, protocol/optimizer code and tests. Its offline runner regenerates raw results, plots and the image-scene, envelope and linked early-evidence workbenches. It has no runtime network requirement, trained weights or embedded PDF JavaScript. Generative AI assisted literature discovery, mathematical exploration, drafting, implementation and testing during preparation. This assistance is separate from the proposed non-learned experimental method. The named human authors are responsible for reviewing the arguments, sources, code and final submission; no independent cryptographic audit or peer review is claimed. Data and code are packaged as arXiv ancillary files with TeX sources [7].
Reproduction and exact witness certificates
From anc/, use Python 3.11 or later in a virtual environment. To reproduce all image, geometry, source-table, revision, capacity, attack and plotting stages, run:
python3 -m pip install -r requirements-lock.txt
python3 tools/reproduce.py --output ../scene-full-new --plots
Choose a fresh output directory: the runner refuses to overwrite prior results or an existing ledger. It first restores and verifies exact source members, then records every stage with plotting, raw measurements, structured reports and regenerated figures. Individual commands are documented in the README. Its deterministic test keys are intentionally public, and its fixed grants expire relative to the fixed test clock. They must never represent real institutional authority. tools/reproduce.py coordinates an isolated full run; dependency locks record the measured environment. The workbench can export a draft JSON scene for structural/source checking, but obtaining real signatures requires the relevant issuer’s informed authorization.
Checking the fractional certificate. Given rational , independently compute every bundle price, each minimum , and . If , (9) certifies using exact arithmetic. This does not certify an integral allocation exists. In the Fano fixture gives bound ; pairwise intersections provide the additional exact upper bound of one. The artifact retains all generated inputs and integer allocations so solver claims can be checked against the independently enumerated finite problem.
Completeness of disclosure. For each selected occurrence, the verifier opens every referenced root evidence record, follows all support IDs in the opened authenticated adjacency lists and recursively demands those records. An induction on DAG depth establishes that every reachable record is opened. Opening their source IDs then covers precisely the charged sources. This argument establishes closure of the admitted graph, not completeness of the archaeological record or an absent-edge claim about the world. A source-list summary lacks the authenticated connectives needed to state that induction.
Inspecting earlier evidence and revisions. The source XML readers keep original sheet/cell addresses and numerical lexemes. scenecrypt.gobekli_data, scenecrypt.social_analysis and scenecrypt.early_holocene_analysis rebuild the new source summaries. scenecrypt.revision_experiments independently rebuilds the 10,000, 100,000 and 1,000,000-record revision branches. A fresh-run atlas is generated from those run outputs, with original datasets read from the immutable input folders. The published record universe excludes the extra skull variant and uses only the twelve OBJ sources in the scaling experiment. Source snapshots, retrieval programs, licenses, unchanged-workbook hashes and decompressed-mesh hashes permit byte-level rechecking independently of the paper’s plots.
Regional source comparisons. The frozen regional release manifest verifies original exports, workbooks, licensed source snapshots and extracts before analysis. New stages regenerate Chogha Mish context/anatomical counts, the strict NISP cohort, exact scope certificates for both typed registries, shared-depth rounding witnesses, the independent partial-transfer audit, all thirteen original-OBJ codec comparisons and the regional companion. The complete source ZIP is strictly below the user-selected 50,000,000-byte cap. All result/display caches are regenerated in the external output directory; no source download is required. REGIONAL_RESEARCH_TR.md gives a Turkish research brief and explicit future falsification conditions. The scope_contrasts API binds an externally admitted question; its policy descriptions are not an automatic historical classification language. Original-source extraction and the current root remain independent verification inputs. The packaged public keys/seeds are experimental fixtures, and no runtime source download or trained model is required.
Language and external transfer. The default runner also executes the SQLite contract study and the complete bundled museum-language experiment with its source-integrity tests. The optional external temporal audit was executed for this release but requires two separately acquired, hash-pinned corpora: provide both --ancient-times-zip and --wikiwars-zip. Their native XML/labels are not redistributed while explicit original redistribution permission remains unresolved. The external evaluator and aggregate receipts are included; the core pipeline remains independently reproducible offline after installing its software/runtime dependencies. Exact JPEG reconstruction additionally requires the trusted POSIX libjxl0.11.2 tools.
References
- [1]Azim Afroozeh, Leonardo Kuffó, and Peter Boncz. ALP: Adaptive lossless floating-point compression. In Proceedings of the ACM on Management of Data, 2024. doi: 10.1145/3626717.DOI
- [2]I. P. Alaeva and A. Yu. Rassadnikov. Archaeozoology of the Zvyagino-4 settlement of the Alakul Culture in the context of different cultural deposits. Vestnik Arkheologii, Antropologii i Etnografii, no. 2(69), pp. 42–57, 2025. URL https://ipdn.ru/_private/a69/42-57.pdf. English title: article in Russian. Context and taxonomic qualifications, p.45; roof-material interpretation, p.53.DOI
- [3]Faruk Alpay and Taylan Alpay. Approval integrity and recovery in llm answer publication, 2026. URL https://arxiv.org/abs/2609.15726.
- [4]Faruk Alpay and Taylan Alpay. Provenance and authorization in autobiographical language models, 2026. URL https://arxiv.org/abs/2610.17935.
- [5]Aris Anagnostopoulos, Michael T. Goodrich, and Roberto Tamassia. Persistent authenticated dictionaries and their applications. In Information Security: ISC 2001, volume 2200 of Lecture Notes in Computer Science, pages 379–393. Springer, 2001. URL https://cs.brown.edu/cgc/stms/papers/isc2001.pdf.DOI
- [6]Apache Parquet. ALP: Adaptive lossless floating-point encoding in Apache Parquet, 2026. URL https://parquet.apache.org/blog/2026/09/22/alp-adaptive-lossless-floating-point-encoding-in-apache-parquet/. 22 September 2026.
- [7]arXiv. Ancillary files, 2026. URL https://info.arxiv.org/help/ancillary_files.html. Accessed 26 October 2026.
- [8]Giuseppe Ateniese, Daniel H. Chou, Breno de Medeiros, and Gene Tsudik. Sanitizable signatures. In Computer Security – ESORICS 2005, pages 159–177, 2005. doi: 10.1007/11555827_10.DOI
- [9]Levent Atici, Sarah W. Kansa, and Justin SE. Lev-Tov. Chogha Mish Fauna, 2010. URL https://opencontext.org/projects/497aded-0c2a-4c62-feef-9079fb09b1a5. Original named table export and overlapping Protoliterate export; CC BY 4.0; primary identifications by Jane Wheeler Pires-Ferreira.
- [10]Nuel D. Belnap. A useful four-valued logic. In J. Michael Dunn and George Epstein, editors, Modern Uses of Multiple-Valued Logic, pages 5–37. D. Reidel, 1977. doi: 10.1007/978-94-010-1161-7_2.DOI
- [11]Arnar Birgisson, Joe Gibbs Politz, Úlfar Erlingsson, Ankur Taly, Michael Vrable, and Mark Lentczner. Macaroons: Cookies with contextual caveats for decentralized authorization in the cloud. In Network and Distributed System Security Symposium, 2014. doi: 10.14722/ndss.2014.23212.DOI
- [12]Pierre Bourdieu. Social space and symbolic power. Sociological Theory, 7(1):14–25, 1989. doi: 10.2307/2026060.
- [13]Kevin D. Bowers, Lujo Bauer, Deepak Garg, Frank Pfenning, and Michael K. Reiter. Consumable credentials in logic-based access-control systems. In Network and Distributed System Security Symposium, 2007. URL https://people.mpi-sws.org/~dg/papers/NDSS07Final.pdf.
- [14]Lee G. Broderick. OXSTAD16 animal bone data, 2021. URL https://archaeologydataservice.ac.uk/archives/collections/view/object.cfm?object_id=1969741. Digital object 1969741; Oxford Archaeology (South) collection, DOI 10.5284/1088114. Created 24 September 2018. CC BY 4.0.
- [15]Lee G. Broderick. OXSTAD16 animal bone spreadsheet metadata, 2021. URL https://archaeologydataservice.ac.uk/catalogue/adsdata/arch-4099-1/dissemination/csv/documentation/OXSTAD16_animal_bone_spreadsheet_metadata.csv. Original catalogue field definitions and 605-record inventory; supporting documentation for ADS object 1969741, collection DOI 10.5284/1088114. CC BY 4.0.
- [16]Harry Buhrman and Ronald de Wolf. Complexity measures and decision tree complexity: A survey. Theoretical Computer Science, 288(1):21–43, 2002. doi: 10.1016/S0304-3975(01)00144-X. URL https://homepages.cwi.nl/~rdewolf/publ/qc/dectree.pdf.DOI
- [17]Lara M. Cassidy, Miles Russell, Martin Smith, Gabrielle Delbarre, Paul Cheetham, Harry Manley, Valeria Mattiangeli, Emily M. Breslin, Iseult Jackson, Maeve McCann, Harry Little, Ciarán G. O’Connor, Beth Heaslip, Daniel Lawson, Phillip Endicott, and Daniel G. Bradley. Continental influx and pervasive matrilocality in iron age britain. Nature, 637:1136–1142, 2025. doi: 10.1038/s41586-024-08409-6.DOI
- [18]Wendy W. Chapman, Will Bridewell, Paul Hanbury, Gregory F. Cooper, and Bruce G. Buchanan. A simple algorithm for identifying negated findings and diseases in discharge summaries. Journal of Biomedical Informatics, 34(5):301–310, 2001. doi: 10.1006/jbin.2001.1029.DOI
- [19]Wenhui Chen, Li Yan, Hoekyung Jung, and Huihuang Zhao. Dolphin: an adaptive lossless compression algorithm for oscillating floating-point time series. Journal of King Saud University Computer and Information Sciences, 38:218, 2026. doi: 10.1007/s44443-026-00553-5.DOI
- [20]Cleveland Museum of Art. Open access, 2026. URL https://www.clevelandart.org/open-access. Collection API and CC0 policy; accessed 26 October 2026.
- [21]Cleveland Museum of Art. Pendant, accession 1987.1, 2026. URL https://www.clevelandart.org/art/1987.1. Open Access collection record; accessed 26 October 2026.
- [22]Coalition for Content Provenance and Authenticity. Content Credentials: C2PA Technical Specification, Version 2.3, December 2025. URL https://spec.c2pa.org/specifications/specifications/2.3/specs/C2PA_Specification.html. Accessed 26 October 2026.
- [23]Adnan Darwiche and Pierre Marquis. A knowledge compilation map. Journal of Artificial Intelligence Research, 17:229–264, 2002. doi: 10.1613/jair.989. URL https://ai.dmi.unibas.ch/research/reading_group/darwiche-marquis-jair2002.pdf.DOI
- [24]Laura Dietrich, Julia Meister, Oliver Dietrich, Jens Notroff, Janika Kiep, Julia Heeb, André Beuger, and Brigitta Schütt. Cereal processing at early neolithic göbekli tepe, southeastern turkey. PLOS ONE, 14(5):e0215214, 2019. doi: 10.1371/journal.pone.0215214.DOI
- [25]Thaddeus Dryja. Utreexo: A dynamic hash-based accumulator optimized for the bitcoin UTXO set. Cryptology ePrint Archive, Paper 2019/611, 2019. URL https://eprint.iacr.org/2019/611.
- [26]Geoff Emberling, Rachael J. Dann, Abbas Sidahmed Mohamed-Ali, Rikke Therkildsen, Sarah M. Duffy, Janice W. Yellin, Sebastian Anstis, and Tim Karberg. In a royal cemetery of kush: Archaeological investigations at el kurru, northern sudan, 2014–15. Sudan & Nubia, 19:54–70, 2015. URL https://www.sudarchs.org.uk/wp-content/uploads/2019/03/SARS_SN19_Emberling_et_al.pdf.
- [27]Robert M. Entman. Framing: Toward clarification of a fractured paradigm. Journal of Communication, 43(4):51–58, 1993. doi: 10.1111/j.1460-2466.1993.tb01304.x.DOI
- [28]Lisa K. Fazio, David G. Rand, and Gordon Pennycook. Repetition increases perceived truth equally for plausible and implausible statements. Psychonomic Bulletin & Review, 26:1705–1710, 2019. doi: 10.3758/s13423-019-01651-4.DOI
- [29]Lester R. Ford, Jr. and Delbert R. Fulkerson. Maximal flow through a network. Canadian Journal of Mathematics, 8:399–404, 1956. doi: 10.4153/CJM-1956-045-5.DOI
- [30]Jane S. Gaastra, Dan Lawrence, and Valentina Tumolo. Provisioning Urbanism: A Comparative Urban-Rural Zooarchaeology of Ancient Southwest Asia. Antiquity, 98(398):363–379, 2024. doi: 10.15184/aqy.2024.24. URL https://www.cambridge.org/core/journals/antiquity/article/provisioning-urbanism-a-comparative-urbanrural-zooarchaeology-of-ancient-southwest-asia/C4E4D2A61C925326C7E88D945ABAD814.DOI
- [31]David Gale, Harold W. Kuhn, and Albert W. Tucker. Linear programming and the theory of games. In Tjalling C. Koopmans, editor, Activity Analysis of Production and Allocation, volume 13 of Cowles Commission Monograph, chapter 19, pages 317–329. John Wiley & Sons, 1951. URL https://cowles.yale.edu/sites/default/files/2022-09/m13-all.pdf.
- [32]Floris Geerts and Antonella Poggi. On database query languages for K-relations. Journal of Applied Logic, 8(2):173–185, 2010. doi: 10.1016/j.jal.2009.09.001. URL https://adrem.uantwerpen.be/bibrem/pubs/GeertsA.pdf.DOI
- [33]Ryan Giordano, Rachael Meager, and Tamara Broderick. An Automatic Finite-sample Robustness Metric: When Can Dropping a Little Data Change Conclusions? Part II: Theory and Intuition. Philosophical Transactions of the Royal Society A, 384(2321):202402614, 2026. doi: 10.1098/rsta.2024.0614. URL https://royalsocietypublishing.org/doi/10.1098/rsta.2024.0614.DOI
- [34]Jean-Yves Girard. Linear logic. Theoretical Computer Science, 50(1):1–101, 1987. doi: 10.1016/0304-3975(87)90045-4.DOI
- [35]Erich Grädel and Val Tannen. Provenance analysis for logic and games. Moscow Journal of Combinatorics and Number Theory, 9(3):203–228, 2020. doi: 10.2140/moscow.2020.9.203. URL https://arxiv.org/abs/1907.08470.
- [36]Todd J. Green, Grigoris Karvounarakis, and Val Tannen. Provenance semirings. In Proceedings of the Twenty-Sixth ACM SIGACT-SIGMOD-SIGART Symposium on Principles of Database Systems, pages 31–40, 2007. doi: 10.1145/1265530.1265535.DOI
- [37]Philip Hall. On representatives of subsets. Journal of the London Mathematical Society, s1-10 (1):26–30, 1935. doi: 10.1112/jlms/s1-10.37.26.DOI
- [38]Stuart Hall. Encoding and decoding in the television discourse. CCCS Stencilled Occasional Paper 7, University of Birmingham, 1973. URL https://epapers.bham.ac.uk/id/eprint/2962/.
- [39]Jerry R. Hobbs, Douglas Appelt, John Bear, David Israel, Megumi Kameyama, and Mabry Tyson. FASTUS: A system for extracting information from text. In Human Language Technology: Proceedings of a Workshop Held at Plainsboro, New Jersey, March 21–24, 1993, 1993. URL https://aclanthology.org/H93-1026/.
- [40]Carl I. Hovland and Walter Weiss. The influence of source credibility on communication effectiveness. Public Opinion Quarterly, 15(4):635–650, 1951. doi: 10.1086/266350.DOI
- [41]Martin Isenburg, Peter Lindstrom, and Jack Snoeyink. Lossless compression of predicted floating-point geometry. Computer-Aided Design, 37(8):869–877, 2005. doi: 10.1016/j.cad.2004.09.015.DOI
- [42]Marcia K. Johnson, Shahin Hashtroudi, and D. Stephen Lindsay. Source monitoring. Psychological Bulletin, 114(1):3–28, 1993. doi: 10.1037/0033-2909.114.1.3.pubmed.ncbi.nlm.nih.gov/8346328
- [43]Simon Josefsson and Ilari Liusvaara. Edwards-curve digital signature algorithm (eddsa). Technical Report RFC 8032, RFC Editor, 2017.DOI
- [44]JPEG XL Project. libjxl0.11.2: Jpeg reconstruction and decoder documentation, 2026. URL https://github.com/libjxl/libjxl/releases/tag/v0.11.2. Released10 February2026. Version-pinned native runtime; exact original JPEG reconstruction is the default for reconstructed JPEG output. Documentation checked27 October2026.
- [45]Richard M. Karp. Reducibility among combinatorial problems. In Raymond E. Miller and James W. Thatcher, editors, Complexity of Computer Computations, pages 85–103. Plenum Press, 1972. doi: 10.1007/978-1-4684-2001-2_9.DOI
- [46]Zachary A. Kissel. Policy-based redactable signatures. Cryptology ePrint Archive, Report 2022/1485, 2022. URL https://eprint.iacr.org/2022/1485.
- [47]David Korn, Joshua MacDonald, Jeffrey Mogul, and Kiem-Phong Vo. The VCDIFF generic differencing and compression data format. RFC3284, 2002. URL https://www.rfc-editor.org/rfc/rfc3284.html.DOI
- [48]Abhiram Kothapalli, Srinath Setty, and Ioanna Tzialla. Nova: Recursive zero-knowledge arguments from folding schemes. In Advances in Cryptology—CRYPTO 2022, volume 13510 of Lecture Notes in Computer Science, pages 359–388. Springer, 2022. doi: 10.1007/978-3-031-15985-5_13. URL https://eprint.iacr.org/2021/370.DOI
- [49]John Kunze, Justin Littman, Liz Madden, John Scancella, and Chris Adams. The BagIt file packaging format (v1.0). RFC8493, 2018. URL https://www.rfc-editor.org/rfc/rfc8493.html.DOI
- [50]Ben Laurie, Eran Messeri, and Rob Stradling. Certificate transparency version 2.0. Technical Report RFC 9162, RFC Editor, 2021.
- [51]Justin Lev-Tov, Sarah W. Kansa, Levent Atici, and Jane C. Wheeler. New Light on Faunal Remains from Chogha Mish, Iran. In Justin Lev-Tov, Paula Hesse, and Allan Gilbert, editors, The Wide Lens in Archaeology: Honoring Brian Hesse’s Contributions to Anthropological Archaeology, volume 2 of Archaeology, pages 443–475. Lockwood Press, Atlanta, 2017. doi: 10.5913/2017956.ch19. URL https://lockwoodpressonline.com/index.php/ebooks/catalog/book/40.
- [52]Feifei Li, Marios Hadjieleftheriou, George Kollios, and Leonid Reyzin. Authenticated Index Structures for Aggregation Queries. ACM Transactions on Information and System Security, 13(4):32:1–32:30, 2010. doi: 10.1145/1880022.1880026. URL https://www.cs.bu.edu/~reyzin/papers/auth-db-agg.pdf.
- [53]Antonia Lourentzaki, Maria Papadopoulou, and Christophe Roche. Modeling an archaeological site – göbekli tepe v.1.0, 2024. URL https://zenodo.org/records/13370343. Unmodified RDF source; DOI registry rights: CC BY-NC-ND 4.0.
- [54]L. Luque, A. Antonini, G. Ferracutti, R. G. Raidou, E. Bjerg, S. Castro, E. Gröler, and M. L. Ganuza. LLUVIA: A level-of-detail approach to uncertainty-aware visual analytics for geosciences. In EuroVis 2026 – Posters and Demos, 2026. doi: 10.2312/evpd.20261002. URL https://diglib.eg.org/items/b5600499-7155-4ef7-b864-7b6a8577a7ee.DOI
- [55]Elena Maini, Laura Strolin, and Matteo Cianfoni. Sheep and Goat Variability along the Edges of the Al-Hajar Mountains in the Prehistory of South-east Arabia. Open Quaternary, 11:7, 2025. doi: 10.5334/oq.149. URL https://reference-global.com/article/10.5334/oq.149.DOI
- [56]Charles Martel, Glen Nuckolls, Premkumar Devanbu, Michael Gertz, April Kwong, and Stuart G. Stubblebine. A General Model for Authenticated Data Structures. Algorithmica, 39(1):21–41, 2004. doi: 10.1007/s00453-003-1076-8. URL https://web.cs.ucdavis.edu/~martel/main/paperlist.html.DOI
- [57]Pawel Mazur and Robert Dale. WikiWars: A new corpus for research on temporal expressions. In Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 913–922. Association for Computational Linguistics, 2010. URL https://aclanthology.org/D10-1089/.
- [58]Ralph C. Merkle. Protocols for public key cryptosystems. In 1980 IEEE Symposium on Security and Privacy, pages 122–134, 1980. doi: 10.1109/SP.1980.10006. URL https://www.ralphmerkle.com/papers/Protocols.pdf.DOI
- [59]Jamie Morgenstern, Deepak Garg, and Frank Pfenning. A proof-carrying file system with revocable and use-once certificates. In Security and Trust Management: STM 2011, volume 7170 of Lecture Notes in Computer Science, pages 40–55. Springer, 2011. doi: 10.1007/978-3-642-29963-6_5. URL https://www.cs.cmu.edu/~fp/papers/stm11.pdf.DOI
- [60]G. M. Morton. A computer oriented geodetic data base; and a new technique in file sequencing. Technical report, International Business Machines Co. Ltd., Ottawa, Ontario, Canada, March 1966. URL https://web.archive.org/web/20190125020453id_/https://domino.research.ibm.com/library/cyberdig.nsf/papers/ODABF9473B9C86D48257798000566A39/$File/Morton1966.pdf. Dated 1 March 1966; archived original report.
- [61]Noam Nisan. CREW PRAMs and decision trees. SIAM Journal on Computing, 20(6):999–1007, 1991. doi: 10.1137/0220062. URL https://epubs.siam.org/doi/pdf/10.1137/0220062.
- [62]Oxford Common File Layout Community. OCFL specification1.1, 2024. URL https://ocfl.io/1.1/spec/. Specification1.1, updated7 November2024.
- [63]Hwee Hwa Pang and Kian-Lee Tan. Verifying completeness of relational query answers from online servers. ACM Transactions on Information and System Security, 11(2), 2008. doi: 10.1145/1330332.1330337. URL https://ink.library.smu.edu.sg/sis_research/778/.DOI
- [64]Sara Perry, Maya Marilena Lara, and Gianpaolo Palma. 3d models of catalhoyuk artifacts, 2020. URL https://zenodo.org/records/3667558. Original source printing-model deposit; CC BY 4.0.DOI
- [65]W. M. Flinders Petrie. The Pyramids and Temples of Gizeh. Field & Tuer, London, 1883. URL https://www.gizapyramids.org/pdf_library/petrie_gizeh.pdf.
- [66]Max Price. The Swine of Sumer (and Beyond): The Place of Pigs in Mesopotamian Cultures from the Uruk Through Old Babylonian Periods. Environmental Archaeology, 2026. doi: 10.1080/14614103.2026.2651646. URL https://www.tandfonline.com/doi/full/10.1080/14614103.2026.2651646.DOI
- [67]Max Price and Noa Corcoran-Tadd. Livestock and the Transition to the Tributary Mode of Production. Cambridge Archaeological Journal, 36:76–91, 2026. doi: 10.1017/S0959774325100140. URL https://doi.org/10.1017/S0959774325100140. First published online 4 September 2025; 2026 journal issue.
- [68]Maria Diletta Pubblico and National Museums Liverpool. 3d model of a cat mummy from thebes, egypt, kept at the world museum, liverpool, uk (56.22.225), 2026. URL https://zenodo.org/records/21760633. Version 1; CC BY 4.0; SEAMS project.DOI
- [69]Alex Rassadnikov. Animal bones from the South Urals and Northern Kazakhstan, 2026. URL https://zenodo.org/records/22714190. Published 11 September 2026. CC BY 4.0; author name as deposited.DOI
- [70]Anders Rundgren, Bradley Jordan, and Samuel Erdtman. Json canonicalization scheme (jcs). Technical Report RFC 8785, RFC Editor, 2020.DOI
- [71]Gabriele Russo et al. Lebanese Faunal Legacy Collections Uncover 400,000 Years of Woodland Resilience and Distinct Hominin Ecological Strategies. Communications Earth & Environment, 7:706, 2026. doi: 10.1038/s43247-026-03926-w. URL https://www.nature.com/articles/s43247-026-03926-w.DOI
- [72]Melina Seabrook, Jesse Wolfhagen, and Suzanne E. Pilaar Birch. Large-scale Carbon and Nitrogen Isotopic Investigations into Caprine Diet and Management at Tepe Yahya, Southeastern Iran. PLOS ONE, 20(9):e0332662, 2025. doi: 10.1371/journal.pone.0332662. URL https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0332662.DOI
- [73]Aryak Sen, Silviu Maniu, and Pierre Senellart. ProvSQL: A general system for keeping track of the provenance and probability of data. In IEEE International Conference on Data Engineering (ICDE), 2026. URL https://pierre.senellart.com/publications/sen2026provsql.pdf.arxiv.org/abs/2504.12058
- [74]Claude E. Shannon. A mathematical theory of communication. Bell System Technical Journal, 27:379–423, 623–656, 1948. doi: 10.1002/j.1538-7305.1948.tb01338.x.DOI
- [75]SQLite Project. Built-in aggregate functions, 2026. URL https://sqlite.org/lang_aggfunc.html. Official documentation, accessed 27 October 2026.
- [76]SQLite Project. EXPLAIN QUERY PLAN, 2026. URL https://sqlite.org/eqp.html. Official documentation, accessed 27 October 2026; plan output is diagnostic and version-dependent.
- [77]Jannik Strötgen and Michael Gertz. HeidelTime: High quality rule-based extraction and normalization of temporal expressions. In Proceedings of the 5th International Workshop on Semantic Evaluation, pages 321–324. Association for Computational Linguistics, 2010. URL https://aclanthology.org/S10-1071/.
- [78]Jannik Strötgen, Thomas Bögel, Julian Zell, Ayser Armiti, Tran Van Canh, and Michael Gertz. Extending HeidelTime for temporal expressions referring to historic dates. In Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC’14), pages 2390–2397. European Language Resources Association, 2014. URL https://aclanthology.org/L14-1655/.DOI
- [79]Steve Teague, Edward Bidulph, and Carl Champness. Saxon, jewish, later medieval and post-medieval occupation on St Aldates and Queen Street, Oxford. Publication report OA reference 6523, issue 2, Oxford Archaeology, 2019. URL https://archaeologydataservice.ac.uk/catalogue/adsdata/arch-841-1/dissemination/pdf/oxfordar1-425870_1.pdf. Revised 8 February 2019. Site code OXSTAD16.DOI
- [80]The London Charter. The london charter for the computer-based visualisation of cultural heritage, 2009. URL https://londoncharter.org/principles.html. Version 2.1.
- [81]The Metropolitan Museum of Art. Statue of a kneeling kushite king, accession 2002.8, 2026. URL https://www.metmuseum.org/art/collection/search/545027. Public-domain image and collection record; accessed 26 October 2026.
- [82]The Metropolitan Museum of Art. The metropolitan museum of art collection api, 2026. URL https://metmuseum.github.io/. Open Access API documentation; accessed 26 October 2026.
- [83]Angela Trentacoste. Animal remains from Forcello, 2015. URL https://opencontext.org/projects/bdc6cb19-e739-404b-8b4f-eeaae3b8aae. Released 21 October 2015. CC BY 4.0.DOI
- [84]Angela Claire Trentacoste. The Etruscans and Their Animals: The Zooarchaeology of Forcello di Bagnolo San Vito (Mantova). PhD thesis, University of Sheffield, 2014. URL https://etheses.whiterose.ac.uk/id/eprint/6553/. Context and phase documentation, pp.72–73; context codes, Table 3.3.2, p.261.
- [85]Shyama Vermeersch, Alexander Ahrens, Dominik Bonatz, Nathalie Kallas, Katharina Schmidt, and Elizabeth Stroud. New Insights into the Diet and Economy of the Central and Northern Transjordan during the Iron Age: Zooarchaeology and Stable Isotopes Analysis of Recent Excavations (Amman Citadel, Tell Zira’a, Tell Bleibil, and Tell Ushayer). Journal of Archaeological Science: Reports, 74:105990, 2026. doi: 10.1016/j.jasrep.2026.105990. URL https://doi.org/10.1016/j.jasrep.2026.105990. Online 4 August 2026; October 2026 issue; source-derived table CC BY-NC 4.0.
- [86]Alexander Wiebel, Sarah M. Brender, and Hella Eckardt. Toward comparative visual analysis of ancient roman millefiori artifacts. In EuroVis 2026 – Posters and Demos, 2026. doi: 10.2312/evpd.20261003. URL https://diglib.eg.org/items/ae6096dc-80f1-4471-ab39-d5f6db6448c7.DOI
- [87]Eren Yüncü et al. Female lineages and changing kinship patterns in neolithic çatalhöyük. Science, 388(6754):eadr2915, 2025. doi: 10.1126/science.adr2915. URL https://escholarship.org/uc/item/5rr9k7gg.DOI
- [88]Eren Yüncü et al. Supplementary data for: Female lineages and changing kinship patterns in neolithic çatalhöyük, 2026. URL https://zenodo.org/records/19297275. Version 4, 30 March 2026; source correction notice; sequence archives not bundled.
- [89]Borut Žalik, Domen Mongus, Blaž Repnik, and David Podgorelec. The contribution of removing reconstructable triangles in the lossless compression of a triangular mesh cad model. Applied Sciences, 16(11):5231, 2026. doi: 10.3390/app16115231.DOI
- [90]Meng Zhang, Dan Lawrence, and Andrew R. Millard. Isotopic Evidence for Changing Diet and Agriculture in China from the Neolithic to the Early Han Period (10–2 ka BP). Quaternary Science Reviews, 373:109725, 2026. doi: 10.1016/j.quascirev.2025.109725. URL https://durham-repository.worktribe.com/output/4918670. Published online 8 December 2025; issue 1 February 2026.DOI
- [91]Quan Zhang, Yanfeng Hou, Xinwei Li, Amy Styring, and Julia Lee-Thorp. Stable Isotopes Reveal Intensive Pig Husbandry Practices in the Middle Yellow River Region by the Yangshao Period (7000–5000 BP). PLOS ONE, 16(10):e0257524, 2021. doi: 10.1371/journal.pone.0257524. URL https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0257524.DOI









