TOMORI: A Tool-Orchestrated Mixing Operator for REAPER Interaction

Introduction

Computer-use agents increasingly connect natural language to actions in desktop and web applications [19]. Generated code provides a flexible interface for such interaction: CodeAct expresses actions as executable programs [17], and BrowserGym supports programmable browser interaction [8]. In 3D design, a user describes a desired scene, and Blender agents write and execute Python scripts to assemble objects and refine the rendered result [6, 7].

Related Work

Computer use and code-based software control. OSWorld evaluates agents in real desktop applications [19], while CodeAct uses executable Python actions and execution feedback for multi-turn tool use [17]. BrowserGym supports browser interaction through programmable actions and structured observations ([8]). In creative software, SceneCraft synthesizes Blender scenes through scripts and visual refinement ([6]), and BlenderAlchemy edits existing Blender states through visually guided program search ([7]). TOMORI applies code-based control to multitrack mixing through generated Lua programs and native project-state feedback.

Language-guided audio editing and DAW agents. AUDIT and Instruct-MusicGen perform instruction-conditioned audio transformations ([18, 20]). WavCraft orchestrates specialist models ([12]), while Text2FX and LLM2Fx-Tools connect language to effect parameters and chains ([2, 4]). DAWZY generates state-aware, reversible REAPER edits and evaluates reliability and user experience ([5]). RIME generates rule-based editing tasks for PO-EMS and evaluates edit graphs, parameters, and final audio, including FAD, KAD, and MERT displacement similarity ([13]).

TOMORI-BENCH studies a complementary setting: model-generated Lua code executes in native REAPER, and the delivered editable project is jointly assessed with its rendered audio against human-reviewed reference transitions. The benchmarks differ in execution environment, operators, reference construction, and delivery requirements; transferring between them requires an explicit mapping of these protocols. Our experiments compare models and editing policies within a shared native-DAW harness.

Automatic mixing and reference construction. Prior work learns or searches mixing parameters ([14, 15, 16]), while GRAFX-Prune recovers processing graphs from multitracks and finished mixes ([9 b]). We use graph inference to construct executable REAPER references, followed by expert refinement. This construction is separate from agent inference; the target mix and future reference states remain hidden.

Tomori

TOMORI connects a language model to the REAPER DAW through a code-based mixing-agent harness. The harness provides project observations and executes model-generated Lua programs as DAW editing actions. Its architecture and interaction workflow are illustrated in Figure 1.

Plugin Architecture

TOMORI utilizes DeepSeek Harness (DSH), a plugin-based agent runtime [3]. DSH provides model interaction, tool calling, and dynamic plugin loading. We develop custom plugins to connect this runtime to REAPER.

The plugins integrate a native REAPER bridge for project-state inspection and Lua execution. They also manage rendering and return execution results to the model. Through these interfaces, the model reads the current project, selects operations, and executes code within the DAW. Audio analysis is available through shell tools and model-written programs. The resulting observations are returned as text, allowing the same tool-calling model to use project information and signal measurements.

Project and Audio Observation

Mixing engineers adjust parameters through repeated listening and inspection of plugin displays, including spectrum analyzers. To approximate this process, TOMORI provides programmatic access to project state and audio measurements.

Our plugins use REAPER’s public APIs to retrieve tracks, routing, effect order, and parameter names and values. The model can also search REAPER documentation to identify the relevant controls. For signal analysis, it writes code or calls FFmpeg to measure levels in selected frequency bands and other quantities relevant to the request. Measurement is optional: the model chooses whether to analyze audio and which quantities to compute. It receives the numerical results as text, without direct audio input.

Code-Based Editing Workflow

Given a mixing request and recent context, the model generates Lua code to modify the current project. The code can insert or remove effects, adjust controls, change routing or automation, and render audio. Each execution returns status, results, and errors. The model can then inspect the updated state and revise its edits before delivery (Appendix § 11.§ 1).

This workflow supports several observation and editing rounds within one request. In the Everytime session, for example, the agent compared vocal and accompaniment band energies, adjusted the accompaniment EQ, and rendered again to measure the change. Another exchange compared sidechain-processed and bypassed outputs while inspecting routing. The final output consists of an editable REAPER project and rendered audio.

Tomori-Bench

Tomori-Bench evaluates language-guided, code-based mixing end to end, from native DAW execution to the delivered project and its rendered audio. It contains 110 edits, each expressed in three instruction forms, and measures agreement in project structure, parameter values, and rendered audio.

Task Representation

We represent a mixing task as a transition between REAPER project states. Let $X_0, \ldots, X_N$ denote the states along a sequence of $N$ reference edits, and $A_i$ the edit taking $X_{i-1}$ to $X_i$. Each state contains source placement, track identity, routing, effect-chain structure, and exposed controls. A case in instruction form $l$ is defined as

where $u_i^l$ is the current instruction and $H_i^l$ contains up to three preceding requests in the same form. The model receives the before-state and requests, while the reference edit and after-state remain hidden. It generates Lua editing code and delivers the resulting project $\hat{X}_i$.

Each case is evaluated independently from its reference before-state. Multiple tool calls and revisions are permitted within the request. Along with normal edits, the sequence includes overshoots and subsequent repairs to assess execution under different project states and adjustment directions.

Reference Construction

Task coverage and instruction styles are informed by Everytime-72, a continuous eight-stem production session containing 72 human requests and corresponding project snapshots. The requests include balance, timbre, spatial placement, automation, processing-chain revision, and delivery recovery. Later requests often correct earlier edits. The request–project correspondence has been audited, and redacted examples are provided in Appendix § 10.

To construct executable references, we use Phoenix_ScotchMorris from MedleyDB, which provides original multitracks and a finished mix ([1]). GRAFX-Prune optimizes effect parameters and prunes processing chains to estimate a graph that approximately reconstructs the mix ([9 b]). We map its effects, order, and routing to native REAPER plugins, rebuild the project, and further fit the parameters. A professional mixing engineer then auditions, refines, and approves the endpoint.

The approved mix is decomposed into small edits, each affecting one or two effects or track controls. These edits include effect insertion or removal, parameter adjustment, and correction of an excessive adjustment. The sequence contains 72 normal edits, 19 overshoots, and 19 repairs. We execute the complete sequence cumulatively to verify that the final project and rendered audio reproduce the approved endpoint. Construction and ordering details are provided in Appendix § 12.

Instruction Forms

Each reference edit is paired with three instructions that vary the implementation detail supplied to the model. We denote these instruction forms by IF0, IF1, and IF2:

• IF0: Specifies the target object, operation or control, adjustment direction, and numerical increments where appropriate.

• IF1: Retains the target object and effect category, with approximate expressions of the adjustment direction or audible problem.

• IF2: Describes the intended sound of the target object, leaving operation selection and parameter settings to the model.

From IF0 to IF2, the forms represent requests from users with decreasing mixing expertise, moving from precise control instructions to descriptions of the intended sound. The instructions therefore become progressively more abstract. For language review, the mixing client checks whether each instruction accurately describes the mixing need addressed by its reference edit. Table 1 illustrates the three forms for one repair edit. All forms share the same reference transition, although IF1 and IF2 do not uniquely specify its numerical settings.

Form	Step 2: English translations of the archived Chinese instructions
IF0	On the left main microphone’s ReaEQ, raise band 2 (around 950 Hz) by 1.0 dB.
IF1	The left main microphone’s EQ was cut too much around 950 Hz; bring a little back.
IF2	The left main microphone sounds too hollow in the middle; restore some body.

Evaluation Metrics

Our end-to-end evaluation assesses reference agreement in both the delivered native project and its rendered audio. Project metrics evaluate the realized project edits and parameter values. Audio metrics evaluate the direction and magnitude of the resulting sound change. During inference, the agent receives execution status, project readback, and optional numerical audio measurements as textual feedback; final-artifact scoring evaluates the outcome of this workflow.

Project Metrics Graph F1. We extract an edit graph by comparing the project before and after execution. Nodes represent effect insertions and removals or control changes; links represent changes to the processing chain. A node matches when its kind, exact operation or control, target track, and all required parameter directions agree with the reference. Graph F1 is defined as

where $M_N$ and $M_E$ count matched nodes and links, and $N_r$, $N_m$, $E_r$, $E_m$ count reference and model nodes and links. Multiset matching penalizes omitted and additional edits. A required parameter that is reversed, unchanged, invalid, or unmatched prevents its node from matching. Links retain track identity and effect order. Scores are macro-averaged over cases using the joint node/direction rule (v3). Implementation details are provided in Appendix § 11.§ 3.

Parameter Precision. Model and reference settings are matched by track, effect, and control. We compute mean absolute error (MAE) over finite matched values and tolerance accuracy over all required parameters (Schaffer and Singh, [13]). The latter counts valid values with absolute error at most $\epsilon$; missing or invalid settings receive no credit. FX and pan use normalized $[0, 1]$ values, while track gain is reported in dB. Lower MAE and higher tolerance accuracy indicate closer reference agreement. Formulas are provided in Appendix § 11.§ 4.

Audio Metrics Reference-Error Reduction. Let $b_i$, $y_i$, and $\hat{y}_i$ denote audio from the before-state, reference after-state, and model after-state. With rendering denoted by $R$, we define

Here $L$ is multi-resolution short-time Fourier transform loss (MR-STFT loss), also used in graph inference (Lee et al., [10]). Audio is loudness-normalized before scoring (Appendix § 11.§ 2). Reference-error reduction is $c_i = \max(0, 1-\ell_i/\ell_i^{\mathrm{noop}})$, reported as a percentage. A value of 80% indicates that editing removes four-fifths of the initial reference error; 0% indicates unchanged or increased error.

Edit Progress. We measure the reference and model changes from the common before-state, $a_i = L(b_i,y_i)$ and $m_i = L(b_i,\hat{y}_i)$. Treating $a_i$, $m_i$, and $\ell_i$ as triangle sides gives

which estimates the model change along the reference direction relative to the reference magnitude. A value of one matches that magnitude, zero indicates no progress, and negative values indicate reversal. Edits in range counts cases with $0.5 \le\rho_i \le1.5$, allowing half to one-and-a-half times the reference change. Excess change counts cases with $\rho_i > 1.5$. Six steps with negligible reference change ($a_i < 10^{-3}$) are excluded, leaving 104 cases per form.

Representation-Space Progress. Since MR-STFT loss is not a Euclidean distance, its projection is approximate. We also compute progress in a learned feature space. Frozen MERT features (Li et al., [11]) are pooled into an audio vector $E(x)$. Following RIME (Schaffer and Singh, [13]), we compare displacements from the common before-state:

This projection measures the component of model change along the reference displacement, normalized by the reference magnitude. We additionally report RIME’s displacement cosine, which measures direction alone. Preprocessing details are provided in Appendix § 11.§ 4.

Experiments

The experiments assess instruction following, rendered-audio outcomes, and model differences within the same Tomori harness. We additionally compare editing and audio-access policies to examine the contribution of the full workflow.

Settings

We select kimi-k3, claude-fable-5, and deepseek-v4-pro for the primary evaluation. Each model completes 330 cases across the three instruction forms, yielding 990 cases in total. The tasks cover EQ, gain, pan, gate, compression, reverb, delay, and width processing. Each case loads the reference before-state in an independent session with up to three recent requests.

Each delivery contains an editable project and two 48 kHz stereo WAVs covering 15–30 and 55–75 seconds. Scores use selected final valid deliveries, including recovery after failed attempts. Metrics are macro-averaged over cases, except tolerance accuracy, which aggregates required controls.

For supplementary comparisons, we evaluate deepseek-v4-flash and kimi-k2.6 on 36 reference steps under IF0 and IF2, with 72 cases per model. The primary models’ corresponding deliveries are reused. Subset selection and paired analysis are described in Appendix F.

Instruction-Following Evaluation

This experiment evaluates how instruction detail affects the execution of reference edits. The project and audio results for the three forms are presented in Table 2.

	Graph F1 ↑	Param. MAE ↓	Error reduction ↑ (%)	Edits in range ↑ MR-STFT (%)	Edits in range ↑ MERT (%)	Excess change ↓ (%)
kimi-k3
IF0	1.0000	0.0352	83.8	95.2	95.2	2.9
IF1	0.9508	0.0811	47.1	57.7	56.7	21.2
IF2	0.8011	0.1112	33.9	49.0	45.2	28.8
claude-fable-5
IF0	0.9841	0.1441	76.0	85.6	85.6	3.8
IF1	0.9305	0.3002	42.6	51.9	52.9	26.0
IF2	0.7777	0.1189	32.8	46.2	42.3	31.7
deepseek-v4-pro
IF0	1.0000	0.0062	85.3	97.1	97.1	0.0
IF1	0.9614	0.0789	48.5	59.6	60.6	16.3
IF2	0.7726	0.1168	32.0	41.3	45.2	30.8

Regarding parameter precision, all three models achieve high tolerance accuracy under IF0. Under IF2, only about one-fifth of normalized parameters match within 0.01. Graph F1 also decreases, indicating that abstract requests affect both operation selection and numerical execution.

Project readback identifies operation substitution and additional processing as common discrepancies. For example, Kimi adjusts pan in IF2 step 1, where the reference adjusts EQ on the same track. Among Kimi’s 40 IF2 cases with imperfect Graph F1, 24 retain all reference nodes but add further edits. Additional processing also accounts for nearly half of Claude’s imperfect cases (Table 8). Chain-link agreement remains high, while node agreement declines.

Reversed adjustments are less frequent than unchanged required parameters. Each model reverses at most four required parameters per instruction form. However, explicit instructions do not eliminate large errors: in IF0 step 99, Claude sets compressor attack to 250 ms when the requested change is from 3 ms to 0.5 ms. These results indicate that high operation agreement alone does not ensure precise parameter execution.

Rendered-Audio Evaluation

This experiment compares each rendered output with its reference after-state and unchanged before-state. Reference-error reduction and edit-progress scores are reported in Table 2.

All three models generally move the sound toward the reference. However, mean error reduction decreases from roughly four-fifths under IF0 to one-third under IF2. Many abstract edits change the sound in the reference direction but with excessive or insufficient magnitude. Among IF2 edits that increase reference error, overshooting is the predominant failure.

The MR-STFT and MERT projections show a consistent trend. Both classify most IF0 edits as within the specified magnitude tolerance, compared with fewer than half under IF2. They assign the same progress category to a large majority of cases, although their model rankings differ under IF2. A stricter MR-STFT criterion, allowing at most 25% deviation from the reference change, reduces the share without changing the trend.

DeepSeek achieves the highest error reduction under IF0 and IF1, while Kimi achieves the highest error reduction under IF2. Kimi also has the highest IF2 Graph F1. Project and audio metrics provide complementary assessments: the former evaluates required edits, while the latter reflects their combined sound, including additional processing.

Ablation Studies

This experiment assesses the editing workflow and access to audio measurements. All conditions use DeepSeek v4 Pro and are paired with its full-workflow deliveries on the same cases. The results are presented in Table 3; detailed policies and audits are provided in Appendix G.

Condition	Form	Graph F1 ↑	Param. match ↑ (%)	Error reduction ↑ (%)	MERT dir. ↑
Same 36 reference steps
Full workflow	IF0	1.000	97.7	85.4	0.894
Single submission (S)	IF0	0.875	72.1	61.3	0.711
Full workflow	IF2	0.701	20.9	30.2	0.753
Single submission (S)	IF2	0.657	16.3	28.7	0.640
No analysis (A)	IF2	0.754	20.9	36.7	0.744
Same 28 steps: signal-dependent and historical-analysis tasks
Full workflow	IF2	0.697	37.2	18.6	0.632
No analysis (L)	IF2	0.754	27.9	22.0	0.638
Forced measurement (M)	IF2	0.645	34.9	17.2	0.653

Editing Workflow The single-submission condition (S) retains pre-edit inspection but permits only one modifying Lua submission, without subsequent observation or revision. Under IF0, parameter matching decreases from 97.7% to 72.1%, and reference-error reduction from 85.4% to 61.3%. Paired analysis supports both decreases. Under IF2, these differences remain uncertain, although MERT direction decreases by .113. The full workflow therefore improves explicit execution. This comparison jointly removes feedback and further editing, leaving their individual effects unresolved.

Audio Access and Measurement Only seven of the 990 primary deliveries include audio measurement, all under IF2. They analyze unprocessed tracks, mainly for gate or compressor thresholds; none analyzes the processed mix or measures after editing.

To examine this behavior, we prohibit analysis (A) on the 36-step IF2 subset. Parameter matching does not decrease. On 28 selected signal-dependent or historical-analysis steps, we compare prohibited analysis (L) with prompted measurement before editing (M). Neither condition shows a clear overall difference in reference-based scores relative to the full workflow, and M shows no clear overall benefit over L. Auditing confirms successful measurements in all 28 M cases, while mean tool calls increase from 11.0 to 18.9. These results distinguish successful measurement from improved final editing; subgroup results remain exploratory.

Comparative Evaluations

We compare all five models on identical cases from the selected subset. The pooled results are presented in Table 4, with form-specific results and paired differences in Appendix § 14.

Model	Graph F1 ↑	Param. match ↑ (%)	Error reduction ↑ (%)
K3	0.870	57.0	58.4
Claude	0.878	52.3	55.0
Pro	0.851	59.3	57.8
Flash	0.851	60.5	58.6
k2.6	0.838	47.7	49.5

Flash performs close to Pro on this subset, although paired analysis does not establish equivalence. Relative to K3, k2.6’s parameter matching decreases by 9.3 percentage points and reference-error reduction by 9.0 points. Paired analysis supports both decreases, while the Graph F1 difference remains uncertain. Thus, similar graph scores can accompany different parameter and audio outcomes. Since subset selection used the primary models’ existing results, these comparisons concern the selected tasks rather than a general model ranking.

Conclusion

In this work, we present TOMORI, a code-based mixing-agent harness that connects language models to native REAPER editing through model-generated Lua programs and custom plugins. TOMORI-BENCH pairs human-reviewed edits with three instruction forms and evaluates native code execution end to end through the delivered project and rendered audio. The experiments show precise execution under explicit instructions and lower reference agreement under perceptual requests. The full workflow improves explicit execution over a single submission, whereas prompted measurement provides no clear overall improvement in reference agreement on the selected abstract cases. These findings characterize instruction-following accuracy in code-based DAW control and highlight the difficulty of translating perceptual goals into precise edits.

Future work. Future work will explore agentic reinforcement learning within TOMORI’s native DAW environment, training policies over sequences of project inspection, audio measurement, Lua editing, and verification. The aim is to learn when to gather evidence, how to revise edits, and when to stop, addressing the observed gap between successful measurement and improved mixing outcomes. Rewards could combine valid project delivery, adherence to explicit editing constraints, perceptual fulfillment of the instruction, and interaction cost. For abstract requests, expert and user preferences could complement reference-based scores to accommodate multiple valid mixing solutions. We will also investigate agents that combine project observations with direct audio input, using controlled comparisons to assess the value of perceptual feedback. Interactive clarification and preference learning could help resolve ambiguous requests and adapt editing decisions to individual mixing goals. These directions require broader data covering multiple songs and longer mixing sessions, with held-out songs, repeated runs, and a common recovery policy to assess generalization, stability, first-attempt reliability, and total interaction cost.

Limitations

While TOMORI supports native-DAW control, several limitations remain in the evaluation of language-guided mixing.

First, a perceptual request does not uniquely specify an operation. Bringing a vocal forward may involve raising its level, lowering the accompaniment, or reducing spectral masking. These implementations can produce similar sound with different edit graphs. Project metrics directly test explicit requirements, but under abstract instructions they measure agreement with one chosen implementation. Neither project nor audio agreement alone establishes whether an alternative satisfies the request.

Second, TOMORI-BENCH contains edits from one song, with one final delivery per model–instruction pair. Broader generalization and generation stability require additional songs and repeated runs. The supplementary subset is selected using existing model results, so it does not support population-level rankings. Scores assess final delivery after recovery; first-attempt reliability and total inference cost require a common recovery policy. Scorer and recovery differences are documented in Appendix § 11.§ 5.

Furthermore, the audio-access comparisons use one model and selected abstract cases. Prompts, rendering permissions, and access policies vary together. The historical-analysis groups are small and behavior-selected. These experiments do not isolate the benefit of measurement content, and uncertain differences do not establish equivalence. Selecting measurements, interpreting results, and translating them into edits remain separate challenges.

Finally, the benchmark focuses on individual edits. Although TOMORI can modify automation and routing, these capabilities are not comprehensively evaluated. Extending the tasks to sidechains, section-aware automation, and continuing client dialogue would assess more of the production workflow.

Ethical Considerations

Native DAW editing can unintentionally alter project content, and model-generated code can perform unintended operations. We evaluate edits in separate case sessions and retain execution and failure records. Project changes and rendered outputs require review before use in production. API inference and audio processing consume computational resources; total energy use and emissions are not reported.

Production dialogue can reveal personal identifiers and private creative material. Reference mix refinement and instruction review were performed by the authors; no external annotators were recruited for these tasks. The Everytime dialogue contributor confirmed consent through messages for research use and publication of de-identified excerpts. Names, filenames, local paths, account identifiers, and private media are removed from the published excerpts. Private recordings and raw dialogue logs are not included in the review materials. No formal ethics-board approval or exemption was obtained.

We cite the creators of reused software, datasets, and models. Our code uses the MIT license; our benchmark project configurations, instruction annotations, and score records use CC BY-SA 4.0, with source attribution and descriptions of the changes. The review archives exclude audio, model checkpoints, and third-party software distributions; reused artifacts remain subject to their original licenses. OpenAI Codex assisted with code implementation, result-analysis and table-generation scripts, figure preparation, and manuscript drafting and revision. The authors take responsibility for the methods, code, manuscript, and conclusions.

Interpreting Inserted-Effect Direction

Inserted effects may have native defaults that differ from the reference-construction values. In IF0 steps 26, 30, 63, and 102, Kimi and Claude retain the native Wet setting of 100%. The reference-construction Wet baseline is 0%, and its target settings are lower than the native default. These unchanged native controls still count as same-direction relative to the virtual initial value. Thus, direction scores measure agreement with the construction baseline. Native readback is required to determine which controls were actually adjusted.

Redacted Everytime-72 Excerpts

We provide excerpts from the human requests in the Everytime session. Reference mix refinement and instruction review were performed by the authors. The dialogue contributor confirmed consent through messages for research use and publication of de-identified excerpts. The text is lightly normalized, with names, filenames, local paths, account identifiers, and private media removed. Each label identifies a de-identified task in the Everytime-72 manifest, which records the source-text hash and corresponding project snapshot.

1. ET001, multi-voice space: “The harmonies still feel untreated. The reverbs are all too large; each voice needs its own EQ, and the harmony image should move slightly off center.”

2. ET003, production strategy: “Do not solve the accompaniment by level alone. Use appropriate compression and sidechain control, balance short, long, and plate reverbs, and ride the changing sections.”

3. ET006, section identity: “Treat the chant section differently: more ethereal, sacred, and extended, while correcting its level and width.”

4. ET009, sequential automation: “Fade the end of one singer’s held note so the next singer can enter, then do the same handoff for the following voice; the curves must not be straight lines.”

5. ET015, spectral accommodation: “The accompaniment was built around another lead. Make frequency room for this lead and open the harmony image slightly.”

6. ET019, staged dynamics: “The compression is too aggressive. Use a staged chain that catches sharp peaks first, broader peaks next, and overall dynamics last.”

7. ET027, delay design: “Move rhythmic space from the long reverb to a tempo-synchronized, ducked delay; keep the long reverb for depth.”

8. ET044, mix-bus correction: “Remove the excessive tape character and rebalance the bus for glue, overall dynamics, and stereo width.”

9. ET052, bounded local edit: “In this section, lift the lead by only 0.1 dB and add no more than 0.2 dB of high frequency for diction.”

10. ET070, section-aware ambience: “Automate the reverbs too: they are too large in quiet passages and too small when the accompaniment grows.”

Reproducibility Details

ScotchMorris is obtained from MedleyDB Sample under the CC BY-SA 4.0 license stated in Zenodo record 1438309 ([1]). Our code is distributed under MIT, while our benchmark project configurations, instruction annotations, and score records are distributed under CC BY-SA 4.0. Package notices identify the source, retain attribution, and describe the construction and anonymization changes. Original audio and third-party software are not redistributed. Redaction and consent for the Everytime excerpts are described in Appendix § 10.

The review materials contain case instructions, project snapshots, scores, and software for execution and evaluation. Audio files, automatic saves, and process backups are excluded. Data manifests record project and media checksums, project identity, audio format, and instruction correspondence. Reference-validation records include the approved endpoint, construction parameters, cumulative readbacks, and language review. These records support aggregate-score verification; recomputing audio metrics requires the corresponding audio files, which are not distributed in the review archives.

GPU-based computation used NVIDIA A100 hardware. No complete computational-budget estimate is available. Selected-delivery tool-call and timing statistics do not include all failed, recovery, and preliminary runs, and cannot establish total inference cost.

Native Execution

The bridge executes commands serially against the pinned project. Each command returns an identifier, status, results, and errors. We retain source code, before/after projects, parameter snapshots, and execution output, including records from partial failures. Failed commands are not automatically retried or rolled back. An unresolved bridge outcome blocks further execution until inspection. These records support verification of code against project changes; they do not establish whether the chosen edit satisfies the request.

Audio Preprocessing

We resample scoring copies to 30 kHz using scipy.signal.resample_poly with the default Kaiser 5.0 filter, preserving the original 48 kHz WAVs. Each excerpt omits its first five seconds, leaving 20–30 and 60–75 seconds. Signals and excerpts are independently normalized to −23 LUFS before the frozen MRSTFTLoss(target,prediction) call.

The loss uses FFT sizes 512/1024/4096, hop sizes 128/256/1024, 96 mel bins, and perceptual weighting. Complete 3.8-second windows are evaluated at a 2.8-second stride, omitting the first second of each window. We average windows, weight the two excerpts equally, and macro-average cases. Edit magnitudes $a_i$ and $m_i$ use the same settings with the before-state as the first argument. Since the loss is asymmetric, $a_i$ differs slightly from $\ell_i^{\mathrm{noop}} = L(y_i,b_i)$.

For Pearson correlation, we subtract channel means and compare corresponding stereo channels at zero lag. Channels and excerpts receive equal weight. We apply no shift search, polarity correction, or channel swap; constant channels yield undefined correlation. Correlation and loudness-normalized loss assess waveform and spectral agreement, while native controls assess track-level changes.

Graph Scoring

The edit graph represents net changes between the original before-snapshot and selected final snapshot. Each inserted or removed FX instance and its configuration forms one node. Existing FX parameter changes and track-control changes form separate nodes. Links encode track-specific effect adjacency and connections to track inputs and outputs. Only added or removed links are compared. Display prefixes and archive suffixes are normalized; different plugins remain distinct operators.

Graph F1 pools node and link match counts and denominators within each case, then macro-averages cases. Two empty graphs score one, while only one empty graph scores zero. Absent links contribute nothing to the denominator. Node and link F1 are reported for diagnosis and are not averaged to obtain Graph F1.

Direction compares the sign of final-minus-baseline with the reference increment. Inserted FX use reference-construction initial values as the baseline, rather than native defaults. If any required parameter is opposite, unchanged, invalid, or unmatched, the entire reference node fails to match, even if its edit identity agrees. Changed links are matched independently. Error categories are reported separately in Table 8. Direction checks only required reference controls; it does not separately test FX enablement or additional Wet settings and does not establish audible contribution. All full-benchmark Graph F1 values use the joint node/direction rule (v3), rescored from archived graph and action records without rerunning models or audio.

	Node F1	Link F1	Opp.	Invalid	Unch.	Unmat.	[email protected]
kimi-k3
	IF0	1.000	1.000	0	0	0	0.934
	IF1	0.921	0.985	0	0	9	2	0.339
	IF2	0.822	0.945	0	0	11	3	0.207
claude-fable-5
	IF0	0.964	1.000	4	1	0	0	0.826
	IF1	0.906	0.976	1	3	7	2	0.306
	IF2	0.795	0.945	2	0	16	1	0.215
deepseek-v4-pro
	IF0	1.000	1.000	0	0	0	0	0.975
	IF1	0.945	0.976	0	0	6	2	0.380
	IF2	0.786	0.909	2	0	13	7	0.223

Parameter and Representation Scoring

Parameters are matched by track, effect, and control identity. Let $p_j$ and $\hat{p}_j$ denote the reference and observed values of required control $j$. FX uses native DAW $[0,1]$ values, and pan is linearly mapped from $[-1,1]$ to $[0,1]$. Let $\mathcal{P}$ contain all required controls and $\mathcal{F}$ those with matched, finite observations. Then

Missing observations are excluded from MAE, so the number of computable controls is also reported. Finite out-of-range values retain their raw errors without clipping. Let $\mathcal{V}\subseteq\mathcal{F}$ contain controls with valid values. Accuracy at absolute tolerance $\epsilon$ is

The denominator includes every required control; unmatched, non-finite, and out-of-range values receive no match credit. Track gain uses the same absolute-error calculation in dB and is reported separately from normalized parameters.

Each model has 363 required normalized parameters (121 per form) and 63 track-gain obligations. Normalized MAE uses 358, 360, and 354 finite observed errors for Kimi, Claude, and DeepSeek, respectively. Tolerance accuracy compares unrounded absolute errors. Separately reported four-decimal matching uses Python rounding; the two matching definitions are computed independently.

MERT-v1-95M is fixed at revision 12af15fef9d0. Each excerpt omits five seconds of warm-up, averages stereo channels, and resamples to 24 kHz. Audio is split into five-second chunks and normalized by the checkpoint extractor; final-layer representations are averaged over frames and then chunks to obtain the excerpt feature. No time alignment or LUFS normalization is applied. Native controls and stereo correlation complement the spatial limitations of mono features.

Delivery Validation

Valid audio requires the specified format and duration, finite decoded samples, and AC RMS above −90 dBFS. Final project and audio deliveries for all 330 cases per model passed the execution-host checks. When an attempt fails, the first valid supplemental delivery in registry order is used without selecting by score. Recovery can resume a saved project with further model calls; scoring uses the final selected delivery.

Archived rows cover all 330 cases per model with matching selection hashes and reference-audio hashes. All 660 MERT excerpts per model have defined scores and share the same checkpoint and preprocessing settings. The scorer versions differ in invalid-value handling and audio-integrity recovery. DeepSeek’s IF2 step 8, marker 20 uses the original hash-pinned 30 kHz scoring copy because the delivered WAV no longer matches its frozen hash; the other 659 excerpts use verified delivery files. Removing this case from all three models changes IF2 percentages of edits in range by at most 0.5 percentage points and preserves their ordering. This is a sensitivity analysis of stored scores, not fresh validation of the external WAV files.

Step Construction

We construct the editing sequence by progressively undoing processing from the approved mix. MR-STFT loss selects changes that increase the difference from the finished mix. Reversing their order produces a sequence from a neutral project toward the target. We additionally introduce excessive adjustments and their corrections, yielding 72 normal edits, 19 overshoots, and 19 repairs.

Task Coverage

Operation family	Steps
EQ	50
Gain/pan	29
Reverb	14
Compression	11
Gate	6
Stereo width	4
Delay	1
Unique steps	110

Supplementary Model Comparison

DeepSeek v4 Flash and Kimi k2.6 use the same TOMORI harness, with project inspection, editing tools, audio-analysis access, three-request history, and managed rendering. Each produces 72 valid final deliveries after recovery. Kimi k2.6 requires eight replacement deliveries following infrastructure interruptions. These counts describe selected outputs and do not measure first-attempt reliability.

The 36-step subset is stratified by operation family and normal/overshoot/repair status: EQ 15, gain/pan 7, compression 5, reverb 5, gate 2, width 1, and delay 1; 22 normal edits, seven overshoots, and seven repairs. With seed 20261010, we generated 500 candidate draws and selected the one minimizing the maximum absolute deviation from the primary models’ full results across IF0/IF2 Graph F1, tolerance accuracy, error reduction, and improvement over the unchanged project. Selection used the archived Graph scores available before joint direction matching; all scores reported here use the revised joint rule. This outcome-informed selection

limits inference to the selected steps, rather than an independent sample of unseen tasks. Table 6 compares all five models on identical cases. Table 7 reports paired differences for Flash versus Pro and k2.6 versus K3. Bootstrap sampling uses 10,000 draws by reference step, keeping IF0 and IF2 together, to assess uncertainty. Parameter matching aggregates required-control counts; other reference-based scores average cases. Comparisons are exploratory and conditional on this subset. Timing excludes previous failed attempts and does not establish total cost or deployment latency.

	Graph F1 ↑	Param. MAE ↓	Param. match ↑ (%)	Error reduction ↑ (%)	MERT dir. ↑	Seconds ↓
kimi-k3
IF0	1.0000	0.0413	93.0	80.3	0.868	145.5
IF2	0.7399	0.1253	20.9	36.6	0.739	154.3
claude-fable-5
IF0	0.9931	0.0853	79.1	75.7	0.841	157.0
IF2	0.7630	0.1271	25.6	34.3	0.745	116.3
deepseek-v4-pro
IF0	1.0000	0.0059	97.7	85.4	0.894	330.1
IF2	0.7011	0.1473	20.9	30.2	0.753	176.6
deepseek-v4-flash
IF0	0.9944	0.0124	95.3	83.2	0.866	89.6
IF2	0.7079	0.1392	25.6	34.0	0.719	115.1
kimi-k2.6
IF0	0.9722	0.0672	83.7	70.2	0.825	226.1
IF2	0.7043	0.1698	11.6	28.7	0.699	229.1

	Graph F1	Param. match (pp)	Error reduction (pp)
Flash – Pro
IF0	-0.006	-2.3	-2.1
IF2	+0.007	+4.7	+3.9
Pooled	+0.001	+1.2	+0.9
k2.6 – K3
IF0	-0.028	-9.3	-10.0
IF2	-0.036	-9.3	-7.9
Pooled	-0.032	-9.3	-9.0

The paired results show no clear difference on the reference-based metrics between Flash and Pro, without establishing equivalence. Flash’s selected attempts finish sooner, although recovery time is excluded. For k2.6 versus K3, the pooled decreases in parameter matching and error reduction are supported, while the Graph F1 difference remains uncertain. These findings apply to the selected subset.

Ablation Protocol and Diagnostics

All conditions use DeepSeek v4 Pro, three-request history, managed evaluator rendering, and the same Graph v3, parameter, MR-STFT, and MERT definitions. Baselines reuse selected full-benchmark deliveries. Final valid deliveries cover 72 S cases (36 steps under IF0 and IF2), 36 A cases (IF2), and 28 cases each for L and M (IF2). They are not additional reference edits or independent songs.

S retains pre-edit inspection but allows only one modifying Lua submission, with no returned readback or subsequent tools. Computation and multiple operations within that submission remain possible; the policy audit verifies one submission per case. A and L prohibit model access to audio files, model-driven rendering, and analysis; evaluator rendering remains available for scoring. M permits audio access and prompts the model to measure relevant original tracks or the current mix before the first edit, then explain how the measurement informs editing. Auditing confirms successful signal measurements in all 28 cases, excluding file-format inspection; it does not validate their relevance or influence on decisions. These conditions assess combined execution or audio-access policies without isolating prompts, rendering permissions, and measurement content.

S and A use the outcome-informed 36-step subset in Appendix § 14. L and M use the same 28 steps: 12 core tasks whose reference settings require signal information, 15 secondary tasks whose adjustment size depends on the signal, and IF2 steps where any primary model historically analyzed audio. Core tasks cover dynamic thresholds or joint EQ frequency/gain changes; secondary tasks cover fixed-band EQ gain or compressor timing. Seven L cases reuse A deliveries and 21 are new; M adds 28 final deliveries. The historical-analysis strata in Table 9 overlap with the reference-selected strata and with each other; the Pro-specific stratum contains five cases.

Condition	Form	Graph F1 $\uparrow$	Param. match $\uparrow$ (%)	Error reduction $\uparrow$ (%)	MERT dir. $\uparrow$
Core signal-dependent tasks (12 cases)
Full workflow	IF2	0.592	29.2	10.8	0.410
No analysis (L)	IF2	0.672	20.8	11.5	0.441
Forced measurement (M)	IF2	0.585	33.3	10.0	0.492
Secondary signal-dependent tasks (15 cases)
Full workflow	IF2	0.828	50.0	26.1	0.787
No analysis (L)	IF2	0.870	38.9	31.9	0.788
Forced measurement (M)	IF2	0.737	38.9	24.2	0.760
Any primary model analyzed audio (6 cases)
Full workflow	IF2	0.667	36.4	11.1	0.563
No analysis (L)	IF2	0.708	27.3	0.5	0.329
Forced measurement (M)	IF2	0.583	36.4	6.9	0.568
Pro baseline analyzed audio (5 cases)
Full workflow	IF2	0.600	44.4	13.3	0.675
No analysis (L)	IF2	0.650	33.3	0.6	0.394
Forced measurement (M)	IF2	0.650	44.4	8.1	0.641

Paired bootstrap analysis uses 10,000 reference-step draws, retaining both forms together for S. It supports the IF0 decreases in Graph F1, parameter matching, MR-STFT error reduction, and MERT direction under S; under IF2, only the MERT direction decrease is supported among these metrics.

A’s observed error reduction increases without improved parameter matching. Overall differences on the reference-based metrics for L and M versus the full workflow remain uncertain, and M shows no clear overall benefit over L. Paired analysis supports increased tool calls under M. Counts concern selected deliveries, not total inference cost including earlier failed attempts.

Historical-analysis strata suggest possible task-specific differences. On the five tasks where Pro’s baseline analyzed audio, MERT direction is .675 for the full workflow, .394 for L, and .641 for M. Paired analysis supports M’s increase over L, while its difference from the full workflow remains uncertain. Table 9 reports all specified strata rather than highlighting only this result. These behavior-selected subgroup comparisons are exploratory and unadjusted for multiple comparisons; they establish neither equivalence, the dispensability of analysis, nor an isolated component benefit. Scores and audits reflect stored final deliveries, rather than fresh local replay of external audio files.