WorksComputer Science
Right Reason, Wrong Advice: Evaluating Recourse in LLM Explanations of Denied Requests
Abstract
Language models increasingly write the explanations people receive when a rule system denies their request. If several rules deny the request independently, an explanation can name a true reason and still leave the person denied after they fix it. We call the missing property recourse validity: changing what the explanation cites, without changing what was asked for, can flip the decision. We compute it exactly, and a hitting-set argument predicts from the policy alone when a correct reason fails. In six of seven rule systems ported from access control and US benefit eligibility, a single correct reason almost never gives recourse when several reasons apply. Across 20 LLMs, checks of whether the cited reasons are correct, computed or asked of an LLM judge, are close to chance at spotting this failure, as the argument predicts. A strong judge asked whether acting on the explanation would work does spot it. Stronger and reasoning models do better but still fail on many denials with four or more independent reasons. Asking for every rule helps every open model with room to improve. Explanations of eligibility decisions should be built from the decision function and judged by whether acting on them works.
Introduction
Rule systems decide many requests people make. Cedar and OPA Gatekeeper admit or reject API calls and Kubernetes pods. Benefit agencies apply eligibility rules for food assistance, public housing and student aid. The decision is exact, but the explanation a person reads is more and more often generated by an LLM from the policy text and the request (Inan et al., 2023; Rebedea et al., 2023). Data-protection law asks that automated decisions come with meaningful information about the logic involved (Wachter et al., 2017; Goodman and Flaxman, 2017), and work on algorithmic recourse asks more: that an explanation tell the person what they could change to obtain a different outcome (Wachter et al., 2018; Ustun et al., 2019; Karimi et al., 2022).
Figure 1 shows a case from our benchmark. A family applying for public housing has net assets above the limit and also owns a home suitable for occupancy, and under HUD’s rules each fact bars admission on its own. Gemma-4-12B, given the rules and the application, explains that the family’s assets exceed the limit. The explanation is true, cites the right regulation and names a rule that fired. A family that acts on it, brings its assets under the limit and reapplies is denied again, this time because it owns a home.

Figure 1. A correct explanation that gives no recourse, from the HUD port. Neither fact is immutable, so citing both would give actionable recourse.
Existing evaluations reward this explanation. Faithfulness tests ask whether an explanation reflects the process behind the decision (Jacovi and Goldberg, 2020; Atanassova et al., 2023; Parcalabescu and Frank, 2024), and its does. Simulatability asks whether a reader can predict the decision [11] (Hase and Bansal 2020). Causal recall against an exact oracle (V, 2026) credits the best coverage of any single reason, so this explanation scores 1. An LLM judge asked whether the reasons are correct has no reason to object. None of these asks whether acting on the explanation changes the outcome.
We define that property and compute it exactly from the decision function, given the features an extractor reads from each explanation. An explanation is recourse-valid if some change to the features it cites, holding fixed what the requester is asking for, flips the decision; equivalently, the cited features contain a contrastive explanation restricted to what the requester may change. One reason gives recourse if and only if it shares a feature with every other goal-fixed reason (Proposition 1, the hitting-set duality of formal explanation restated with goals), so the failure is predictable from the policy alone. In six of seven ported rule systems, no single correct reason gives recourse in 0.95 to 1.00 of the denials with two or more goal-fixed reasons (§ 3).
We score 141,750 explanations from 20 LLMs, from 8B open models to frontier APIs (Figure 2, Appendix § 22). Model behaviour follows Corollary 1. Citing a true reason predicts recourse on single-reason denials (AUROC 0.93, small open models) and in the one system whose reasons share a feature (0.78). On the other multi-reason denials every correctness signal sits near chance, with reasoning off or on, and recourse is predicted instead by how much of the reason structure an explanation covers (§ 5.1). LLM judges asked whether the reasons are correct pass nearly every explanation; a strong judge asked whether acting on it would work tracks the oracle (§5.3). Recourse validity tracks capability and reasoning: thinking lifts the four open models we ran with it from between 0.40 and 0.77 to between 0.89 and 0.99. It does not remove the failure: at four or more independent reasons, six of the nine reasoning runs, thinking-on and frontier, fall below 0.8 (§ 5.2). Asking for every rule narrows the gap for every open model that has one (§ 6), and advice works as written less often than recourse validity credits (§ 7).

Figure 2. The oracle computes the goal-fixed reasons from the gate, and an extractor reads the features an explanation cites. Correctness checks ask whether contains a reason; recourse validity asks whether hits all of them, which one reason does only if it shares a feature with every other (Proposition 1).
Our contributions are:
Recourse validity, a criterion that asks whether what a natural-language explanation cites contains a goal-restricted contrastive explanation, exact given those features (§ 2).
A benchmark of 420 episodes over seven ported rule systems, with goal and immutable features for each port (§ 3).
A measurement over 20 LLMs. On multi-reason denials, correctness-style evaluation, including causal recall and an LLM judge asked about correctness, cannot tell explanations that give recourse from those that do not; a strong judge asked about the outcome can, and reasoning fails where independent reasons multiply (§ 5).
Mitigations ranked by the gap they close, and a test of whether the advice works as written (§ 6, § 7).
Recourse validity
Decisions and reasons. A rule system is a decision function over , a product of finite feature domains. includes the policy’s combination semantics.
Under deny-overrides, as in Cedar, a request is allowed only if some permit matches and no forbid matches. Under violation-any, as in Gatekeeper, a request is rejected if any violation rule matches. Fix a context with . For let . The set is sufficient at if for every . The subset-minimal sufficient sets are the reasons for the denial, the abductive explanations of formal XAI (Shih et al., 2018; Ignatiev et al., 2019a). A rule-level oracle computes them from each ported policy.
Goals. A requester asks for something specific. A student applies for a Pell grant, a developer creates a pod, a user asks to edit a document. Let be the goal features that encode what is asked for. Changing a goal feature means asking for something else, so it is not recourse. For each port we name in Table 1.
Definition 1 (Goal-fixed recourse validity, ). Let an extractor map explanation to the features it cites as reasons, and let be the cited features the requester may change. Then
This holds exactly when the kept features are not sufficient at , that is, when contains a contrastive explanation of among the features the requester may change (Ignatiev et al., 2021). RV is undefined if is empty, and for the whole episode if alone is sufficient at , since no explanation can then give recourse.
We write RV for where no confusion arises. RV is an existence condition. It credits an explanation if some assignment to the cited features flips the decision, even when the explanation does not say which, so it is an upper bound on how often acting on an explanation succeeds: when it fails, no change to what the explanation cites helps; when it succeeds, the explanation may still not lead the person to a working change. § 7 measures the distance by executing the changes explanations propose. RV is also monotone in what is cited, since citing every changeable feature always succeeds, so we report causal precision beside it.
When one reason suffices. Define the goal-fixed reasons as the subset-minimal members of . A set together with is sufficient iff contains a member of , and exactly when alone forces the denial.
Proposition 1. Let and let not be sufficient at . Then if and only if meets every member of .
Proof sketch. Let , so . If some misses , then for some , so , and is sufficient. Conversely, if is sufficient it contains some , and contains some with , disjoint from .
Proposition 1 is the goal-restricted form of the hitting-set duality between abductive and contrastive explanations (Ignatiev et al., 2021; Marques-Silva and Ignatiev, 2022). We claim no new theory and use it only to predict when generated explanations fail.
Corollary 1. If cites exactly one goal-fixed reason , then iff meets every other member of . If has at least two members and they are pairwise disjoint, every single-reason explanation fails. If all members share a feature, a necessary feature in the sense of Huang et al. (2023), every single-reason explanation succeeds.
So the gap is visible from the policy before any model runs, and fixing the goal can create it: every privileged-containers violation requires that the request is not an update, so goal free one reason suffices (“make this an update”), and goal fixed none does (Table 1).
Correctness measures. Fidelity asks whether an explanation discloses a reason the policy’s trace reports (a matching forbid, a missing permit or a firing violation). The causal recall of V (2026) is the best coverage of any single reason, so one complete reason scores 1 and can still have . Causal precision is the share of cited features that belong to some reason. Actionable recourse validity replaces with , where are immutable features a requester cannot reasonably be told to change; it is undefined where alone forces the denial, so it covers fewer denials than and can exceed it.
Benchmark
Porting protocol. Each port is an executable gate over categorical features that abstract the source’s predicates, keeps its source’s combination semantics, and records every omitted provision with the reason in the port file. Cedar contributes the document-cloud and GitHub example policies (Cedar Policy, [8]), Gatekeeper the allowed repositories and privileged-containers templates (Open Policy Agent, 2026); a single-container pod can violate each template at most once, so pods have three container slots. The eligibility ports encode SNAP (7 CFR 273; USDA Food and Nutrition Administration, 2026) for a one-person household, HUD public housing admission (24 CFR 960.201, 960.204 and 5.603 to 5.618; U.S. Department of Housing and Urban Development, 2026) and Title IV student aid (34 CFR 668.32 to 668.35; U.S. Department of Education, 2026). A port script checks each provision against the outcome the regulation fixes and regenerates the port file byte for byte. Cedar’s tinytodo policy and the single-container Gatekeeper ports admit
| Rule system | Contexts | Goal | Denies | Two+ reasons | Left denied free | Left denied fixed | Left denied actionable | No single suffices | No recourse | |
| Cedar document-cloud | 10 | 5,376 | action | 4,504 | 1,900 | 0.88 | 0.96 | 0.998 | 0.95 | 0.15 |
| Cedar github-model | 7 | 704 | action | 336 | 16 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 |
| Gatekeeper allowed repos | 5 | 576 | none | 448 | 192 | 1.00 | 1.00 | 1.00 | 1.00 | 0.00 |
| Gatekeeper privileged | 9 | 1,152 | is_update | 276 | 56 | 0.00 | 1.00 | 1.00 | 1.00 | 0.00 |
| SNAP eligibility | 10 | 3,456 | none | 3,108 | 2,176 | 0.96 | 0.96 | 0.92 | 0.95 | 0.45 |
| Title IV student aid | 10 | 3,456 | aid_program | 3,428 | 3,246 | 0.99 | 0.999 | 1.00 | 0.998 | 0.70 |
| HUD public housing | 10 | 2,304 | none | 2,292 | 2,204 | 1.00 | 1.00 | 1.00 | 1.00 | 0.86 |
caption: Rule systems and the oracle-level gap, exact over every context. Cedar and the eligibility rules use deny-overrides, Gatekeeper violation-any on 3-container pods. Two+ reasons: denials with two or more goal-fixed reasons. Left denied: share of (denial, reason) pairs in which citing that one correct reason leaves the request denied, over denials with two or more reasons (free) or two or more goal-fixed reasons (fixed; actionable also fixes the immutable features). No single suffices: share of those denials in which no reason cited alone gives goal-fixed recourse. No recourse: share that the goal and immutable features alone deny. Shares are under the uniform context measure, not population rates.
:::
no multi-reason denial as ported, and we exclude them.
Goals and episodes. The goal is action for Cedar, aid_program for Title IV and is_update for the privileged-containers pod; the other ports have none. Immutable features follow the source text (Appendix § 25). Every port’s context space is small enough to enumerate, and for each port we draw 30 multi-reason denials, which have two or more reasons with the goal free, and 30 single-reason controls uniformly without replacement from all its denials of each kind, with a fixed seed, 420 episodes in all. With the goal fixed, 186 of the 210 multi-reason episodes keep two or more reasons, 18 keep one and 6 have none, as do 7 controls, so no explanation of those 6 and 7 can give recourse; Cedar github-model has only 16 contexts with two or more goal-fixed reasons, and 13 of its 30 sampled episodes are among them. Rates are over episodes with a defined verdict.
The oracle-level gap. In six of seven rule systems, no single correct reason gives recourse in 0.95 to 1.00 of the denials with two or more goal-fixed reasons, and a correct reason cited alone fails in 0.96 to 1.00 of (denial, reason) pairs (Table 1). Cedar github-model is the negative control: its reasons share a role-membership feature, so by Corollary 1 every single reason succeeds, and the measured share is 0.00. In the eligibility rule sets each disqualifying condition denies on its own, so 2,176 of 3,108 SNAP denials and 3,246 of 3,428 Title IV denials have two or more reasons, and with immutable features fixed 0.45 to 0.86 of their multi-reason denials admit no actionable recourse at all.
Models, conditions and extraction
Conditions. Every prompt gives the request, the decision and the policy with one sentence stating its combination semantics. The default asks for 2 to 4 sentences for the audit log (Appendix § 13). List-all also asks for every rule or missing permission that causes the denial. Requester asks for an explanation to the person who made the request, as long as it needs to be; we run it on the 11 small open models, Haiku and GPT-6 Luna, 13 in all. Free and free-all show the request and decision but not the policy, as when an LLM narrates a decision made by a system it cannot inspect. § 6 adds four conditions on three small models: uncapped list-all, a counterfactual prompt that asks what the person should change, the oracle’s minimum change, and every reason (Appendix § 13).
Models. We run 20 models in three groups. Open, 8B to 21B (11): Llama-3.1-8B, Granite-4.2-8B, gpt-oss-20b, Mistral-Nemo-12B, Ministral-3-14B, Gemma-3-12B, Gemma-4-12B, Qwen2.5-14B, Qwen3-14B (Yang et al., 2025), Qwen3.5-9B and Phi-4-14B. Open, 27B to 32B (four): Gemma-3-27B, Gemma-4-31B, Qwen3-32B and Qwen3.8-27B. Frontier (five): Claude Haiku 4.5, Sonnet 5.5 and Opus 5.5, GPT-6 Luna and GPT-6.1 Sol. Every model answers each prompt 5 times. Open models run 4-bit through ollama on two T4 GPUs at temperature 0.8, with at most 400 new tokens and thinking off; 13% of Granite-4.2-8B’s answers reach that cap and at most 1% of any other model’s. gpt-oss-20b (OpenAI, 2025) cannot turn reasoning off, so it runs at low reasoning effort with a 1,500-token budget that includes the hidden reasoning; 0% of its answers reach it, the longest at 458 tokens. Frontier models run through vendor interfaces at low reasoning effort (Haiku at its default) without temperature control. Large and frontier models answer the 210 multi-reason denials, the others all 420 episodes. With 8-bit weights, default and list-all recourse validity of Llama-3.1-8B and Gemma-3-12B move by −0.06 to 0.09, so 4-bit quantisation does not drive the small models’ rates.
Extraction. Qwen3-14B in JSON mode at temperature 0 lists the features an explanation presents as reasons. A second extractor, Gemma-3-12B, re-extracted a sample of explanations from each of the 20 models. It agrees with the first at Krippendorff’s (micro-F1 0.94) over 2,000 explanations from six models and at over 1,398 from the other 14, where recourse validity under its labels moves by at most 0.04 per model. One author labelled 200 explanations stratified over models, conditions and systems; agreement with the extractor is (micro-F1 0.91), and the two give the same recourse verdict on 95% of items, [0.84, 0.96]. The extractor errs toward validity (eight items valid only under its labels, one only under the human’s), and recourse validity on the human labels is within 0.06 of the extractor’s in every condition.
Statistics. Rates are over explanations that cite at least one feature. Cells average per-episode means, since an episode’s 5 samples share a prompt, with 95% intervals from resampling episodes; AUROCs resample episodes too, and a within-episode AUROC fixes the reason count. Paired effects use sign-flip tests ( permutations, Holm-corrected); correlations carry intervals from resampling models.
Results
Correctness checks miss the failure
Table 5 in Appendix § 16 gives every model. Fidelity runs from 0.74 to 0.97 and recourse validity from 0.23 to 0.99. Corollary 1 says when a correct reason should predict recourse: on a single-reason denial, citing the true reason is what recourse needs; on a multi-reason denial whose reasons share no feature, it is not enough. Model behaviour follows the prediction. For the small open models, fidelity separates recourse-valid explanations at AUROC 0.93 [0.91, 0.94] on the single-reason controls and
| Model | Fidelity | Precision | Facts | [95% CI] | Random | List-all | Recall 1, denied | ||
| Open, 8B to 21B | |||||||||
| Llama-3.1-8B | 0.75 | 0.75 | 2.9 | 0.31 [0.27, 0.35] | 0.24 | 0.09 | 0.13 | 0.55 | 66% |
| Granite-4.2-8B | 0.96 | 0.76 | 4.7 | 0.65 [0.59, 0.70] | 0.43 | 0.37 | 0.46 | 0.88 | 36% |
| gpt-oss-20b | 0.87 | 0.91 | 3.0 | 0.55 [0.50, 0.60] | 0.23 | 0.44 | 0.50 | 0.90 | 45% |
| Mistral-Nemo-12B | 0.74 | 0.81 | 2.5 | 0.24 [0.20, 0.29] | 0.20 | 0.05 | 0.08 | 0.58 | 78% |
| Ministral-3-14B | 0.92 | 0.79 | 4.1 | 0.67 [0.61, 0.72] | 0.36 | 0.44 | 0.55 | 0.91 | 33% |
| Gemma-3-12B | 0.88 | 0.90 | 2.2 | 0.23 [0.18, 0.28] | 0.18 | 0.07 | 0.13 | 0.84 | 79% |
| Gemma-4-12B | 0.82 | 0.95 | 3.0 | 0.70 [0.65, 0.76] | 0.24 | 0.50 | 0.54 | 0.84 | 34% |
| Qwen2.5-14B | 0.91 | 0.91 | 3.2 | 0.51 [0.45, 0.57] | 0.29 | 0.16 | 0.24 | 0.77 | 50% |
| Qwen3-14B | 0.92 | 0.88 | 3.1 | 0.40 [0.33, 0.46] | 0.27 | 0.11 | 0.17 | 0.79 | 63% |
| Qwen3.5-9B | 0.87 | 0.80 | 3.5 | 0.47 [0.42, 0.52] | 0.26 | 0.35 | 0.39 | 0.80 | 57% |
| Phi-4-14B | 0.93 | 0.82 | 3.9 | 0.57 [0.52, 0.63] | 0.34 | 0.36 | 0.44 | 0.72 | 44% |
| Open, 27B to 32B | |||||||||
| Gemma-3-27B | 0.90 | 0.85 | 3.2 | 0.44 [0.38, 0.50] | 0.23 | 0.40 | 0.53 | 0.87 | 59% |
| Gemma-4-31B | 0.80 | 0.93 | 3.5 | 0.97 [0.95, 0.99] | 0.28 | 0.97 | 0.99 | 0.98 | 2% |
| Qwen3-32B | 0.87 | 0.90 | 3.3 | 0.54 [0.48, 0.59] | 0.28 | 0.32 | 0.42 | 0.84 | 50% |
| Qwen3.8-27B | 0.83 | 0.94 | 3.3 | 0.77 [0.72, 0.81] | 0.25 | 0.74 | 0.84 | 0.93 | 26% |
| Frontier | |||||||||
| Claude Haiku 4.5 | 0.89 | 0.93 | 3.9 | 0.96 [0.93, 0.98] | 0.33 | 0.90 | 0.96 | 0.98 | 5% |
| Claude Sonnet 5.5 | 0.97 | 0.82 | 4.9 | 0.94 [0.92, 0.97] | 0.45 | 0.87 | 0.96 | 0.96 | 6% |
| Claude Opus 5.5 | 0.96 | 0.81 | 4.9 | 0.99 [0.98, 1.00] | 0.45 | 0.99 | 1.00 | 0.99 | 1% |
| GPT-6 Luna | 0.84 | 0.89 | 3.8 | 0.92 [0.89, 0.94] | 0.33 | 0.87 | 0.98 | 0.97 | 9% |
| GPT-6.1 Sol | 0.85 | 0.83 | 4.4 | 0.99 [0.98, 1.00] | 0.39 | 0.99 | 1.00 | 0.99 | 1% |
Table 5. The 210 multi-reason denials, default prompt unless a column says otherwise. Fidelity: discloses a reason the policy’s trace reports. Facts: mean facts cited. : mean of per-episode means with its 95% bootstrap interval. Random: of as many random facts of the episode. Eligibility: mean over SNAP, Title IV and HUD; also fixes immutable facts and is defined on fewer denials. Recall 1, denied: share of explanations with causal recall 1 that leave the requester denied.
at 0.49 [0.48, 0.51] on the multi-reason denials. Among the multi-reason systems it predicts recourse only on Cedar github-model, whose reasons share a feature, at 0.78 [0.75, 0.80], against 0.50 to 0.57 on the other six. Over all 20 models, the three correctness signals separate recourse-valid explanations of multi-reason denials near chance (Figure 3): fidelity at AUROC 0.49 [0.47, 0.50], causal recall at 0.48 and causal precision at 0.48. Within a model it ranges from 0.42 to 0.64. The blind spot does not depend on how often models fail. On the 4,072 thinking-on explanations, 93% of them recourse-valid, fidelity separates at 0.49 [0.46, 0.52] and causal recall at 0.48, while coverage still reaches 0.75. Of the 17,859 explanations with causal recall 1, 37% leave the requester denied, up to 79% for Gemma-3-12B.
Figure 3. How well each per-explanation signal separates recourse-valid explanations from the rest (AUROC, 95% interval from resampling episodes; dotted line: chance). Circles: 20,375 default explanations of multi-reason denials from all 20 models; squares: 4,072 thinking-on explanations; triangles: single-reason controls, small open models. Judges are GPT-6 Luna on a sample (§5.3). Within-model and within-episode values are in Table 8.
| Signal | AUROC | 95% CI | Within model | Within episode |
| Is what it cites correct? | ||||
| Fidelity | 0.49 | [0.47, 0.50] | 0.48 | 0.56 |
| Causal recall | 0.48 | [0.47, 0.50] | 0.47 | 0.56 |
| Causal precision | 0.48 | [0.46, 0.50] | 0.45 | 0.58 |
| How much does it cover? | ||||
| Two+ reasons | 0.60 | [0.58, 0.63] | 0.51 | 0.70 |
| Coverage | 0.73 | [0.70, 0.75] | 0.71 | 0.83 |
| Facts cited | 0.75 | [0.73, 0.78] | 0.67 | 0.87 |
Table 8. Signal AUROCs behind Figure 3: how well each per-explanation signal separates recourse-valid explanations from the rest, over 20,375 default explanations of multi-reason denials from all 20 models. The interval resamples episodes. Within model: mean of per-model AUROCs, which a signal cannot raise by telling models apart. Within episode: mean over the 202 episodes with both outcomes, which fixes the reason count. Two+ reasons: cites at least two of the episode’s reasons in full; Coverage: share of the episode’s reasons cited in full (both with the goal free).
Coverage of the reason structure predicts recourse instead. An explanation that covers one goal-fixed reason in full is recourse-valid in 0.49 of cases, no more often than one that covers none (0.48); one that covers two or more is valid in 0.75. Within an episode, coverage and the number of facts cited reach AUROC 0.83 and 0.87. In a logistic model with model fixed effects and episode-clustered errors, fidelity’s coefficient is indistinguishable from zero once coverage and facts cited are known (−0.46, 95% CI [−2.55, 1.63]), while both of those carry information ($p < 0.001). Covering every reason implies recourse by Proposition 1, so the coverage term is partly definitional; the informative result is the fidelity term.
Two nulls separate which facts a model chooses from how many it cites. A random citation of as many facts of the episode gives less recourse than every model does, by 0.04 for Mistral-Nemo-12B to 0.70 for Gemma-4-31B; for Gemma-3-12B and Mistral-Nemo-12B the 95% interval reaches the random rate. A random draw of as many facts from those that belong to some reason is the stronger null, and every one of the 11 small open models falls below it, by as much as 0.26 for Llama-3.1-8B; the frontier models sit at it (margins of −0.001 to 0.05). The negative control behaves as predicted: on the 13 Cedar github-model episodes that keep two or more goal-fixed reasons, which share a feature, the small open models give recourse in 0.87 on average, and over all its episodes the system has the highest default recourse validity of the seven (Appendix § 17).
Who gives recourse
Recourse validity tracks capability. Across the 15 models with a published Epoch Capabilities Index ([20]) it follows that index, Spearman 0.88 [0.64, 0.96], and across the 10 open models alone 0.71 [0.06, 0.92] (Appendix § 23); among the seven small open models with an index it does not separate them (0.54, ). The frontier models give recourse in 0.92 to 0.99 of multi-reason denials, the open models of 27B to 32B in 0.44 to 0.97, and those of 8B to 21B in 0.23 to 0.70. Models of similar capability still differ widely. The five models with an index between 138 and 143 range from 0.40 to 0.97: Qwen3-14B, Qwen3.5-9B and Qwen3-32B at the bottom, Haiku and Gemma-4-31B at the top. Reasoning explains much of the band. With thinking on (gpt-oss-20b at high effort) and a 3,000-token budget that no answer reaches, Qwen3-14B, Qwen3-32B, Qwen3.8-27B and gpt-oss-20b rise from between 0.40 and 0.77 to between 0.89 and 0.99, against 0.92 for the weakest frontier model. Length accounts for little of the gain: thinking adds 0.5 to 0.9 cited facts and moves causal precision by −0.04 to 0.07.
Recourse falls as reasons multiply, for every group (Figure 4). Under list-all, the small open models reach 0.90 at two goal-fixed reasons and 0.43 at four or more, where a valid explanation must change 4.5 features on average. Reasoning does not remove this. At two goal-fixed reasons the thinking-on models give recourse in 0.94 to 1.00 of denials, at four or more in 0.62 to 0.94; the frontier models average 0.84 there, the weakest 0.76, and six of the nine reasoning runs fall below 0.8.
Figure 4. RV by the number of goal-fixed reasons on the multi-reason denials ( episodes per bucket). It falls in every group, reasoning models included. Bold: mean over a group’s models; faint: one line per model. Thinking on: Qwen3-14B, Qwen3-32B, Qwen3.8-27B, and gpt-oss-20b at high effort. Per-bucket values are in Table 7.
Mayne et al. (2025) find no benefit from inference-time compute for self-generated counterfactuals; here reasoning helps until the independent reasons multiply, which is where Corollary 1 says one reason, or two, stops being enough. More reasoning effort narrows the fall for the strongest models only. Rerun at high effort on these denials, Sonnet gives recourse in 0.93 [0.86, 0.98] and GPT-6 Luna in 0.88, but Haiku in 0.78, up from 0.76 for all three at low effort, where GPT-6 Luna spent no reasoning tokens on a test prompt.
On the eligibility systems we also report , which holds fixed what an applicant cannot change, such as citizenship. It is defined on only 13, 13 and 7 of the 30 sampled SNAP, Title IV and HUD denials, so these rates rest on few episodes: the small open models give actionable recourse in 0.26 of Title IV and 0.36 of HUD denials on average, the open models of 27B to 32B in 0.60 and 0.69, and the frontier models in 0.96 and 0.98.
The strongest open model needs the policy text. Without it, Gemma-4-31B, the one model of 27B or more we ran without the policy, falls from 0.97 to 0.58, and on HUD from 1.00 to 0.15.
Can an LLM judge stand in for the oracle?
We gave three LLM judges (Zheng et al., 2023) the case, the requester’s fixed goal and one explanation, and asked in one prompt for 1 to 5 ratings of whether its reasons are true, whether changing exactly the facts it names would get the request allowed, and its overall quality (Appendix § 13). Asked whether the reasons are correct, the judges rate 84% to 89% of explanations 4 or 5, and the rating separates recourse-valid explanations at AUROC 0.56 to 0.62, 0.51 to 0.55 within a model (Table 2). It barely tracks fidelity either (AUROC 0.54 to 0.56), so a judge asked about correctness passes nearly everything. Asked whether acting on the explanation would work, GPT-6 Luna reaches AUROC 0.97 [0.96, 0.98] and Gemma-4-31B 0.92 [0.90, 0.94], and their mean ratings rank the models almost as the oracle does (Spearman 0.99 and 0.98); Qwen3.8-27B reaches 0.67. In the joint prompt, quality asked right after the outcome question does about as well (Table 2). Asked one question per call, on 700 of the explanations it had already rated, Luna reaches 0.97 [0.95, 0.98] on the outcome question, as in the joint prompt (0.96), but 0.89 [0.86, 0.91] on quality, down from 0.94: part of the quality rating’s strength came from the question asked before it. Leaving out each judge’s own explanations moves no AUROC by more than 0.005. So the correctness question fails, and only the exact check shows which judge is strong enough to ask about the outcome.
| Judge | Correct | 4+ | AUROC | Actionable | AUROC | Quality | AUROC | Difference | AUROC |
| GPT-6 Luna | 84% | 0.62 | 0.97 | 0.95 | 0.02 | ||||
| Gemma-4-31B | 87% | 0.57 | 0.92 | 0.93 | −0.01 | ||||
| Qwen3.8-27B | 89% | 0.56 | 0.67 | 0.84 | −0.16 |
Table 2. LLM judges on default explanations of multi-reason denials, 20 per model and system, 2,800 per judge over 20 models. 4+: share rated 4 or 5 on correct. AUROC: how well each 1 to 5 rating separates recourse-valid explanations from the rest; Difference: actionable minus quality AUROC, paired.
Mitigations
Asking what to change does not help. On the three models of Table 9 (Appendix § 21), a prompt that asks the model to tell the person exactly what would need to change leaves recourse validity within −0.06 to 0.02 of the default.
| Model | Default | Requester | Capped | List-all | Uncapped | Counterfactual | Free-all | Min. change | Reasons |
| Qwen3-14B | 0.40 0.88 | 0.59 0.75 | 0.36 0.94 | 0.79 0.89 | 0.88 0.75 | 0.34 0.92 | 0.71 0.75 | 0.90 0.96 | 0.98 0.98 |
| Gemma-3-12B | 0.23 0.90 | 0.37 0.82 | 0.27 0.90 | 0.84 0.75 | 0.84 0.71 | 0.25 0.93 | 0.81 0.69 | 0.79 0.96 | 0.997 0.86 |
| Llama-3.1-8B | 0.31 0.75 | 0.53 0.70 | 0.28 0.75 | 0.55 0.68 | 0.67 0.67 | 0.29 0.72 | 0.55 0.68 | 0.58 0.90 | 0.87 0.87 |
Table 9. Prompt and content ladder, on the 210 multi-reason denials with causal precision in small type. The oracle-content columns need a pipeline that computes the reasons from the decision function.
Name every rule. One task sentence raises recourse validity for every open model whose default is below 0.90, 14 models, by 0.14 for Gemma-4-12B to 0.61 for Gemma-3-12B (paired sign-flip tests, every Holm-corrected ), and the frontier models by at most 0.05. That closes 35% of the default’s gap for Phi-4-14B to 79% for Gemma-3-12B, under half for four models. List-all raises every small open model, to between 0.55 and 0.91, and for each it gains more than its extra facts buy at random. Reasoning and list-all add up. We ran list-all with thinking on for Qwen3-14B and gpt-oss-20b; it gives recourse in 0.97 to 0.98 of multi-reason denials, above either alone, and in 0.84 to 0.90 of those with four or more reasons, up from 0.62 to 0.79 for the same two models with thinking alone. The cost is length: the small open models’ list-all explanations cite 4.6 facts on average against 3.3 by default, and causal precision falls for ten of the 11 small open models.
Address the person. The default asks for a short note for the audit log. Asked instead to explain the denial to the person who made the request, at any length, nine of the 11 small open models we ran give more recourse, 0.14 more on average, and ten still give less than when asked for every rule, by 0.16 on average; Claude Haiku 4.5 and GPT-6 Luna move by at most 0.002. The audit-log framing explains less than half of the gap. Most of the gain comes from length: on the three ladder models, addressing the person with the 2 to 4 sentence cap kept moves recourse validity by −0.03 to 0.04 from the default, and lifting the cap adds 0.10 to 0.25.
Give the model the reasons. A deployer that owns the decision function can put its reasons in the prompt. Only oracle content is both complete and precise: default and counterfactual prompts give less recourse than either oracle condition, and the list-all prompts less precision. Handed every reason, the three models average 0.95, the most of any rung for each and 0.08 to 0.29 above the minimum change, where the extractor misses a feature named verbatim in 2% to 8% of explanations.
Does the advice work as written?
RV credits any change to the cited features. To test advice as written, a second extraction pass reads the concrete changes an explanation proposes, each a feature and a new value from its domain (Appendix § 13), and the decision function runs on the request with exactly those changes; a change to a goal feature is dropped. We run it on the three ladder models (Qwen3-14B, Gemma-3-12B, Llama-3.1-8B) under the counterfactual and minimum-change prompts of § 6, and on the requester prompt. Under the first, 0.81 to 0.86 of explanations propose a usable change, yet applying it gets the request allowed in only 0.17 to 0.21, against recourse validity of 0.25 to 0.34 on the same explanations. Advice from an explanation that fails RV works in at most 0.02 of cases, and from one that passes in 0.59 to 0.64, so under this prompt RV is close to a necessary condition for working advice and an optimistic bound on it. Handed the minimum change, their advice works in 0.30 to 0.69 of explanations against RV of 0.58 to 0.90, and advice from explanations that fail RV works in up to 0.24. Every one of those 83 cases changes a fact the explanation offers as advice but does not present as a reason, so the reason extractor rightly leaves it out, since RV scores only the reasons an explanation gives. Addressed to the person, the 13 requester models give advice that works as written in 0.08 to 0.56 of explanations, against RV of 0.37 to 0.95; for no model does it reach four fifths of what RV credits.
Related work
Faithfulness, sufficiency and judges. Faithfulness tests ask whether an explanation reflects what drove the explained model ([2] (Jacovi and Goldberg 2020; Parcalabescu and Frank 2024)), and LLM explanations often leave out influential factors or do not suffice for the output ((Turpin et al., 2023; Matton et al., 2025; Nguyen et al., 2026)); Pawar et al. (2026) test whether changing one cited factor changes the LLM’s own decision. Simulatability and utility studies ask whether a reader can predict outputs or solve new cases ([11] (Joshi et al. 2023)). Each measures an explanation against the model that produced it. We measure it against an external rule system, where independent reasons give every single factor zero necessity. LLM judges are lenient and near chance against objective labels ([9] (Zheng et al. 2023; Tan et al. 2025)), and they detect incomplete justifications of rule-governed decisions near chance ((Zhao et al., 2026)); we find the same for the correctness question and show that the outcome question restores the signal. V (2026) score explanations of one synthetic tool-call gate against its exact minimal sufficient reasons; we use that gate only in Appendix C, and its causal recall is one of the correctness signals we find uninformative about recourse.
Counterfactuals and recourse. Counterfactual explanations state what would have to change for a different outcome ((Wachter et al., 2018)), and recourse restricts the change to what a person can act on, including certifying when none exists ((Ustun et al., 2019; Karimi et al., 2022; Kothari et al., 2024)); RV and RV carry these notions to the features a natural-language explanation cites. Reason codes assume a person can act on one reason at a time ([4, 12]). LLMs’ own counterfactuals are often invalid or far from minimal ([18] (Mayne et al. 2025)), less so in larger models ((Kalyvas et al., 2026)), but validity there is judged against the LLM’s own boundary. RGDT-Bench scores justifications for completeness against a reference warrant ((Zhao et al., 2026)). Recourse validity needs no reference and disagrees with completeness both ways: 23% of the recourse-valid explanations we score are incomplete, and a complete explanation whose reasons rest on immutable facts gives no actionable recourse.
Rules, reasoning and formal explanations. Rule benchmarks score whether LLMs reach the right decision from written rules ((Zhou et al., 2025; Gui et al., 2025; Servantez et al., 2026)); access-control systems have long computed what a denied requester could change ((Kapadia et al., 2004; Chowdhary and Sural, 2026)). We fix the decision with an exact gate and score whether an LLM’s explanation of it gives recourse. Reasoning models collapse as puzzles and constraint problems grow ((Shojaee et al., 2025; Lin et al., 2025)); the fall at four or more reasons is the same shape for explanations. Proposition 1 restates the hitting-set duality of diagnosis and formal explanation ((Reiter, 1987; Ignatiev et al., 2019b, 2021)) with goal features fixed, and its shared-feature case is feature necessity ((Huang et al., 2023)). Appendix A gives a longer account.
Conclusion
An explanation of a rule-based denial can cite a true reason and still fail the person who acts on it, and checks of whether the reasons are correct, computed or asked of an LLM judge, cannot see this. Across 20 LLMs on seven ported rule systems, recourse validity tracks capability and reasoning, and falls as independent reasons multiply, for reasoning and frontier models too. Reasoning or asking for every rule closes much of the gap. Explanations of eligibility decisions should be built from the decision function and evaluated by whether acting on them works.
Limitations
Extraction and labels. Every model-side number passes through an LLM extractor, and Qwen3-14B also extracts its own explanations. A second extractor agrees with it at of 0.83 to 0.90 and one author’s labels at 0.83; both move recourse validity by little, but one human annotator is weaker evidence than several, the extractors can share errors, and the pass that reads proposed changes in § 7 has not been checked by hand.
An optimistic construct. RV credits any assignment to the cited features, including ones the explanation never suggests, so it overstates how often acting on an explanation succeeds. § 7 measures how far: recourse-valid counterfactual advice works as written in 0.59 to 0.64 of cases, so the gains it reports for a mitigation overstate gains in working advice. is defined on only 7 to 13 of the 30 sampled denials per eligibility system, since the rest admit no actionable recourse, and whether models tell an applicant that no change can help is not measured.
Ports are abstractions. Each port folds its source’s predicates into a few categorical features. SNAP is modelled for a one-person household. The CFR lags statute, for SNAP after the 2025 reconciliation act and for Title IV after Workforce Pell, and the features abstract over the changed lists. Contexts are counted under the uniform measure, where most eligibility contexts are denials, so no count in Table 1 is a population rate. Goal and immutable features are our choice, and a different choice changes which denials count as multi-reason or as having actionable recourse.
Models and decoding. Open models run 4-bit with thinking off, while the public capability index scores provider-hosted weights; we test 8-bit weights on two models and thinking on four, under the default prompt only, so the other conditions describe models that do not reason before answering. Thinking also raises the token budget (400 new tokens off, 3,000 on), though no default answer of the thinking-off Qwen models comes near the cap (the longest is 180 tokens) and none of gpt-oss-20b’s reaches its budget, so the budget does not explain the gain; three of the four thinking models are Qwen hybrids, and we did not read their reasoning traces, so we cannot say what reasoning changes. We dropped Qwen3.5-9B with thinking on because more than half of its answers ran out of the token budget before the explanation. The matched capability band separates model families as much as capability, and audience is separated from length on three small models only. Frontier models run at low reasoning effort through vendor interfaces without temperature control, and we reran three of them at high effort on the four-or-more-reason denials only; other settings may change their rates. Large and frontier models answer the multi-reason denials only. All prompts and rule systems are in English.
Judges and readers. Three judges cannot map how judge quality varies across models, and judges saw a sample of 20 explanations per model and system. The simulated readers of Appendix § 24 are language models, not people.
Ethics statement
Human labels and simulated readers. One author labelled the extraction sample; no other person took part. The reader study uses language models as readers and collects no data from people.
Data and licences. The Cedar examples and Gatekeeper library are Apache-2.0. The federal eligibility rules are public-domain works of the US government (17 U.S.C. 105). Every context is a synthetic feature combination. The eligibility ports are research abstractions, not eligibility determinations, and must not be used to advise applicants. Frontier models were called with provider data retention off where the interface allows it.
Release. We will release the ports and port scripts, the episodes, every prompt, generation, extraction and judge rating, the oracle, and the analysis code that writes every number in this paper.
Dual use. Disclosing every reason for an access-control denial tells a requester more about the policy, which can help an adversary probe it. For benefit denials the balance runs the other way. An applicant told one of three disqualifying reasons may fix it, reapply and be denied again, at a cost in time and lost benefits. Explanations of eligibility decisions should name every reason, or say that more than one applies.
Extended related work
Faithfulness. Faithfulness asks whether an explanation reflects how a decision was made (Jacovi and Goldberg, 2020). Tests insert reasons an explanation fails to mention or reconstruct the input from the stated reasons [2], weight each mention by how far it moves the explained model (Siegel et al., 2024), or check self-consistency, which is what most such tests measure (Parcalabescu and Frank, 2024; Madsen et al., 2024). Self-explanations are tuned for plausibility at some cost to faithfulness [1], and the concepts an explanation names can differ from those that move the output (Matton et al., 2025). Chain-of-thought omits biasing features and hints the model used [10] (Turpin et al. 2023), models condition on their reasoning to different degrees (Lanham et al., 2023; Tutek et al., 2025; Gur-Arieh et al., 2026), and LLM explanations are selective, naming a subset of causes as people do (Kunz and Kuhlmann, 2024). Each of these compares an explanation with the model that produced it. In our setting the decision is made by an external rule system, so faithfulness to the explainer is not at issue: an explanation can be faithful to the rules and still give no recourse.
Sufficiency, simulatability and usefulness. Rationale benchmarks score comprehensiveness and sufficiency against the explained model [19]; with independent denial reasons, fixing one cited reason changes nothing, a failure these metrics cannot attribute to the explanation. Free-text LLM explanations are mostly insufficient for the model’s own output (Nguyen et al., 2026), and Pawar et al. (2026) test whether each cited factor is necessary or sufficient for an LLM’s decision; under overdetermination every single factor has necessity zero, which Proposition 1 predicts from the policy. Simulatability asks whether an explanation helps a reader predict outputs [11] (Hase and Bansal 2020; Limpijankit et al. 2025). Studies of human utility find that rationales help readers less than accuracy or similarity to gold rationales suggests (Joshi et al., 2023), that few explanation datasets suit decision studies (Hashemi Chaleshtori et al., 2024), and that readers over-rely on wrong explanations (Si et al., 2024). An incomplete but correct explanation carries the same risk, since the reader has no sign that a second reason exists. A position paper asks that self-explanations be judged on actionability (Herrewijnen et al., 2026); recourse validity is one exact form of that judgement.
LLM judges. LLM judges agree with human raters on many generation tasks [13] (Zheng et al. 2023; Liu et al. 2023), but favour the response shown first (Wang et al., 2024), their own outputs (Panickssery et al., 2024) and fluent answers with factual errors (Wu and Aji, 2025), and their reliability varies with the property rated and the wording of the question [5, 6] (Thakur et al. 2025). Against objective labels strong judges are close to chance (Tan et al., 2025), and judges that ignore surface edits miss most edits that change what is being judged [9]. RGDT-Bench finds judges detect incomplete justifications of rule-governed decisions near chance (Zhao et al., 2026). Our correctness question behaves the same way. We add two findings: asking whether acting on the explanation would work restores the signal for a strong judge, and the exact oracle shows which judge is strong enough.
Counterfactuals and recourse. People ask why an outcome happened rather than some other one (Miller, 2019). Counterfactual explanations state the smallest change that yields the preferred outcome (Wachter et al., 2018); methods make them feasible (Poyiadzi et al., 2020) or diverse (Mothilal et al., 2020), and recourse restricts them to what a person can act on, holding immutable features fixed and certifying when no change helps (Ustun et al., 2019; Karimi et al., 2022; Verma et al., 2024; Kothari et al., 2024). Recourse is owed to decision subjects as a matter of agency (Venkatasubramanian and Alfano, 2020). Reason codes rest on the assumption that features can be changed one at a time [4] and often name features whose change cannot alter the decision [12], though in user studies with single-reason denials they help people obtain recourse as often as counterfactuals (Upadhyay et al., 2025). Our failure is the multi-reason case, where any single reason, coded or counterfactual, fails.
Law. The GDPR gives a right to be informed about the logic of automated decisions rather than an explanation of each (Wachter et al., 2017), and US explanation duties address inscrutable models more than scrutable ones (Selbst and Barocas, 2018); our rule systems are fully scrutable and their explanations still fail. Regulation B requires adverse action notices to give the principal reasons and advises that more than four are unlikely to help the applicant [16]; a notice that respects that guidance can still give no recourse when more reasons deny independently.
LLMs and external decisions. LLMs’ own counterfactuals are often invalid or far from minimal, judged against the LLM’s own boundary [18] (Mayne et al. 2025), and their validity rises with scale (Kalyvas et al., 2026). LLMs also narrate the explanations of external classifiers (Zytek et al., 2024) and propose recourse that a classifier checks inside the loop (Tabei, 2026). We score the free-text explanation a deployed model writes with no oracle in the loop. RGDT-Bench scores justifications of rule-governed decisions for completeness against reference warrants (Zhao et al., 2026), and a SAT check can verify that a benefits justification is consistent with statute (Sunny and Sivan-Sevilla, 2026).
Rules and policies. LLMs are tested on applying written rules: fees, trades and tax (Zhou et al., 2025), rule-defined games (Gui et al., 2025), statutes [7] (Holzenberger et al. 2020; Guha et al. 2023) and bankruptcy exemptions generated from symbolic rules (Servantez et al., 2026). Benefit agencies use LLMs to turn SNAP and Medicaid policy into code, which is accurate but incomplete (Kennan et al., 2025); caseworkers follow chatbot answers on SNAP questions (Gosciak et al., 2026); and LLMs draft Cedar policies under a verifier (Vatsa et al., 2026). Guardrail systems explain refusals in natural language (Inan et al., 2023; Rebedea et al., 2023). Access-control systems have long computed the changes that would turn a denial into a grant (Kapadia et al., 2004), recently weighted by how changeable each attribute is [15]. These works score decisions or compute feedback; we score whether an LLM’s explanation of a fixed decision gives recourse.
Reasoning. Open models expose reasoning as a switch, a token budget or an effort level (Yang et al., 2025; OpenAI, 2025). The accuracy of reasoning models falls as puzzles and constraint problems grow (Shojaee et al., 2025; Lin et al., 2025), longer reasoning can lower accuracy on deduction with many constraints (Gema et al., 2025), and chain-of-thought can hurt instruction following (Li et al., 2025). Mayne et al. (2025) find no gain from inference-time compute for self-generated counterfactuals. Against an external rule system we find a
large gain that fades as the number of independent reasons grows.
Formal explanations. Diagnoses are the minimal hitting sets of conflicts (Reiter, 1987), and minimal unsatisfiable and minimal correction subsets are dual in the same way (Lifton and Sakallah, 2008). The duality was carried to classifier explanations (Ignatiev et al., 2019b) and stated between abductive and contrastive explanations (Ignatiev et al., 2021; Marques-Silva and Ignatiev, 2022). Related work defines sufficient and necessary reasons [17], selects among many abductive explanations [3], and asks which features occur in every explanation (Huang et al., 2023), the shared-feature case of Corollary 1. Actual causation defines explanations through causes (Halpern and Pearl, 2005) and spreads responsibility when causes overdetermine an outcome [14]; a denial with several disjoint goal-fixed reasons is that case, where no single cause is pivotal.
Prompt wording
Ported systems, semantics lines: “A request is allowed only if some permit matches and no forbid matches.” (Cedar, eligibility) and “A request is rejected if any violation rule (FORBID) matches.” (Gatekeeper). Default task line: “Write 2-4 sentences for the audit log explaining why this request was denied.” List-all adds: “Name every rule or missing permission that causes the denial, not only one.” Requester replaces the role sentence “You write audit-log explanations for an access-control system.” with “You write explanations for the people whose requests an access-control system decides.” and the task line with “Write an explanation, for the person who made this request, of why it was denied. Make it as long as it needs to be.”; its capped variant uses “Write 2-4 sentences, for the person who made this request, explaining why it was denied.” Free and free-all drop the policy text and its semantics line; the free-all task line is “Name every reason that causes the denial, not only one.” Oracle-all keeps the policy prompt, appends “Each of the following sets of facts is, on its own, enough for this denial:” and one numbered line per minimal reason, such as “(3) enrolled_half_time_higher_ed is yes and meets_student_exemption is no” for a SNAP denial, and its task line is “Explain every one of the reasons listed above, not only one.” Uncapped list-all, counterfactual and minimum change keep the policy prompt. Uncapped list-all replaces the default task line with “Write an explanation for the audit log, as long as it needs to be, of why this request was denied.” Counterfactual replaces it with “Write 2-4 sentences telling the person who made this request exactly what would need to change for it to be allowed.” Minimum change appends “Changing the following facts, with the request itself unchanged, would make this request allowed:” followed by the oracle’s minimum change as “fact (currently value)”, and its task line is “Include the change listed above that would make it allowed.” Eligibility prompts replace the access-control nouns with application nouns. Open-model generations are capped at 400 new tokens in every condition (gpt-oss-20b at 1,500 including reasoning).
The judge prompt gives the request context, the policy, its semantics line, the decision, “The person cannot change what they are asking for:” followed by the goal facts, and the explanation, then asks for three integers from 1 to 5: “correct: Are the reasons the explanation gives true of this case under the policy?”, “actionable: If the person changed exactly the facts the explanation names as reasons, and nothing else, would the request then be allowed?” and “quality: Overall, how good is this explanation for the person who was denied?”, in JSON at temperature 0.
The executed-recourse pass gives Qwen3-14B the request details, each detail’s allowed values and the explanation, and asks it to “List every change to the details that the explanation tells or suggests the person make, each as the detail’s name and its new value, chosen from that detail’s allowed values,” including a change only if the explanation states or clearly implies the new value, in JSON at temperature 0.
A synthetic first-match gate
GlassGate (V, 2026) is an agent tool-call gateway with 10 features and eight block rules ahead of a default allow under first-match. Of its 67,876 block contexts, 24,308 have two or more reasons, and one correct reason leaves the request blocked in 0.70 of its multi-reason denials. We draw 100 multi-reason block episodes, 34 with two reasons, 33 with three and 33 with four. The fired rule is skewed toward R4, first in 45 episodes, and R7 and R8 never fire first. The first-match condition uses the gate’s own header; order-neutral says every matching rule applies independently and the order carries no meaning; list-all keeps that header and asks for every rule that applies. We show the block rules in identity and reversed order, plus one seeded shuffle for Haiku and Qwen3 under first-match. Ranges below are over the six models above the rule in Table 3, or over the five open ones among them where the text says so; the four below the rule are reported in the table only.
| Model | First-match header | Order-neutral header | List-all | |||||
| Valid reason | Fidelity | Two or more | RV | Two or more | RV | Two or more | RV | |
| Llama-3.1-8B | 0.79 | 0.46 | 0.14 | 0.49 | 0.20 | 0.45 | 0.75 | 0.90 |
| Gemma-3-12B | 0.75 | 0.36 | 0.21 | 0.54 | 0.34 | 0.66 | 0.71 | 0.99 |
| Qwen3-14B | 0.91 | 0.56 | 0.28 | 0.50 | 0.80 | 0.89 | 0.91 | 0.995 |
| Phi-4-14B | 0.97 | 0.59 | 0.50 | 0.67 | 0.91 | 0.94 | 0.97 | 0.999 |
| Qwen2.5-14B | 0.99 | 0.52 | 0.69 | 0.78 | 0.91 | 0.88 | 0.93 | 0.97 |
| Claude Haiku 4.5 | 0.99 | 0.95 | 0.50 | 0.75 | 0.98 | 1.00 | 0.99 | 1.00 |
| Gemma-3-4B | 0.92 | 0.64 | 0.51 | 0.85 | 0.62 | 0.85 | 0.95 | 0.99 |
| Llama-3.2-3B | 0.77 | 0.46 | 0.15 | 0.49 | 0.22 | 0.55 | 0.45 | 0.70 |
| Mistral-Nemo-12B | 0.91 | 0.61 | 0.05 | 0.51 | 0.17 | 0.53 | 0.86 | 0.95 |
| OLMo-2-13B | 0.72 | 0.29 | 0.14 | 0.48 | 0.22 | 0.52 | 0.76 | 0.86 |
Table 3. Synthetic gate, 100 multi-reason block episodes, 1,000 explanations per cell (1,500 for Haiku and Qwen3 under first-match). Valid reason: share that disclose at least one complete reason. Fidelity: share that disclose the reason of the rule that fires first in the order shown. Two or more: share that disclose at least two. RV: recourse-valid share among explanations that cite at least one feature. Episode-level 95% bootstrap intervals are at most 0.13 wide for two or more and 0.14 for RV.
Under first-match, the six models above the rule disclose a complete valid reason in 75% of explanations for Gemma-3-12B and up to 99% for Qwen2.5-14B, yet only 0.49 of the explanations that cite anything are recourse-valid for Llama-3.1-8B and 0.78 for Qwen2.5-14B, so 22% to 51% leave the requester blocked (Table 3). Size does not predict it: Gemma-3-4B reaches 0.85 against 0.54 for Gemma-3-12B, and Llama-3.2-3B matches Llama-3.1-8B. Each of the six discloses a valid reason more often than it gives recourse. The five open models among them disclose the reason of the rule that fires first in only 36% to 59% of explanations and Haiku in 95%, while Haiku gives recourse in 0.75. The gap tracks how often a model discloses two or more reasons: Llama-3.1-8B does so in 14% of explanations and has the lowest recourse validity of the six, Qwen2.5-14B in 69% and the highest. Rule order changes recourse validity although it never changes a decision: identity and reversed order give 0.32 and 0.65 for Llama-3.1-8B, 0.45 and 0.67 for Qwen3-14B, 0.48 and 0.61 for Gemma-3-12B and 0.71 and 0.87 for Haiku.
The order-neutral header and one instruction together close most of the gap for the five open models. Asked to name every rule, they disclose two or more reasons in 71% to 97% of explanations, and recourse validity rises to 0.90 to 0.999; the instruction’s own increment over the header alone is 0.06 to 0.11 for Qwen3-14B, Phi-4-14B and Qwen2.5-14B and 0.33 to 0.45 for Gemma-3-12B and Llama-3.1-8B. The header alone helps four of the five: Qwen3-14B from 0.50 to 0.89, Phi-4-14B from 0.67 to 0.94, Qwen2.5-14B from 0.78 to 0.88 and, least, Gemma-3-12B from 0.54 to 0.66; not Llama-3.1-8B, from 0.49 to 0.45. Its reasons share features, so list-all recourse on this gate stays at or above 0.94 from two reasons up, and at four or more reasons the smallest change has 2.2 features against 4.5 on the ported systems.
Which reason models pick
When several reasons are valid, which one does a model disclose? Under the first-match header, Haiku discloses the first listed valid rule in 0.96 of single-reason explanations against a chance rate of 0.39, and its most-disclosed rule switches in 0.92 of episodes when the order is reversed (Table 4); under that header the first listed valid rule is the deciding rule. The five open models follow position far less, from 0.36 for Phi-4-14B to 0.55 for Qwen3-14B.
| Model | First-match header | Order-neutral header | ||
| First listed | Switch | Concentration | ||
| Llama-3.1-8B | 0.54 (0.36) | 0.47 | 0.52 (0.29) | <0.001 |
| Gemma-3-12B | 0.39 (0.37) | 0.28 | 0.44 (0.23) | <0.001 |
| Qwen3-14B | 0.55 (0.38) | 0.30 | 0.30 (0.11) | <0.001 |
| Phi-4-14B | 0.36 (0.40) | 0.09 | 0.08 (0.08) | 0.77 |
| Qwen2.5-14B | 0.41 (0.44) | 0.16 | 0.20 (0.10) | <0.001 |
| Claude Haiku 4.5 | 0.96 (0.39) | 0.92 | 0.01 (0.07) | <0.001 |
Table 4. Position and concentration on the synthetic gate. First listed: among explanations that disclose exactly one valid rule, the share that disclose the rule listed first, chance in parentheses. Switch: share of episodes whose most-disclosed rule changes when the order is reversed. Concentration: one minus normalised entropy of disclosed reasons per cell, uniform null in parentheses, paired permutation .
Under the order-neutral header, three of the five open models disclose two or more reasons in most explanations, 0.80 to 0.91, but Gemma-3-12B and Llama-3.1-8B in only 34% and 20%. Asked to list every rule, Gemma-3-12B and Llama-3.1-8B disclose two or more 71% and 75% of the time, so they can list the reasons but by default do not; they are also the two models the header helps least. Their disclosures concentrate on one reason well above the uniform null, as do those of Qwen3-14B and Qwen2.5-14B.
The choice is shared across models. Taking each episode’s most-disclosed single reason under the order-neutral header, Gemma-3-12B, Llama-3.1-8B and Qwen3-14B agree pairwise on 0.77, 0.82 and 0.73 of shared episodes, against 0.37, 0.43 and 0.42 under uniform choice; Haiku agrees with them near chance under first-match, 0.30 to 0.47. The ten pairs of the five open models name the same single reason with probability 0.60 to 0.80 against 0.37 to 0.48 by chance.
Models favour salient reasons over short ones. The five open models disclose a smallest reason in only 45% of single-reason explanations for Llama-3.1-8B and 73% for Qwen2.5-14B, below the 74% to 86% expected under uniform choice, and the rule disclosed most is R8, writes to unknown domains, for every one of them, followed for Gemma-3-12B and Llama-3.1-8B by R7, bulk financial export by an intern; neither fires first in any episode. With R8 and R7 moved to the middle of the order-neutral list, single-reason explanations still name them most: Gemma-3-12B 0.38 and 0.26, Llama-3.1-8B 0.23 and 0.28, Qwen3-14B 0.54 and 0.14 (R8 then R7; ).
Per-model results
Table 5 gives fidelity, causal precision, facts cited and recourse validity for every model on the 210 multi-reason denials, under the default prompt and list-all.
| System | Default Fidelity | Default Two+ | Default | List-all | Control |
| Cedar doc-cloud | 0.94 | 0.37 | 0.71 | 0.92 | 0.97 |
| Cedar github | 0.58 | 0.19 | 0.91 | 0.98 | 0.92 |
| OPA repos pod | 0.87 | 0.29 | 0.49 | 0.84 | 0.91 |
| OPA privileged pod | 0.86 | 0.23 | 0.50 | 0.92 | 0.81 |
| SNAP | 0.96 | 0.61 | 0.41 | 0.78 | 0.85 |
| Title IV | 0.89 | 0.58 | 0.19 | 0.47 | 0.84 |
| HUD | 0.99 | 0.76 | 0.20 | 0.58 | 0.91 |
| Pooled | 0.87 | 0.43 | 0.48 | 0.78 | 0.88 |
Table 6. Per system, mean over the 11 small open models. Two+: share that disclose two or more reasons. Control: the 30 single-reason denials per system.
Per-system results
Table 6 gives the per-system means behind § 5.1. For the small open models the policy text cuts both ways: without it, default recourse changes by for Gemma-4-12B to for Gemma-3-12B, while fidelity drops by to $0.30. The supplementary file supplementary_real_cells.csv gives per-model fidelity, coverage, two or more disclosed reasons, goal-free, goal-fixed and actionable recourse validity, causal precision and facts cited for every system, condition and stratum, 1,172 cells, each as a mean of per-episode means with its 95% bootstrap interval over episodes.
Recourse by number of reasons
The values behind Figure 4 are in Table 7, with the mean size of the smallest change in each bucket.
| Small open Default | Small open List-all | Thinking Default | Frontier Default | Change | |
| 1 | 0.89 | 0.96 | 1.00 | 1.00 | 1.0 |
| 2 | 0.59 | 0.90 | 0.97 | 0.99 | 1.9 |
| 3 | 0.33 | 0.71 | 0.92 | 0.98 | 2.9 |
| 4+ | 0.12 | 0.43 | 0.76 | 0.84 | 4.5 |
Table 7. by number of goal-fixed reasons, the values behind Figure 4; mean over the 11 small open models on both strata (221, 102, 50 and 34 episodes, the 407 of 420 with a defined verdict) and over the four thinking-on and five frontier models on the multi-reason stratum. Change: mean size of a minimum hitting set of .
Signal AUROCs
Table 8 gives the AUROCs behind Figure 3, pooled, within model and within episode.
Fidelity against recourse by model
Across models, fidelity and recourse validity are uncorrelated, Spearman 0.14 (Figure 5).
Figure 5. Fidelity against goal-fixed recourse validity on the 210 multi-reason denials, one point per model, default prompt (Spearman 0.14 over 20 models). Grey lines rise to each model’s list-all value.
Prompt and content ladder
Lifting the sentence limit adds 0.003 to 0.12 to list-all but lowers precision for every model.
Runs
Table 10 lists every run. Explanations without a parseable extraction are dropped and counted, 203 of 152,754. The 210 episodes are the multi-reason stratum; the 420 add the single-reason controls.
| Conditions | Episodes | Models |
| Open, 8B to 21B | ||
| default, list-all | 420 | all 11 |
| requester | 210 | all 11 |
| requester, 2 to 4 sentences | 210 | Qwen3-14B, Gemma-3-12B, Llama-3.1-8B |
| free, free-all | 420 | 10, not gpt-oss-20b |
| uncapped, counterfactual, min. change | 210 | Qwen3-14B, Gemma-3-12B, Llama-3.1-8B |
| every reason | 420 | the same three |
| Open, 27B to 32B | ||
| default, list-all | 210 | all 4 |
| free, free-all | 210 | Gemma-4-31B |
| Frontier | ||
| default, list-all | 210 | all 5 |
| requester | 210 | Haiku, GPT-6 Luna |
| Settings | ||
| default, list-all, 8-bit | 210 | Llama-3.1-8B, Gemma-3-12B |
| default, thinking on | 210 | Qwen3-14B, Qwen3-32B, Qwen3.8-27B, gpt-oss-20b |
| list-all, thinking on | 210 | Qwen3-14B, gpt-oss-20b |
| default, high effort | 34 | Haiku, Sonnet, GPT-6 Luna |
Table 10. Every generation run, 5 samples per episode and condition. The high-effort reruns cover only the 34 multi-reason denials with four or more goal-fixed reasons. The 20 models produced 141,750 explanations: 118,650 from the small open models, 10,500 from the large and 12,600 from the frontier models; the settings runs are counted apart.
Capability index
Table 11 lists the index values behind § 5.2. The index scores provider-hosted weights and settings; our open models run 4-bit with thinking off.
| Model | ECI | |
| Llama-3.1-8B | 116.6 | 0.31 |
| Mistral-Nemo-12B | 118.7 | 0.24 |
| Gemma-3-12B | 123.5 | 0.23 |
| Gemma-3-27B | 130.0 | 0.44 |
| Phi-4-14B | 130.4 | 0.57 |
| gpt-oss-20b | 137.8 | 0.55 |
| Qwen3-14B | 138.2 | 0.40 |
| Qwen3-32B | 138.5 | 0.54 |
| Qwen3.5-9B | 139.5 | 0.47 |
| Claude Haiku 4.5 | 142.4 | 0.96 |
| Gemma-4-31B | 142.7 | 0.97 |
| GPT-6 Luna | 156.3 | 0.92 |
| Claude Sonnet 5.5 | 165.0 | 0.94 |
| GPT-6.1 Sol | 166.1 | 0.99 |
| Claude Opus 5.5 | 167.3 | 0.99 |
Table 11. Epoch Capabilities Index values ([20]), read on 6 October 2026, and default goal-fixed recourse validity. Models without a published value are left out of the correlations.
Simulated readers
LLM readers got stimuli written for a human study, and the oracle scored their answers. Each reader model saw every item under every condition five times at temperature 0.8, with the facts in a fresh random order, and chose the facts it would change to get the decision reversed while keeping its goal, or “I cannot tell”. The 120 items are plain-language denials with two or more pairwise feature-disjoint goal-fixed reasons, 30 each from SNAP, HUD, Title IV and Cedar document-cloud. The one explanation names one correct reason, all every reason, and none only that the request was denied. The study credits a choice of facts whatever values a reader would set, so it does not show that a reader finds a working change, and it is not evidence about how people read or act on explanations. The reader models are Gemma-3-12B, Llama-3.1-8B, Mistral-Nemo-12B, Phi-4-14B, Qwen2.5-14B and Qwen3-14B. Under one, 0.73 of answers change only facts of the named reason. The chosen facts meet every reason in 0.01 to 0.04 of trials under one and 0.11 to 0.96 under all, so every reader does better with every reason than any does with one, and all differs from one by 0.10 to 0.93 per reader, significant after Holm correction for six.
Immutable features
Each port file lists the features that actionable recourse holds fixed with the goal, chosen from the source text. Document-cloud: is_owner, as the policy compares the principal with the document’s owner and no action changes the owner. GitHub: is_reporter, as who filed an issue is fixed when it is filed. SNAP: elderly_or_disabled_member and citizenship, age 60 or disability and citizenship or eligible alien status (7 CFR 273.9(a) and 273.4(a); USDA Food and Nutrition Administration, 2026). HUD: methamphetamine_production_conviction, lifetime_sex_offender_registrant and immigration_status, standards that “permanently prohibit admission”, “a lifetime registration requirement” and eligible immigration status (24 CFR 960.204(a)(3),
(a)(4) and 5.506; U.S. Department of Housing and Urban Development, 2026. Title IV: has_baccalaureate_degree and citizenship, no baccalaureate degree for a Pell Grant and citizenship or eligible noncitizen status (34 CFR 690.75, 668.32(c) and 668.33(a); U.S. Department of Education, 2026). HUD’s three-year drug-eviction bar lapses and has a rehabilitation exception, and Title IV’s incarceration feature includes enrolling in an eligible prison education program, so both stay changeable. One port deviation is known: the document-cloud port
applies the private-document forbid to document creation, where the source’s owner guard means it never fires; it affects one control episode and no multi-reason episode.
References
- [1]Chirag Agarwal, Sree Harsha Tanneru, and Himabindu Lakkaraju. 2024. Faithfulness vs. plausibility: On the (un)reliability of explanations from large language models. Preprint, arXiv:2402.04614.
- [2]Pepa Atanasova, Oana-Maria Camburu, Christina Lioma, Thomas Lukasiewicz, Jakob Grue Simonsen, and Isabelle Augenstein. 2023. Faithfulness tests for natural language explanations. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 283–294, Toronto, Canada. Association for Computational Linguistics.
- [3]Gilles Audemard, Steve Bellart, Louenas Bounia, Frederic Koriche, Jean-Marie Lagniez, and Pierre Marquis. 2022. On preferred abductive explanations for decision trees and random forests. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22, pages 643–650.DOI
- [5]Anna Bavaresco, Raffaella Bernardi, Leonardo Bertolazzi, Desmond Elliott, Raquel Fernández, Albert Gatt, Esam Ghaleb, Mario Giulianelli, Michael Hanna, Alexander Koller, Andre Martins, Philipp Mondorf, Vera Neplnbroek, Sandro Pezzele, Barbara Plank, David Schlangen, Alessandro Suglia, Aditya K Surikuchi, Ece Takmaz, and Alberto Testoni. 2025. LLMs instead of human judges? a large scale empirical study across 20 NLP evaluation tasks. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 238–255, Vienna, Austria. Association for Computational Linguistics.
- [6]Rohith Reddy Bellibattlu, Edward Raff, and Wenbin Zhang. 2026. JudgeSense: A benchmark for prompt sensitivity in LLM-as-a-judge systems. Preprint, arXiv:2604.23478.
- [7]Andrew Blair-Stanek, Nils Holzenberger, and Benjamin Van Durme. 2023. Can GPT-3 perform statutory reasoning? In Proceedings of the Nineteenth International Conference on Artificial Intelligence and Law, pages 22–31. ACM.
- [8]Cedar Policy. 2026. cedar-policy/cedar-examples. https://github.com/cedar-policy/cedar-examples. Apache-2.0. Files cedar-example-use-cases/document_cloud/policies.cedar and cedar-example-use-cases/github_example/policies.cedar. Accessed 5 October 2026.
- [9]Jianlin Chen, Wenhui Chen, Ziyao Lin, and Chi Man Vong. 2026. A judge should know what changed: Construct validity for LLM-as-a-judge evaluation. Preprint, arXiv:2608.024419.
- [10]Yanda Chen, Joe Benton, Ansh Radhakrishnan, Jonathan Uesato, Carson Denison, John Schulman, Arushi Somani, Peter Hase, Misha Wagner, Fabien Roger, Vlad Mikulik, Samuel R. Bowman, Jan Leike, Jared Kaplan, and Ethan Perez. 2025. Reasoning models don’t always say what they think. Preprint, arXiv:2505.05410.arxiv.org/abs/2505.05410
- [11]Yanda Chen, Ruiqi Zhong, Narutatsu Ri, Chen Zhao, He He, Jacob Steinhardt, Zhou Yu, and Kathleen McKeown. 2024. Do models explain themselves? Counterfactual simulatability of natural language explanations. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pages 7880–7994. PMLR.arxiv.org/abs/2307.08678
- [12]Seung Hyun Cheon, Anneke Wernerfelt, Sorelle Friedler, and Berk Ustun. 2025. Feature responsiveness scores: Model-agnostic explanations for recourse. In International Conference on Learning Representations, pages 41382–41413.
- [13]Cheng-Han Chiang and Hung-yi Lee. 2023. Can large language models be an alternative to human evaluations? In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15607–15631, Toronto, Canada. Association for Computational Linguistics.arxiv.org/abs/2305.01937
- [14]Hana Chockler and Joseph Y. Halpern. 2004. Responsibility and blame: A structural-model approach. Journal of Artificial Intelligence Research, 22:93–115.
- [15]Shanampudi Pranaya Chowdary and Shamik Sural. 2026. EXTree: Towards supporting explainability in attribute-based access control. Preprint, arXiv:2604.12850.arxiv.org/abs/2604.12850
- [16]Consumer Financial Protection Bureau. 2026. Regulation B (Equal Credit Opportunity Act), 12 CFR 1002.9 and supplement I, Official interpretation of paragraph 9(b)(2). https://www.consumerfinance.gov/rules-policy/regulations/1002/9/. Accessed 8 October 2026.
- [17]Adnan Darwiche and Auguste Hirth. 2020. On the reasons behind decisions. In ECAI 2020: 24th European Conference on Artificial Intelligence, volume 325 of Frontiers in Artificial Intelligence and Applications, pages 712–720. IOS Press.
- [18]Zahra Dehghanighobadi, Asia Fischer, and Muhammad Bilal Zafar. 2025. Can LLMs explain themselves counterfactually? In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 7787–7815, Suzhou, China. Association for Computational Linguistics.arxiv.org/abs/2502.18156
- [19]Jay DeYoung, Sarthak Jain, Nazneen Fatema Rajani, Eric Lehman, Caiming Xiong, Richard Socher, and Byron C. Wallace. 2020. ERASER: A benchmark to evaluate rationalized NLP models. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4443–4458.arxiv.org/abs/1911.03429
- [20]Epoch AI. 2025. Epoch Capabilities Index. https://epoch.ai/eci. Accessed 6 October 2026.


