PaperlayerView exact revision on arXiv

Source-linked AI summary

No Judgment Without a Reason: Counterfactual Receipts for Versioned AI Evaluators

Ye Chen, Weining Zhang

arXiv:2608.20938v1cs.AI

TL;DR

AI evaluators can gate actions, route reviews, and supply training feedback, so revised judgments need accounts based on grounds, norms, and authority rather than labels alone. The paper formalizes counterfactual judgment receipts and finds that structured outputs can appear nearly perfect while failing equivalent source orders and unseen multi-source transitions.

  • Problem

    AI evaluators operate in oversight pipelines where their labels decide release, routing, and training feedback, making the reliability and scalable supervision of these evaluators a live concern.

  • Method

    The paper models judgment transitions through grounds, norms, and authority, using executable counterfactual hybrids to recover minimal accounts and distinguishing receipt prediction from certification by execution.

  • Results

    Structured outputs can look nearly perfect on locked tests while failing equivalent source orders and inverse transitions; under unseen multi-source composition, verdicts survive while receipts approach zero.

  • Takeaways & Limitations

    Reason-aware evaluation should separate prediction from certification and assess transformation consistency alongside standard accuracy when auditing evaluators.

  • Takeaways & Limitations

    ReasonBench uses executable adjudicators, while real institutions involve tacit practice, conflicting records, discretionary exceptions, and contested authority.

Abstract

from arXiv · show

Evaluators often produce correct labels via flawed reasoning, a critical failure for agentic systems gating actions, routing reviews, or supplying training feedback. Standard evaluation only verifies final label correctness, ignoring whether judgment changes stem from valid evidence, consistent rules, or proper rule applicability. We formalize evaluator reasoning accountability via three core sources: grounds, norms, and authority. Varying these sources yields an eight-cell counterfactual judgment cube to characterize judgment updates. We define judgment receipts as minimal source replacement sets that reproduce revised verdicts to explain judgment transitions. We derive certification cost bounds for black-box evaluators and present ReasonBench, a policy and logical reasoning benchmark with verifiable receipts covering 19,520 cases and 7,200 controls. In frozen evaluations, Qwen3-1.7B reaches 98.41% receipt accuracy, while cube prediction scores 96.99%, a consistent 1.42-point drop validated by Qwen3-0.6B replication. Strong standard accuracy masks severe robustness flaws. Meaning-preserving source permutations reduce valid receipt recovery to 54.8% and 49.2% for direct and cube prediction. Models trained on simple single-source changes retain 93.75% verdict accuracy but recover only 7.16% of receipts for complex multi-source updates. Permutation retraining boosts consistency to 96.6% yet worsens cube prediction deficits. Structured counterfactual supervision fails to guarantee robust reasoning. We show reason-aware evaluation must decouple prediction and certification, reporting transformation consistency alongside standard accuracy for trustworthy evaluator auditing.

1 A judgment and its account

Evaluator judgments can change for reasons that final-label accuracy does not reveal. The paper frames each change as an accountability problem over evidence, rules, and authority, requiring complete minimal accounts of the transition.

  • 1 A judgment and its account: Operational evaluators grant permission, route attention, and generate training feedback, so their reasoning reliability matters beyond label correctness.The paper positions evaluator outputs inside release, review-routing, and training-feedback loops.
  • 1 A judgment and its account: A deny-to-approve transition may reflect changed grounds, amended norms, or newly applicable authority, distinctions erased by final-label benchmarks.Free-text explanations can describe the revised case without identifying which declared change produced the new verdict.
  • 1 A judgment and its account: The accountability vocabulary targets fluent rationales that may fail to track computation and version changes whose sufficient causes are otherwise untested.The paper names attribution failures such as misattribution, drift, and inertia.
  • 1 A judgment and its account: The paper defines reason as a typed, counterfactually executable replacement in declared sources, not a mental state or persuasive paragraph.A changed judgment incurs a debt of reasons discharged by listing every minimal sufficient replacement.
  • 1 A judgment and its account: The framework separates prediction from certification: grounds, norms, and authority form a judgment cube, while executed counterfactuals certify predicted accounts.The contributions also include paired consistency and the ReasonBench benchmark with executable states and exact receipts.
  • 1 A judgment and its account: A receipt establishes sufficiency only inside an executable evaluator; it does not establish that rules are just, evidence true, or authority legitimate.The paper treats inspectable accounts as the start, not the whole, of accountability.

2 Evaluator states

Evaluator states expose three typed interfaces: accepted grounds, governing norms, and authority over applicability. This decomposition supports executable counterfactual accounting while making the interface’s scope and assumptions explicit.

  • 2 Evaluator states: An evaluator state is represented as I_t = (G_t, N_t, A_t), combining grounds, norms, and authority for version t.The three components are the declared local sources of judgment.
  • 2 Evaluator states: Grounds contain accepted evidence, norms map grounds to dispositions, and authority records which norms govern particular cases.The interface corresponds to evidence inputs, decision rules, and applicability conditions.
  • 2 Evaluator states: For deterministic adjudicators, differing judgments require at least one changed local source, although changed sources may leave the verdict unchanged.No-op revisions therefore serve as invariance controls against flipping merely because an amendment appears.
  • 2 Evaluator states: The three-part state is an interface rather than an ontology, and deployments may expose more or fewer source types at exponential cost in the number of types.Tacit practice, conflicting records, and contested authority must first be represented at an explicit interface.
  • 2.2 Authority as a jurisdiction gate: Authority is distinct from a norm: it determines whether a norm may govern, while the norm determines what follows on the merits.Collapsing them turns lack of jurisdiction into a merits decision.
  • 2.2 Authority as a jurisdiction gate: The benchmark represents absent authority as referral, so non-jurisdiction escalates the case instead of producing a merits ruling.This is the paper’s benchmark-specific treatment, not a broader legal theory.

3 Judgment receipts

Judgment receipts enumerate minimal source replacements that reproduce a revised verdict, while the judgment cube executes every old/new hybrid. The paper separates these audit objects from model predictions and measures consistency under transformations.

  • 3 Judgment receipts: The judgment cube evaluates all 2^3 old/new hybrids of grounds, norms, and authority, including interior cells that reveal joint replacement effects.Its endpoints show whether the verdict changed; interior cells identify sufficient single-source or combined replacements.
  • 3.2 Complete minimal accounts: A judgment receipt is the complete family of inclusion-minimal source replacement sets whose execution reproduces the revised verdict.If grounds and norms independently suffice, both {G} and {N} remain in the receipt.
  • 3.2 Complete minimal accounts: Receipts preserve multiplicity and transition-level interventions, making them formally distinct from single-decision prime-implicant explanations.They use typed source replacements and are weaker than implicants without monotonicity.
  • 3.3 Exact enumeration and its cost: With three sources, a receipt family contains at most three minimal accounts, and no receipt exists when endpoint verdicts agree.The bound follows from the antichain structure of minimal sufficient subsets.
  • 3.3 Exact enumeration and its cost: Exact black-box certification requires querying all 2^k − 2 interior hybrids in the worst case when endpoint labels differ and no structure is assumed.The lower bound concerns certification, not whether a model should emit every cube cell.
  • 3.4 Prediction is not certification: A model-generated receipt or cube is only a prediction; certification requires executing the relevant counterfactuals through a trusted adjudicator.A predicted cube may enforce internal consistency but cannot provide external correctness.
  • 3.5 Paired consistency under controlled transformations: Paired consistency tests whether transformed predictions obey a known relation, exposing errors that unpaired accuracy cannot detect.For source-order permutations or irrelevant distractors, endpoints and receipts should remain identical; violated relations prove at least one paired prediction is wrong.

4 Failure accounting

The paper separates evaluator change failures into drift, inertia, and misattribution, showing that attribution errors can persist even when endpoint verdicts are correct. It also frames reason debt as selective attribution risk requiring coverage and gold-change recall.

  • Drift is a false-positive change, inertia is a false-negative change, and misattribution is an incorrect receipt on a gold change.
  • Misattribution can persist despite correct old and revised endpoint verdicts.The experiments therefore report receipt error conditional on both endpoint verdicts being correct.
  • Reason debt is selective risk for the attribution task and is undefined when the accepted set is empty.
  • Reason debt must be reported with accepted coverage and gold-change recall because universal abstention can simulate perfect accounting while preserving inertia.
  • An evaluator matching revised verdicts on changed cases and preserving old verdicts elsewhere matches the revised evaluator everywhere.

5 ReasonBench

ReasonBench makes evaluator explanations executable across organizational policy and logical reasoning cases, with controlled source mutations and paired controls. Its frozen split and audits constrain leakage while preserving diverse receipt structures and distribution tests.

  • ReasonBench executes candidate explanations by providing old and revised evaluator states, running all eight hybrids, and storing exact receipt families.
  • Organizational stratum: The organizational stratum translates 45 audited clauses into typed executable policy rules with structured grounds and authority scopes.
  • Logical stratum: The logical stratum independently samples 800 compatible positive-rule worlds and uses facts, implications, and an explicit authority mask.
  • Splits and controls: The locked test contains 2,400 examples from 102 independent units, with prompts never treated as independent observations.
  • Splits and controls: Among main cases, 10,888 have gold endpoint changes and 8,632 are stable; minimum receipt sizes range from one to three, with 3,470 receipt-ambiguous cases.
  • Splits and controls: Each locked example receives inverse-transition, source-order permutation, and irrelevant-distractor controls, totaling 7,200 transformations without adding independent observations.
  • Distribution tests: The frozen protocol tests operator, linguistic, and proof-depth shifts, while a post-freeze cardinality experiment evaluates generalization from single-source to multi-source changes.
  • Audit: An independent pre-freeze audit found zero label errors, zero source-unit leakage, and zero prompt overlap, while reproducing all 7,200 controls.

6 Experimental design

The experiments compare direct and cube-based targets under frozen training, parsing, outcome, uncertainty, and selective-acceptance protocols. Cube outputs are coherent by construction, so correctness and certification remain separate questions.

  • Learning targets: All learned methods receive identical old and revised grounds, norms, and authority descriptions, differing only in serialized targets.
  • Learning targets: Direct verdict, direct receipt, rationale receipt, and judgment cube targets produce verdicts, receipts, rationales, or eight ordered cell outputs respectively.
  • Learning targets: Cube-derived receipts cannot contradict predicted cells, but this coherence is a formatting property rather than evidence that the cells are correct.
  • Primary comparison: The primary comparison tests whether cube supervision improves revised-verdict and exact receipt-family accuracy over direct receipt supervision, especially for multi-source and ambiguous cases.
  • Training: The primary backbone is Qwen3-1.7B with LoRA adaptation, using five seeds with identical data, optimization, and decoding.
  • Outcomes: The co-primary outcomes are revised-verdict accuracy and exact receipt-family accuracy, complemented by endpoint, cube, error, consistency, and selective-debt diagnostics.
  • Inference: Uncertainty uses hierarchical bootstrap resampling over five paired seeds and independent units, with Holm adjustment for the two locked-test co-primary comparisons.
  • Selective evaluation: Selective thresholds are fitted on calibration data across 20%–100% accepted-change coverage, and locked-test debt is reported with realized coverage and correctly accepted gold-change recall.

7 Results

The locked-test results show that strong ordinary accuracy does not establish correct receipt reasoning, and full-cube supervision does not improve prediction over direct receipts. Performance varies by structure and distribution shift, while paired consistency controls expose substantial presentation sensitivity.

  • Surface diagnostics: 53.8% revised-verdict and 44.2% exact receipt accuracy come from the majority diagnostic, while character TF-IDF reaches 90.0% and 61.7%.The character model uses more training data than the neural runs, and its lexical signal does not approach the learned receipt results.
  • Full-cube supervision: −1.42 percentage points is the cube-minus-direct effect on exact receipt accuracy, rejecting the frozen superiority claim for cube supervision.The 95% interval is [−2.87, −0.30] with Holm-adjusted p = .002.
  • Structural subgroups: −5.69 points is the cube-minus-direct receipt effect on six organizational locked clauses, compared with −0.07 points on the logical locked stratum.The six organizational clauses can flag a problem but cannot estimate how organizational policies behave in general.
  • Verdict–receipt separation: 97.30% receipt accuracy is achieved by direct receipt on gold changes per locked seed, while the cube’s revised-verdict gap is 0.90 points.The changed-case analysis separates endpoint correctness from receipt attribution.
  • Verdict–receipt separation: 1.66% of eligible direct seed-example observations and 1.85% of cube observations have wrong receipts despite both endpoints being correct.Correct revised and old verdicts therefore do not entail correct attribution of the source changes.

7.6 Permutation-augmented training (post-hoc)

Permutation augmentation largely restores order consistency, but its benefits split sharply by task stratum and do not eliminate the cube’s weaknesses. The post-hoc evidence remains exploratory, with costs concentrated in organizational clauses.

  • Stratified effects: 99.40% direct and 96.82% cube order consistency on logical cases contrasts with 87.79% and 67.94% on organizational clauses.The augmented cube still violates the required relation on roughly a third of organizational pairs.
  • Consistency repair: +40.76 points for direct receipt and +40.67 for cube order consistency shows that augmentation restores the targeted relation.The implied lower bound on mean pairwise error falls from 22.1% to 1.7% for direct prediction.
  • Target trade-offs: −23.73 points separates augmented cube from direct receipt on organizational clauses, versus −0.02 on logical worlds.The joint target degrades most where it was already weakest, and six clauses cannot support a general organizational-policy claim.
  • Interpretation: 96.6% direct order consistency after retraining indicates that frozen near-50% consistency was largely a training-distribution property.The untargeted inverse relation remains broken, so repairing one transformation does not establish general consistency.
  • Selective accounting: 81.97% accepted-change coverage with 0.27% reason debt for direct receipt and 85.54% with 0.16% for cube show selective acceptance can concentrate errors.Near full coverage, debt rises to 2.26% and 3.48%, while correct gold-change recall reaches 98.60% and 94.58%.

8 What the experiment changes

The experiments separate receipt prediction, transformation consistency, and executable certification rather than treating them as one capability. High locked accuracy is narrower than robust reasoning, while structured cube supervision does not reliably improve generalization.

  • Certification: The executed cube identifies the complete minimal receipt family exactly and is worst-case query-optimal for unrestricted black-box certification.This guarantee assumes a trusted adjudicator and declared interface.
  • Prediction targets: A larger joint cube target is not automatically a more robust learner because any cell error can corrupt the derived receipt or minimality.Direct receipt training compresses the target to the scored object and can exploit benchmark regularities more efficiently.
  • What accuracy means: 98.4% locked receipt accuracy measures prediction on a familiar rendering, not presentation-independent understanding of grounds, norms, and authority.Order-consistency rates near 50% provide evidence against the stronger interpretation.
  • Operational implication: A learned model can propose receipts and route audits, but trusted execution must carry the auditable commitment.The experiments do not estimate deployment latency, audit scheduling, reviewer time, or saved cost.
  • Evaluation design: Fidelity, consistency, and certification are distinct questions, and a high value on any one implies neither of the others.Locked accuracy, paired transformations, and executed cubes respectively answer these separate questions.
  • Experimental revision: Direct receipt wins the locked comparison, while both targets fail order and inverse controls and cube supervision shows no advantage at higher receipt complexity.The post-freeze composition gain is confined to a narrow organizational subset and does not reverse the primary verdict.

9 Related work

The paper connects receipts to explanation, causality, metamorphic testing, selective prediction, and auditability while distinguishing its version-transition semantics from those neighboring frameworks.

  • Evaluator biases: Source-order fragility is a version-transition analogue of position and formatting biases, measured against an executable gold account.The comparison links evaluator robustness to known shortcut features and context-position sensitivity.
  • Explanation and counterfactuals: Receipts retain minimal sufficiency but explain transitions between evaluator versions through typed source replacements rather than token subsets.The complete minimal family is retained, and the intervention set is executed through a reference adjudicator.
  • Causality and provenance: A receipt is a minimal replay of a revised judgment from a specified old state, not a claim about metaphysical, legal, or moral causation.This distinguishes the framework from actual-causality and provenance objects.
  • Selective prediction: Reason debt adapts selective prediction’s coverage-error trade-off to accepted evaluator changes while retaining gold-change recall.Low debt cannot be achieved simply by rejecting every amended case.
  • Metamorphic evaluation: Paired transformations attach metamorphic relations to executable version-transition semantics, so violations certify a prediction error.The relation is stronger than a heuristic concern because the required output relation is known.
  • Auditability: A judgment receipt is one narrow machine-checkable artifact within accountable-algorithm and internal-audit work.It contributes a verifiable lifecycle record without requiring source disclosure alone.

10 Scope and limitations

The framework’s guarantees depend on executable adjudicators, a complete declared source interface, compatible deterministic components, and benchmark conditions that limit external validity. Several deployment and human-use questions remain untested.

  • External validity: Constructed adjudicators provide exact counterfactuals but limit external validity because real institutions contain tacit practice, conflicting records, discretionary exceptions, and contested authority.The method’s strength and its external-validity limit arise from the same executable setup.
  • Source interface: Three typed components can miss real sources when policy is hidden in weights, authority changes evidence meaning, or retrieval and adjudication are entangled.The declared interface must therefore be audited rather than assumed complete.
  • Compatibility: Hybrid coherence is guaranteed in ReasonBench but not across database migrations, ontology changes, or redesigned procedures.Those transitions require an explicit compatibility map or should be recorded as uncertifiable under the old interface.
  • Certification cost: Exact unrestricted certification costs 2^k cells, making fine-grained decompositions impractical without additional assumptions such as monotonicity or sparsity.Three sources are practical, but larger decompositions can become expensive.
  • Benchmark composition: The organizational locked test has six independent clauses, so aggregate results can hide stratum imbalance and cannot represent organizational policy generally.The logical stratum is larger but templated.
  • Experimental scope: The primary comparison uses one model family, one LoRA configuration, one serialization per target, fixed cell order, and ordinary autoregressive decoding.The 0.6B replication tests capacity in only one direction, while augmentation uses three seeds and one rate.
  • Human validity: Receipts are exact under benchmark semantics, but practitioner completeness, auditor preferences, and effects on appeals remain untested.No human-validity study establishes whether the interface improves real oversight practice.
  • Deployment evidence: No deployed two-plane system was compared, leaving execution latency, audit scheduling, reviewer time, and saved cost unmeasured.The prediction/certification separation is supported formally and predictively, not by deployment trials.

11 Conclusion

A versioned evaluator’s judgment should remain replayable from inspectable changes in grounds, norms, or authority, with counterfactual execution certifying the account. The empirical evidence shows that structured outputs can restore one consistency relation while leaving another broken, so prediction and certification must remain distinct.

  • A versioned evaluator’s judgment has an account in accepted grounds, governing norms, and authority, whose old/new hybrids expose minimal explanations for the transition.Executing the counterfactual table makes every minimal account recoverable.
  • Both receipt and cube targets can appear nearly perfect on locked tests while failing equivalent source orders, inverse transitions, and unseen multi-source composition.Under unseen multi-source composition, verdicts survive while receipt recovery approaches zero.
  • Permutation augmentation largely restores the targeted relation but leaves the next relation broken, with the structured target paying the largest price where it was already weakest.The reported split shows repair is selective rather than general.
  • Reason-aware evaluation should certify reasons by executing counterfactuals while measuring learned outputs as predictions against consistency, coverage, and reason debt.The required standard keeps changed actions replayable from inspectable declared grounds, rules, or authority.
Loading 2608.20938v1…