Same Verdict, Different Evidence: Auditing AI Judges

TL;DR for operators

Two AI judges can select the same correct answer while providing explanations of substantially different reliability. Treating those judgments as equally valuable overlooks information that matters for evaluation, post-training, and decision auditing.

DeCert, introduced by Sun and colleagues,¹ evaluates whether a justification uses valid evidence, addresses material issues, and adequately supports its verdict. It also generates training preferences between explanations that reach the same correct conclusion.

Across two 8B model families, this supervision improves several judgment and rationale-quality measures. For organizations building AI evaluation or quality-assurance systems, the practical opportunity is to assess explanations independently of outcome correctness. However, the method still requires reference verdicts, evaluation policies, and fallible automated analyzers. Better observable explanations do not establish that models reveal their internal reasoning.

Two correct judgments can have very different explanations

Consider an automated evaluation system comparing two candidate AI responses. Two judges independently select Response A, and both choices agree with the reference preference. An outcome-based scoring system assigns them equal credit.

Yet one explanation identifies the relevant errors and correctly compares the responses. The other contains a false assessment of Response B. If both judgments enter a training dataset, the system has no outcome-based reason to prefer the sounder explanation.

The paper provides a concrete example. In its Appendix E case study, two judgments correctly select Response A and both receive the maximum Decision Support score of 1.000. Their other scores differ sharply:

Source: Appendix E, Table 9.

The weaker explanation incorrectly accepts Response B’s erroneous light-travel-time calculation. Its final preference remains correct, but the supporting argument is substantially less reliable.

This distinction becomes consequential when AI judgments serve as evidence for model selection, preference training, or documented decisions. Verdict correctness measures whether the system selected the preferred response. It does not establish the quality of the argument presented for that selection.

Three checks determine whether the justification supports the verdict

DeCert treats a generated explanation as a decision certificate: a set of claims whose evidence and logical relationships can be inspected.

The framework evaluates three complementary properties.

Soundness measures whether observations are grounded in the candidate responses and whether inferences follow from their supporting claims. Each rationale is decomposed into an acyclic dependency graph. Unsupported premises weaken downstream claims, so a valid-looking conclusion cannot fully compensate for an invalid supporting step.

Coverage measures whether the explanation addresses important considerations. An independent analyzer identifies up to six material issues using the request, candidate responses, and evaluation policy, without seeing the sampled rationales or verdicts. Claims are matched to those issues using importance weights and validity adjustments. Repeated generic assertions therefore cannot substitute for addressing distinct material concerns.

Decision Support measures whether the verified statements adequately justify the declared preference. The framework derives admissible comparison intervals for evaluation criteria and calculates the least favorable support margin consistent with those intervals.

These intervals express logical uncertainty about what the verified rationale establishes; they are not statistical confidence intervals. Unresolved criteria retain uncertainty rather than automatically becoming ties.

The three scores are combined through a geometric mean:

A serious deficiency in one component consequently depresses overall certificate quality.

The case study illustrates why the components are distinct. Both rationales achieve maximum Decision Support, but the weaker one contains unsupported reasoning and incomplete coverage. Checking whether the conclusion follows from asserted criterion assessments is not equivalent to validating every premise behind those assessments.

Agreement on the correct answer becomes additional training data

The framework’s most consequential training contribution is what it does when sampled judges agree.

For approximately 12,000 training comparison instances, the researchers generate eight candidate judgments per instance: three from GPT-4o and five from Qwen3-8B. Each judgment contains a preference verdict and an explanation.

DeCert ranks these judgments using both verdict correctness and certificate quality. The scoring rule gives correctness priority, ensuring that a correct verdict always outranks an incorrect one.

When sampled judgments disagree in correctness, the framework selects stronger and weaker candidates for training. When all judgments are correct, it can still distinguish them by certificate quality. It retains a rationale preference pair when the score difference exceeds a threshold, set to 0.25 in the default configuration. All-incorrect pools are discarded.

This produces approximately 10,000 supervised fine-tuning examples and 8,000 direct preference optimization (DPO) pairs, including about 3,000 all-correct rationale pairs.

The rationale-pair ablation tests whether this additional supervision matters.

Source: Section 4.3, Figure 3.

These are ablation results, not comparisons with unadapted base models. They support the contribution of same-verdict rationale preferences within the reported training configuration.

For post-training teams, the implication is specific: examples with identical correct labels need not be redundant if their explanations differ measurably in quality.

Judgment accuracy improves, but the comparison baseline matters

The researchers train Llama-3.1-8B and Qwen3-8B using LoRA-based supervised fine-tuning followed by DPO.

On the held-out HelpSteer3 evaluation, original-order accuracy rises from 63.9% to 73.0% for Llama and from 65.2% to 71.5% for Qwen. These improvements, 9.1 and 6.3 percentage points, are relative to the respective unadapted base models.

The more informative comparison for the supervision method is against outcome-only training:

Source: Section 4.2, Table 1. PCA means position-consistent accuracy, requiring correctness under both candidate response orders.

DeCert improves both models on the two aggregate out-of-distribution benchmarks. Its Qwen configuration also achieves the highest corresponding aggregate scores among the compared open generative judges.

The gains are narrower than the largest improvements against unadapted base models, and individual benchmark categories do not improve uniformly. The evidence supports comparative effectiveness under these configurations, rather than universal superiority across judging tasks.

Evidence sensitivity exposes improvements that accuracy alone misses

A judge may remain accurate while repeating an explanation that no longer fits the evidence. The paper therefore tests whether generated rationales respond appropriately to decision-relevant changes.

Across 1,000 counterfactually edited response pairs, Llama’s reported Evidence Adaptation Rate increases from 64.5% to 88.9%, while its Stale Rationale Rate decreases from 34.0% to 11.1%. For Qwen, the corresponding changes are 87.2% to 93.4% and 30.5% to 7.3%.

A separate rationale-quality evaluation against human atomic-rationale checklists shows average precision increasing from 0.2212 without rationale preference pairs to 0.2624 with them.

Position robustness also improves. On HelpSteer3, Llama’s position-consistent accuracy increases from 45.2% to 60.0%, and Qwen’s from 53.6% to 62.9%, relative to their base models. The 14.8-point Llama increase is the paper’s largest reported gain in this measure, not a uniform improvement across benchmarks.

The appendix provides further diagnostic evidence. Controlled rationale edits show that Soundness detects 90.5% of fabricated-evidence modifications, Coverage detects 64.1% of missing-key-issue modifications, and Decision Support detects 95.5% of reversed-verdict modifications.

These perturbation tests help establish what the certificate components detect. They do not establish that the scoring system identifies every naturally occurring reasoning defect.

The verification system is itself a source of uncertainty

DeCert does not require reference explanations, but it is neither label-free nor independent of external assessment.

Its material-issue analyzer, GPT-4o-mini, achieves an F1 score of just 0.288 against human-annotated issues in the reported 500-example comparison. Its recall of 0.293 approaches GPT-4o’s 0.299 at substantially lower reported API prices, but the extraction quality remains imperfect.

Incomplete issue identification can distort Coverage, allowing an explanation to appear comprehensive despite omissions. The controlled tests also show that geometric aggregation detects harmful edits more sensitively than arithmetic averaging while producing somewhat more false positives on harmless modifications.

These trade-offs matter when designing automatic rejection thresholds: aggressive detection can also discard acceptable explanations.

More fundamentally, a verified natural-language explanation is not necessarily a faithful account of the model’s internal computations. The experiments demonstrate improvements in observable justification quality and evidence responsiveness, not mechanistic interpretability.

The evidence is also limited to pairwise judging, two 8B model families, the reported evaluation policies, and a defined set of benchmarks. Enterprise performance remains an open validation question.

Where decision certificates could change enterprise workflows

Cognaptus inference: The most immediate application is in systems that already compare alternatives against explicit criteria.

For AI evaluation teams, certificate checks could help determine which judgments deserve inclusion in preference datasets. Instead of treating every correct verdict as equally informative, a pipeline could prioritize judgments whose supporting claims are grounded, sufficiently complete, and consistent with the chosen response.

For document-review or compliance-analysis systems, similar checks could evaluate whether a recommendation addresses required considerations and whether its conclusion remains justified when evidence is incomplete. Such applications would need domain-specific criteria, calibrated verification, and independent evaluation before deployment.

For model governance teams, the distinction suggests maintaining separate measurements for decision correctness, explanation adequacy, response-order consistency, and sensitivity to changing evidence.

The paper does not establish financial returns, compliance improvements, or production reliability for these applications. It provides a tested method for generating and evaluating justification-quality signals that could be adapted to those settings.

The value lies in distinguishing judgments that look equally correct

DeCert shows that correct judgments can conceal consequential differences in their supporting explanations. Its contribution is to make those differences measurable and usable for post-training.

The results indicate that supervising observable justification quality can improve AI judging beyond verdict-only training. They also establish an important boundary: the certificate is only as dependable as its evaluation criteria and verification components.

For systems that use AI judgments as training data or decision records, correctness and justification quality deserve separate treatment.

Cognaptus: Automate the Present, Incubate the Future.