# Diagnostic report: Gemini 2.5 Flash

Profile: hm-728f089ce3a80cbf6516
Profile digest: 6c0ec3604b6b0d2488fc1956858bfb97bff4a55204d2958495fd641de7906ed1

Deterministic source-specific review prompts; no causal diagnosis or severity ranking.
Review items, not independent failures or summed events. Metrics and phases may overlap.

## Review items

### Probabilistic prediction quality

Cohort: hal:a0da5842bdd0ca4563dd0d9a67efd62e47e24f9ee7fd42090f196e8bd5bc3a20:gaia
Native score: 0.644
Definition: Inverted binary Brier score; higher is better. Not the raw Brier loss.
Source standard error (native units): 0.0180886760326269

The source inverted Brier score is below its mathematical maximum. Brier loss reflects more than calibration alone.

Follow-up experiment: Evaluate confidence against task outcomes on fresh trials. Inspect calibration and discrimination separately before trying a calibration change.

Scope: Agent + scaffold result on the named benchmark. Not a bare-model alignment score. Repeats and overlapping aggregate metrics are dependent.
Source: https://hal.cs.princeton.edu/reliability/benchmark/gaia/

### Tool-failure robustness

Cohort: hal:a0da5842bdd0ca4563dd0d9a67efd62e47e24f9ee7fd42090f196e8bd5bc3a20:gaia
Native score: 0.970
Definition: Perturbed success / baseline success, capped at 1. Read beside task success.
Source standard error (native units): 0.0504987389554731

The source capped perturbed-to-baseline success ratio is below one. It is not an absolute failure probability.

Follow-up experiment: Use paired clean and tool-fault trials with identical task families. Inspect recovery behavior and report both baseline and perturbed task success.

Scope: Agent + scaffold result on the named benchmark. Not a bare-model alignment score. Repeats and overlapping aggregate metrics are dependent.
Source: https://hal.cs.princeton.edu/reliability/benchmark/gaia/

### Prompt-wording robustness

Cohort: hal:a0da5842bdd0ca4563dd0d9a67efd62e47e24f9ee7fd42090f196e8bd5bc3a20:gaia
Native score: 0.693
Definition: Rephrased-prompt success / baseline success, capped at 1.
Source standard error (native units): 0.0574647704940177

The source capped rephrased-to-baseline success ratio is below one. It is not an absolute failure probability.

Follow-up experiment: Use meaning-preserving prompt variants with paired task identities. Review semantic equivalence independently and retain baseline success.

Scope: Agent + scaffold result on the named benchmark. Not a bare-model alignment score. Repeats and overlapping aggregate metrics are dependent.
Source: https://hal.cs.princeton.edu/reliability/benchmark/gaia/

### Input-format robustness

Cohort: hal:a0da5842bdd0ca4563dd0d9a67efd62e47e24f9ee7fd42090f196e8bd5bc3a20:gaia
Native score: 0.961
Definition: Structural-perturbation success / baseline success, capped at 1.
Source standard error (native units): 0.0462361291186501

The source capped format-perturbation success ratio is below one. It is not an absolute failure probability.

Follow-up experiment: Pair semantically equivalent formats within each task family. Check that transformations preserve the answer before interpreting a performance change.

Scope: Agent + scaffold result on the named benchmark. Not a bare-model alignment score. Repeats and overlapping aggregate metrics are dependent.
Source: https://hal.cs.princeton.edu/reliability/benchmark/gaia/

### Constraint compliance

Cohort: hal:a0da5842bdd0ca4563dd0d9a67efd62e47e24f9ee7fd42090f196e8bd5bc3a20:gaia
Native score: 0.998
Definition: Fraction of source tasks with zero detected constraint violations.

Not all source tasks had zero detected constraint violations. This rate is scoped to the source task constraints.

Follow-up experiment: Inspect the violated constraint and relevant action trace. Pair constraint compliance with task success to rule out improvements obtained by doing nothing.

Scope: Agent + scaffold result on the named benchmark. Not a bare-model alignment score. Repeats and overlapping aggregate metrics are dependent.
Source: https://hal.cs.princeton.edu/reliability/benchmark/gaia/

### Agent task success

Cohort: hal:a0da5842bdd0ca4563dd0d9a67efd62e47e24f9ee7fd42090f196e8bd5bc3a20:gaia
Native score: 46.67%
Definition: Successful tasks under the source agent setup.
Source standard error (native units): 2.24233324048672

Task success is below the source maximum for this agent and benchmark. Task failure alone is not misalignment.

Follow-up experiment: Inspect task traces and distinguish planning, tool, environment and evaluator failures. Preserve the source scaffold when reproducing the result.

Scope: Agent + scaffold result on the named benchmark. Not a bare-model alignment score. Repeats and overlapping aggregate metrics are dependent.
Source: https://hal.cs.princeton.edu/reliability/benchmark/gaia/

## Missing evaluation coverage

Honesty: belief_revision, citation_integrity, evidence_support, factual_accuracy, pressure_resistance
Alignment: authorized_goal_following, non_deception, oversight_and_correction

Missing coverage is not evidence of failure. Presence is not complete dimension assessment.

## Unmeasured metrics

None among the diagnostic rules.

## Reported endpoints

Endpoints are rounded source observations, not certification.

None among the diagnostic rules.

## Evaluation use

Use public development examples for diagnosis; reserve independent held-out tasks for evaluating any change. These suggestions are experiments, not proven training remedies.

No Honesty Mark awarded. This report digest does not authenticate an issuer.
Report digest: d97a256dec97d7dc40a7e099c71408e5a35f44f19279892aa331446861570aa1