Reports and evidence¶

Reports answer three separate questions: what was configured, what happened, and why a verdict was assigned.

Verdicts¶

A model assessment alone is capped at Potential.

FrameFuzz comparison¶

FrameFuzz reports add a comparison section without changing ordinary report parsing. It shows the template pack/version, random seed, semantic hash, case order, context and binding status, reviewed payload, scorer provenance, per-case outcome, warnings, and one campaign conclusion. confirmed_framing_gap requires a deterministic framed hit, an explicit-case miss, and verified isolation for every case. Direct API campaigns are capped at potential_framing_gap.

An edited template, contaminated clean control, missing scorer, or incomplete campaign is inconclusive. no_gap_observed means only that this configured comparison did not show a gap; it is not a security guarantee. See FrameFuzz.

Local Core reports¶

Each exported run has its own artifact directory containing:

• session.json — structured configuration, turns, evaluations, scorer results, and hashes;
• report.html — a self-contained human-readable report;
• scenario.json — replayable setup without captured evidence or credentials.

The Reports workspace lists the artifact store on disk and opens JSON results as bounded, text-only data. HTML is downloaded rather than injected into the extension because it can contain hostile target output.

Direct API reports¶

Direct mode stores up to 50 bounded reports in extension-owned IndexedDB in the current Chrome profile. Reports can be opened, downloaded as JSON, and deleted individually. The provider key is never included. Removing the extension removes this browser-local history.

What to review¶

For every material turn, check:

1. the objective, tactic, and hypothesis;
2. the exact payload that was approved or automatically authorized;
3. the captured target response;
4. observed signals and evaluator reasoning;
5. every deterministic scorer result, including not-applicable checks;
6. the terminal stop reason and configured limits.

Treat reports as potentially sensitive. They may contain selected target responses and disclosed data. Keep them under the same retention and access policy as other penetration testing evidence.

Evaluation reports¶

The Core CLI can reduce a complete matrix of stored session reports into a frozen strategy evaluation. Its JSON and script-free HTML contain aggregate measurements and snapshot hashes, not copied payloads or responses. Evaluation never reruns a target or provider. See Evaluate and share strategies.