The Report

The Report

Each run writes to --out (default runs/<timestamp>/):

• report.html — human-readable, self-contained (no external assets).
• report.json — machine-readable.
• evidence.jsonl — the full append-only step log (Evidence Schema).

report.html

A summary scoreboard plus one card per scenario:

• Scoreboard — exposure score (0–100), agents exploited / total, cost breaches, actions captured, errors, target, timestamp.
• Per scenario — the attack (category, goal, channel); the Maps to line (ATLAS id + technique name, OWASP id + name); the plain-English verdict; the canary-backed proof (winning payload, target reply, captured tool call, and where the canary surfaced); a collapsible step-by-step timeline; and a concrete fix.

Verdicts are proof-based — see How It Works for the full table.

report.json

{
  "run_id": "...",
  "exposure_score": 0-100,
  "summary": {"total", "exploited", "metered", "mirror_only", "errors",
              "target_host", "target_identity", "started_at"},
  "scenarios": [
    {"scenario_id", "verdict", "verdict_meaning", "exploited", "steps",
     "hit_canaries", "category", "goal", "channel", "atlas", "atlas_name",
     "owasp", "owasp_name", "success_condition", "remediation",
     "proof": {"payload", "reply", "tool_calls", "canary_location"},
     "timeline": [{"step", "channel", "payload", "reply", "tool_calls", "verdict"}]}
  ]
}
```

Exposure score

round(100 * sum(0.7 + 0.3/steps for each proven scenario) / total). A proven finding (success or metered) in one step counts fully; slower ones count down to 0.7; no proven findings → 0.

Re-rendering

aphasia report --run-dir <dir> rebuilds the report from evidence.jsonl. Because final capped/mirror_only verdicts are not written as rows, a re-render shows per-step verdicts; the live run report is faithful.