needs-human-run: commit one sanitised evaluation run per track (garak / pyrit / laaf)

[emmanuelgjr]

What

evals/ ships runnable security-test profiles for three tools (Garak, PyRIT, LAAF) mapped to OWASP GenAI entries, but no committed run evidence — nothing in the repo demonstrates that a single profile has ever executed end to end. T-ENG05 (#38) parameterized every target and documented the procedure; this issue is the part only a human can do.

Why an agent can't do it

Producing a real run requires all three, none available to an automated contributor here:

1. Tools installed — pip install garak pyrit + the LAAF repo (git+https://github.com/qorvexconsulting1/laaf-V2.0.git).
2. A credential + written authorisation to run adversarial prompts against a model deployment (see the authorisation notice atop evals/README.md).
3. Budget — roughly $0.10–1.00 per track with a small hosted model.

Fabricating or hand-editing results is explicitly out of bounds (evals/samples/README.md, hard rule "never fabricate results").

Definition of done

Follow evals/samples/README.md and commit one sanitised run per track into evals/samples/{garak,pyrit,laaf}/, each with a RUN.md provenance block (run date, tool version, target label, authorisation, outcome, sanitisation done). Then:

• evals/samples/garak/ — report.jsonl per profile + SUMMARY.txt (no hit logs)
• evals/samples/pyrit/ — per-script summary blocks (no echoed model output)
• evals/samples/laaf/ — crosswalk-report.md + per-stage JSON with payload/response arrays stripped
• Update the "Committed run evidence" row in evals/README.md
• Revisit evals/THRESHOLDS.md — the first real pass rates are the moment to review the DRAFT thresholds (SME sign-off)

The evals.test.mjs guard already requires a RUN.md beside any committed sample, so a results dir without provenance fails CI.

Context

• PR: T-ENG05: parameterize eval targets, surface thresholds, add run-evidence runbook #38 (T-ENG05)
• Thresholds needing SME review: evals/THRESHOLDS.md (25 rows, all DRAFT)