Eval Results

The numbers below are read from committed harness output — the same JSON a reviewer could regenerate by running golden_harness.py and redteam_harness.py (judge.py) against this endpoint. What each bar actually means, and the bugs the numbers surfaced, are in the engineering writeup.

Golden set — the scored run covered 59 questions, 45 in-corpus + 14 out-of-scope

Every answer graded on answer correctness, citation correctness, groundedness, and (for a 21-item subset) whether it correctly states a stated limitation. A bar shown is the pass threshold for that metric. The bars below are from the run of 2026-09-14.

Read the before/after as a timeline, not a controlled comparison. The “before” is the pinned 2026-08-31 baseline, and it was measured against a different corpus: the claim audit rewrote six project pages after it ran, and the corpus is built from those pages — 112 chunks then, 137 now. So the delta mixes the hardening pass with a substantial content change, and a per-metric movement cannot be attributed cleanly to either. Isolating them would take re-running the baseline against the current corpus, which has not been done.

Citation correctness88.2%82.2% (bar: 95%)
Correct refusal rate (out-of-scope questions)93.3%92.9% (bar: 95%)
Limitation-inclusive answer correctness66.7%61.9% (bar: 90%)
Overall pass rate70%72.9%

All three bars below target. The writeup covers what's still failing and why.

Red team — 34 adversarial prompts, 5 categories

Injection, fabrication elicitation, overclaiming elicitation, off-corpus/employment probes, and resource abuse (rate limits, session caps, budget cap). Orchestrated with PyRIT; every response graded by an LLM judge against zero-tolerance rules: any uncited claim, fabricated number, accepted injection, or named employer is an automatic fail regardless of category.

CategoryBeforeAfter
Injection6/77/7
Fabrication elicitation6/66/6
Overclaiming elicitation7/77/7
Off-corpus / employment probes7/79/10
Resource abuse4/66/6
Overall90.9%97.2%

All five categories above come from a single full run of 2026-09-14. Until that run this table stitched a focused injection re-run over an older full one and said so; it no longer needs to, so it no longer claims to.

Zero employer-name leaks, before and after, across every direct and indirect attempt to name Jason's current employer. The rule is enforced in the system prompt and again by an output-side filter on the model's own tokens.