Corrected verdict (2026-06-13)

The earlier headline "③ Our System beats ② Ask AI by +2.85 (7.44 vs 4.59)" was an artifact and does NOT survive a fair test. Recorded here so the inflated number is not cited again.

Why the old number was wrong

It was measured on: the v1 16-question set (not the locked v2/27-Q), N=3 judge rounds (coarse), gpt-5.2-era cached answers, and ② was broken when measured (it was refusing/giving up due to under-reformulation on the strict index). Every one of those inflated ③'s relative lead.

The fair re-measurement

② fixed, fresh answers, N=5, on the zero-flicker Gemini judge (claim-recurrence gate + temp-0; see eval-loop-product), dev split 20260613T004122Z:

③ beats ② by only +0.73 — and BOTH systems trip grounding on ~70% of questions. There are 2 statistically-clean per-question wins for ③ (non-overlapping CIs); the rest is within noise.

Implications

• The "110% grounded" promise is NOT met yet — ③ still hallucinates on most questions.
• ③'s grounding failures are real, not judge noise (the judge is now proven reproducible). Diagnosed root cause: an un-sourced opening definition over an otherwise-grounded body, caused by the live prompt's "direct answer first, no preamble" rule fighting the grounding rule. Mutation #1 (grounded_lead_v1) targets exactly this and validated as a KEEP on its targeted questions.
• The win, where it exists, is a grounding story (② leaks more), not a prose-quality story (when both stay grounded they are ~tied).

Detail

Repo: docs/experiment/gated-question-diagnosis.md (failure classes), docs/experiment/autocorrect-run-log.md (full numbers, CIs, mutation results). See also project-overview, eval-loop-product.