ticket(#000008,#000006,index): §12.6 reminder A/B verdict + cross-refs
§12.6 in #000008 captures the 2026-05-03T12-38-53Z reminder-only
A/B (apply_caps=False, reminder=True) on the same 9-question
broad subset:
Mode | Strict-rate | Mean ratio | UNGROUNDED
----------+----------------+-------------------+-----------
quote | 0.56 → 0.52 | 0.900 → 0.845 | 0 → 0
pointer | 0.00 → 0.00 | 0.473 → 0.643 | 9 → 6
JSON | 0.19 → 0.22 | 0.524 → 0.735 | 7 → 1
Pointer-mode violation deltas:
FORMAT_COLLAPSED 2 → 0 (-100%)
NO_EVIDENCE_POINTER 9 → 6 ( -33%)
TITLE_MISMATCH 10 → 15 ( +50%) ← side effect
TOO_MANY_CLAIMS 7 → 8 ( +14%)
§10.8 gate verdict: MET. Both FORMAT_COLLAPSED and NO_EVIDENCE_
POINTER cleared the 5pp floor (−7pp absolute / −11pp absolute
respectively). Strongest signals are mean-ratio improvements
(+17pp pointer, +21pp JSON) — grounded rows ground BETTER under
reminder. JSON-mode UNGROUNDED dropped 7 → 1, a 22pp redistribution
from "didn't ground" to "partially grounded".
Caveat: TITLE_MISMATCH increased (+50%). Reminder may pressure
Hermes to cite *something* rather than say "no evidence", picking
up wrong-source citations as a side effect.
Recommendation: §10.8 gate met but hold default flip until §12.7
(cap-only) and §12.8 (cap+reminder) cells run, per §10.8 "if A+B
together outperform either alone by ≥5pp: land Option D".
Cross-references:
- TICKETS.md index: #000008 status flipped to "open · phases 0–4
landed; bench A/B in progress".
- #000006 rolling log: cross-reference to #000008's bench cycles
+ the d24291b classifier-defect fix surfaced from the
distribution scan.
russell@unturf.com
fa8d93c8ef