# Independent r7 engineering acceptance

**Pass for fresh comparison freeze. This is not launch qualification.**

Runner SHA-256:
`911298ba21797fe0704874ed85c8f306193de2a69d130ea7ebb32b18bb9064f7`.

The only change from independently reviewed r6 is normalization of `triggers`,
`triggered`, and `triggering` to `trigger`. No question IDs, answer values, or
reference access were added.

- All 24 independent static controls pass.
- All 30 answerable consumed questions now retain every required source span
  in their actual isolated field candidates. This is an availability check,
  not model scoring; the ten unknown cases are outside that check.
- The separate r7 targeted run passed all nine questions: the seven original
  failures, Q13, and Q14. Independent readback verified 18 exact grounded
  citations, 36 recorded model requests, source stability, and each answer
  below the 90-second deadline.
- The preserved r6 full run passed 39/40, with Q13 the sole failure. All seven
  original failures passed that run.

There was no full r7 40-question inference run. The r6 full run and r7 targeted
run must remain separately reported. Earlier r6 semantic controls and native
reviews remain evidence for their exact versions; the r7 change is limited to
the reviewed token family. The scorer remains byte-identical to r4.

The reviewer made no model calls, candidate edits, scorer edits, reference
edits, or fresh-corpus reads. Finite role and inferred-document-scope limitations
remain as documented. The fresh 240-attempt comparison and final audit remain
separate requirements before any launch conclusion.