`grade` is a band label over `p_win` and nothing else. `servedGrade.gradeFor`
reads exactly { p_win, refused, refusal_reason, insufficient_data,
factor_adjustment } -- no line, no odds, no edge, no side -- and
`gradeFreeze.TOLERANCES.MAX_LETTER_MISMATCH` is 0, so the served letter MUST
equal the band of the served probability.
The intraday refresh re-graded an adversely-moved prop AT THE CURRENT LINE and
then copied only `res.grade` onto the row for the LOCKED line, discarding the
p_win that produced it. Two defects in one write: the letter stopped matching
the probability beside it, and the letter answered a different proposition than
the claim it was stamped on.
MEASURED on ledger_entries in the current model era
(engine1@2026-08-07-fullwindow): 30,047 unrevised rows carry ZERO letter/p_win
mismatches; all 9 intraday-revised rows are mismatched. The revision was the
sole producer of incoherent authoritative rows. The pre-cutover era is
uninterpretable here (its `grade` was the engine index, not a p_win band, and
mismatches 49.4% of the time WITHOUT any revision) -- which is exactly why the
control was run before quoting a rate.
Simulated over the 57 historical revisions: in the current era, keeping the
published grade restores coherence on 9 of 9.
So the adverse branch now records the move and stops there. No grader is
invoked, no ledger revision is applied, and the grade-rank comparator is
deleted rather than parked -- a dead comparator beside the code is how the
mutation gets re-wired. A production coherence gate compares every output row
to its input on every BELIEF field and, on any divergence, publishes the
untouched slate instead.
History is untouched: the 57 existing `revised_from_grade` rows keep their
values, and the UI that renders them is unchanged. Forward-only.
Lineage is not touched and stays OFF.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CQJeAG8vcDoL5zkiaJyVb8
