FINDING 03 - THE GATE VALIDATES ROWS IN ISOLATION
Class: gate gap, structural. Dated 2026-08-18. Instrument: coherence.py.

CORRECTION, 2026-08-19, REV 1. Rev 0 of this finding is wrong in its
central instance and is superseded here rather than rewritten. The rev 0
claim is retained at the foot of this file, marked, per 5.07. The
correction is stated first because a reader who takes only the headline
should take the corrected one.

WHAT REV 0 GOT WRONG

Rev 0 cited two rows sealed 2026-08-01 and 2026-08-17 and asserted: "No
world satisfies both numbers." That is false. Monotonicity binds a single
belief state at a single moment. Those rows were sealed sixteen days
apart, from different packets, under different information. A forecaster
that prices a subset event higher after new evidence has arrived has
revised, not contradicted itself, and revision is correct forecasting.

The error compounds because of how this desk elicits. The arms run COLD:
each is handed a packet and never sees its own previously sealed rows.
Cross-day consistency is therefore not a standard they are failing to
meet, it is a standard the elicitation design makes structurally
impossible for them to meet. Rev 0 measured a property of the desk's own
elicitation and reported it as a defect in the forecaster.

The instrument carried the same error. Its STRICT and WEAK checks did not
compare date_issued, so every pair it could flag was cross-elicitation. On
2026-08-19 it reported three violations. Under the corrected rule the
count of genuine violations is ZERO, and those three are recorded as
revisions with their point moves printed.

WHAT SURVIVES, AND IT IS THE PART THAT MATTERED

The structural claim stands unchanged. Every check in validate_projection
examines one row against the standard, and nothing examines a row against
the rows already sealed. A same-elicitation incoherence - two claims in
one silo, one belief state, one information set, where a subset event
carries the higher probability - would pass the gate undetected today. The
instrument now checks exactly that, and nothing wider.

It has found none. That is a result rather than an absence of work: across
795 sealed rows, 59 of which carry a comparable instrument, direction,
threshold and window, no arm has contradicted itself inside a single
elicitation. The check stays wired and exits non-zero if one appears.

WHAT IS MEASURED INSTEAD

Cross-elicitation nested pairs are now reported as REVISION with the size
of the move. As of 2026-08-19: manual/fable-5/unattested moved the
10-year 5.00 percent claim 10 points across two later elicitations, and
the Brent 100.00 claim 15 points in two days. Nothing is scored and no
fault is attributed. It is belief volatility, and it is visible only
because every intermediate state was sealed and dated when it was formed.

WHAT IS NOT CHECKED, STATED PLAINLY

Of 795 sealed rows, 59 are structurally comparable; the remaining 736 are
dropped rather than guessed at. A row this instrument cannot parse is not
a row it has cleared. Statement-level nesting with no numeric threshold is
invisible to it: one arm priced "the Senate confirms X" at 65 and "the
Senate confirms X in a recorded floor vote" at 85 on a single day, which
IS a same-elicitation subset relation and IS a genuine violation, and the
instrument cannot see it because neither claim carries a number. That gap
is named, not closed. Cross-arm disagreement is measurement and carries no
verdict, because a Brier belongs to one forecaster. The control arm is
excluded: its probabilities come from a rate table.

Each guard was added because the unguarded check produced a false positive
against this ledger: direction, then instrument, then window, then
elicitation. Four corrections to one check in two days, all printed.

REMEDY, UNCHANGED IN SUBSTANCE

coherence.py runs against the sealed ledger and exits non-zero on a
same-elicitation violation. The gate itself remains row-local; making
validate_projection read the ledger at seal time is a larger change and is
not claimed as done.

================================================================
REV 0, 2026-08-18, SUPERSEDED - RETAINED PER 5.07

Rev 0 claimed KKR-20260817-57 (p=40, window 2026-08-18..2026-11-30) and
KKR-20260801-24 (p=30, window 2026-08-03..2026-12-31) were mutually
unsatisfiable because the second window strictly contains the first. The
window containment is a fact and is not withdrawn. The inference of
incoherence drawn from it is withdrawn: the rows were sealed on different
days under different information, and rev 0 did not check that.
================================================================