Rank 5 on the docket, and the first item this arc whose expected outcome is
LOSING capability.
conservative_floor is a one-sided Wilson lower bound, and Wilson assumes
independent trials. CORE's pipeline is deterministic, so replaying an identical
case is not a second trial — it is the same trial with a guaranteed outcome.
The deduction sealer folded the raw practice corpus and counted every replay,
so bands recorded 720/720 committed whether that was 720 decisions or 28
decisions seen 26 times each. conservative_floor(720,720)=0.990868 cleared
theta_SERVE=0.99; conservative_floor(28,28)=0.808413 is not close.
RESULT: 25 licensed bands -> 4. Twenty-one demoted.
Survivors: en_conditional_chain, en_disjunctive, en_verb_fact,
en_verb_universal — the only bands whose corpus holds enough independent cases.
What the demotion did and did not do, stated exactly. NO ANSWER CHANGED. No
answer became wrong. `wrong` stayed 0 across all 25 bands. The engine is
exactly as correct as it was. What changed is the CLAIM attached to the answer:
21 bands now serve the same sound conclusion prefixed with "(reasoned, but I
haven't yet earned a verified track record on arguments of this shape)". The
reasoning did not get worse; the boast did.
The producer is hardened, which is the half that stops this recurring.
seal_ledger now refuses outright — assert_sealed_evidence_distinct compares the
ledger about to be written against the corpus's distinct-case count and raises
BEFORE anything is written. The curriculum sealer has carried this guarantee
since ADR-0264 R9; the deduction sealer had no equivalent, which is exactly how
the exposure arose. Catching it at seal time rather than in an audit matters:
an audit finds a padded ledger after it is committed, trusted, and gating a
live flag, and unwinding that then needs a ruling — it needed one.
A HOLLOW GUARD OF MY OWN, caught by sabotage rather than trusted. The first
version compared all_gold_problems() with distinct_gold_problems() — two
functions that agree with each other by construction, and neither of which is
what build_ledger folds. Reverting build_ledger to the raw corpus walked
straight past it AND re-sealed the inflated artifact. Same defect class the
ledger already suffered: a guard checking something adjacent to the thing it
protects. It now takes the BUILT LEDGER as input, because the invariant is
about the artifact. Re-running the identical attack: refused, wrote nothing.
Pinned as test_sealer_refuses_a_ledger_that_counts_replays.
A SECOND MISS, caught by the gate rather than by me. My blast-radius scan
grepped for deduction_serve_license / deduction_serve_ledger / 720 and found
nine files. It missed tests/test_construction_inventory.py, which asserts
deduction SURFACE STRINGS and never mentions the ledger — so a pattern search
over the wrong noun could not see it. The full gate found it. Recorded because
"I scanned for the blast radius" is worth exactly as much as the scan's key,
and mine was the wrong key.
That test is the #138 fabrication defect pin, and it is updated WITHOUT being
weakened. Both surfaces now carry the disclosure prefix, and the fabrication
survives untouched — which is a sharper finding than before: a disclosure
prefix looks like it might cover the hazard, and it does not. The user is still
told they said "furthermore". Exact-match assertions kept on the full strings
so neither the fabrication nor the licence state can drift unnoticed, plus an
isolating assertion that the only difference between clean and leaked recitals
is the fabricated premise.
Four records corrected in the same PR, each of which would otherwise have gone
on asserting capability that no longer exists:
- CAPABILITY_LEDGERS' manifest note ("25 bands at 720/720 wrong=0")
- the published CLAIMS.md claim — caught immediately by G-22's own pin, which
is that fix earning its keep three days after landing
- test_volume_honesty.py's module docstring, which said the exposure was
"NOT silently repaired ... changing it is Shay's ratification, not a
test's". R-13 IS that ratification, so the line went stale by being acted on
- the exposure inventory's framing: it now pins the repair, not the gap
test_volume_honesty's sealed-vs-producer assertion is strictly stronger than it
was. It compared sealed committed counts to the producer's committed counts and
both were the inflated 720 — the assertion held perfectly while the number it
agreed on was the wrong number. Two records agreeing is not evidence that
either is right. It now pins committed == distinct, plus a non-vacuity check.
Nine tests encoded the unearned expectation and were updated to the honest one.
Closes G-19 and H-1.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Wcw2pnMBwyvmNyQg4uPEt4
Builds the producer `chat/curriculum_serve_license.py` already named as the
only writer of the ledger it reads — `evals.curriculum_serve.practice.runner
.seal_ledger` — which did not exist. Until now, authoring curriculum could
not earn a license at any volume.
A band's case space is ENUMERATED, not authored: every exam question that
routes to that subject (resolve_domain) whose relation is in that family, one
committed case per distinct query atom. So `committed == distinct` is not a
property the generator maintains, it is the same fact twice — the structural
form of ADR-0264 R9, and why this producer cannot develop the deduction
producer's exposure (CASES_PER_BAND=720 filled from spaces as small as 28).
CASES_PER_BAND=660 follows `evals/determination_estimation` per
DIVISION-OF-WORK §4: a ceiling on committed volume, not a quota.
11 bands, wrong=0, inflation 1.0 everywhere.
THE LEDGER IS DELIBERATELY NOT COMMITTED. The plan's Phase-1 exit criterion
asked for "a real (still-unearned) ledger"; that state is unreachable.
Reliability is commitment precision and a correct UNKNOWN is a commitment, so
conservative_floor(660,660)=0.990046 clears theta_SERVE on non-commitments
alone. Sealing today licenses FOUR bands on evidence that is 99.0%-99.98%
correct non-commitment — which ADR-0262 §5.1 rules unacceptable and which
ClassTally structurally cannot see (no verdict axis). Writing the artifact is
therefore a ratification and belongs to Phase D's explicit human-run verb.
The open outcome-mix ruling is now blocking, not academic.
Two corrections to my own earlier records, both measured:
- ADR-0264 §4.2 sized bands from per-TERM exclusivity; the router's predicate
is per-PAIR, which is looser. Every ceiling was understated, and
systems_software_causal is 720 not 630 — flipping it from "cannot reach
657" to "can". 4 of 11 bands can reach the floor, not 3. Now pinned as
measured values in BAND_ATOM_SPACE.
- `then` and `therefore` are taught philosophy_theology lemmas AND control
words of the argument reader (`therefore` is the conclusion marker). 164
routable atoms refuse compiled_premises_unreadable — coverage misses, never
wrongs — in the band Phase F retargets to. The vocabulary boundary is not
screened against the reader's grammar; a pack teaching and/or/not/if would
open a much larger hole the same way. Recorded, not fixed (§6.5), pinned as
a two-lemma bound.
Registers curriculum_serve in AUDIT_SOURCES, retiring the declared `None`
that was Phase C's forcing function, and adds the producer to `deductive`.
[Verification]: in-worktree on canonical CPython 3.12.13 with `uv sync
--locked`: smoke 571 (569 + 1 test + 1 parametrization), deductive 310
(291 + 19). Mutation-checked: injecting the deduction padding pattern into
band_cases turns 5 tests red including the R9 audit and the seal guard.
curriculum_serve lane unchanged at n=32 correct=32 wrong=0 anti_recall=5;
all 11 lane SHA pins match; no ledger written, so no serving change.
Phase B of the curriculum-license-loop arc. Stacked on the ADR-0264 branch.
The invariant: conservative_floor is a one-sided Wilson lower bound and Wilson
assumes INDEPENDENT trials. CORE's pipeline is deterministic, so replaying an
identical case is not a second trial -- it is the same trial with a guaranteed
outcome. A ledger's `committed` count is an upper bound on its evidence; the
defensible figure is its distinct-decision count.
Phase B was scoped as a precaution for a curriculum producer that does not
exist yet. Running the instrument against the producers that DO exist found it
already violated:
21 of the 25 ratified deduction_serve bands do not clear theta_SERVE=0.99 on
distinct evidence. Three inflate 28 distinct cases into 720 committed
(honest floor 0.8084 vs claimed 0.9909). deduction_serving_enabled was
ratified ON 2026-07-24 and is live.
The estimation producer (ADR-0175) is clean -- 660 committed, 660 distinct,
zero repeats, evidently chosen just above the 657 a perfect record needs. So
this is a regression from a standard already established in the repo, not an
architectural gap. `claimed` is identical for all 25 deduction bands because
every band commits 720 with a perfect record: the gate sees 25
indistinguishable passes while honest floors span 0.8084..0.9909.
NOT REPAIRED. The ledger is SHA-sealed, ratified, and gating a live flag, so
re-sealing it is Shay's ratification, not a test's. wrong=0 still holds -- no
band answered anything incorrectly. What is not established is reliability at
the volume claimed. Measured, pinned in both directions, written down.
Added:
- core/reliability_gate/evidence.py -- pure measurement, imported by no serving
path. volume_for_theta() DERIVES 657 from conservative_floor rather than
restating it, so a WILSON_Z change cannot leave a stale literal behind.
- tests/test_volume_honesty.py -- 13 tests. AUDIT_SOURCES is checked against
CAPABILITY_LEDGERS, so a new licensed capability cannot ship unaudited; the
curriculum entry is a declared None that fails the moment its ledger exists.
That is Phase C's forcing function against repeating the pattern.
- docs/research/distinct-evidence-audit-2026-07-25.md -- full audit + the open
outcome-mix ruling for Shay.
- ADR-0264 R9 amended with the measured exposure.
Decision key is case TEXT deliberately: a tighter key (normalized atom) would
collapse spelling variants and report MORE inflation, so text-identity
under-reports and every number is a floor on the real gap.
Outcome mix is deliberately NOT pinned. ClassTally has no verdict axis, the
deduction producer already balances 240/240/240 by construction, and imposing
a per-verdict-class floor would retroactively fail all 25 bands on a criterion
no ADR has ratified (smallest per-band class count is 120). Recorded for a
ruling with numbers instead.
[Verification]: uv sync --locked on canonical CPython 3.12.13; in-worktree
smoke 569 passed (556 baseline + 13), deductive 285 passed. Registered in the
`smoke` curated suite -- not left to `full`, which gates nothing. Mutation-
checked: doubling CASES_PER_BAND -> 2 failed; reporting committed as distinct
-> 4 failed; gating on `claimed` instead of `honest` -> 3 failed; tree restores
to 13 passed.
Claude