Wave 1 — delete — run before more Wave 2 enforcement, correcting the method
inversion this arc had been committing: the plan sequences scrub -> DELETE ->
enforce, and Wave 1 was ruling-blocked while Wave 2 was not, so enforcement
landed first on structures deletion might have removed.
PR-3 (R-7 ruled A)
docs/gaps.md and docs/audit/substrate-liveness-ratchet.md carry historical
banners that name H-9's own mechanism as the reason they exist: an instrument
that looks authoritative converts "I should check" into "I already checked".
The ratchet's seven OPEN items — W-003, W-005, W-007, W-008, W-009, W-017,
W-018, every one L10-chained — migrated into G-5 WITH their dependency
chains, because that reasoning is the ratchet's lasting contribution and its
status column is not.
One deliverable could not be executed as written and is recorded rather than
claimed: L12 has NO in-repo generator to drop it from. The system map is
local and gitignored (D5 — a regeneratable index carrying no authority), so
the phantom existed only there and in one taxonomy row. That row now records
the ruling instead of the flag.
PR-3b — new, and it corrected its own author
Added when PR-4's two ratchets made the question unavoidable: what exactly
are they policing? Measured: 21 curated suites, 9 holding exactly ONE file,
12 holding <=4, and 2 reachable from any gate.
I proposed cutting ~11. The evidence supported FOUR. cognition, teaching,
packs and algebra are named in AGENTS.md's own pre-merge-gate instruction;
fast, pulse and proof in the CLI's help text; phase5, phase6, adr-0024, math
and formation in READMEs and ADRs. Deleting those breaks documented commands
— worse than the sprawl. Taking the 4 the evidence supports, rather than the
11 I had already said out loud, is the whole point.
Deleted: refusal, margin, rotor, inner-loop. Zero --suite references anywhere
in code, docs, ADRs, workflows or CLI help. Per-phase ADR-0024 aliases
offered so reviewers could run a phase independently; nothing ever did. All
seven of their files remain covered by adr-0024 — verified BEFORE the cut —
so nothing was orphaned and no coverage moved.
21 -> 17 suites. Pin 3's gate-unreachable gap 19 -> 15 BY DELETION, which
costs no gate time at all, versus promotion which costs it every run. An
alias nobody calls is not a curation decision; it is one more name that has
to be kept true, and two ratchets were policing these.
Registers: H-9 CLOSED, PR-3 and PR-3b marked LANDED, G-5 carries the seven
migrated items, the taxonomy's L12 row records the ruling.
Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Wcw2pnMBwyvmNyQg4uPEt4
PR-4's membership ratchet closed "no new orphans". Within the hour this arc's
first full-tree run found tests/test_ratification_ceremony.py RED, in the
curated suite `teaching`, invoked by NO gate tier. Membership was satisfied and
bought nothing — which is exactly the objection PR-4 itself raised against
G-7's original formulation: a curated suite nobody runs is the same
non-guarantee as `full` wearing a different name. The fix had the same blind
spot as the thing it fixed.
Measured: 21 curated suites, 2 gate-reachable (smoke, deductive). The gap is
NINETEEN, not the handful first estimated.
Shipped as baseline + ratchet, sized to the measurement rather than the rule:
19 `post-merge-only` justifications written in one sitting is the ceremony
failure already rejected for the 749-file membership baseline. The pin freezes
the gap, fails on growth, enforces both directions (a promoted suite must leave
the baseline), pins the count, and — critically — VERIFIES ITS DECLARED GATE
SET AGAINST THE SHELL (pre-push / local-ci.sh), so it cannot go vacuous the way
a scan that silently matches nothing does (N-1's class, three false conclusions
in this repo so far).
Two sabotages observed red: a new curated suite with no gate caller, and a
GATE_SUITES value drifted from the scripts it claims to describe. The second
correctly cascaded into two failures, which is the anti-vacuity guard working.
What pin 3 deliberately does NOT do: decide which of the 19 belong on the gate.
That costs gate time and is a decision, not an engineering task. Recorded as the
open half of G-7. `teaching` is the concrete first candidate — it just shipped a
red ratification pin that no gate ran.
Process note: a gate run was invalidated by editing the tree mid-run for the
third time today. Stopped with TaskStop rather than pkill this time, which
reaped cleanly instead of leaving the orphaned wait-loop that ran for 2h16m
earlier. All edits now land before any run starts.
Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Wcw2pnMBwyvmNyQg4uPEt4
Ruled: fix, don't document. The argument that settled it — a red test on a
serving-path contract left with an explanatory note IS the H-8 failure mode, a
document acknowledging a gap instead of closing it, which becomes the next
assessor's finding.
Two distinct failures, and the second is the serious one.
1. THE CEREMONY PINS WERE STALE AGAINST CORRECT CODE (H-8e's class).
ChainRecord.polarity defaults to AFFIRMATIVE and is DELIBERATELY OMITTED
from affirmative rows (ADR-0264 R1; teaching/ratification.py:119). Both pins
built ChainRecord(**{k: row[k] for k in __slots__}), so they KeyError'd on
every correctly-serialized row. Now built from the row's own keys, letting
absent fields take their dataclass defaults — which RESTORES the guarantee
rather than weakening it. Consequence of the staleness: ratification-corpus
byte compatibility and unreviewed-status refusal were UNVERIFIED since
ADR-0264 R1 landed.
2. CLAIMS.md PUBLISHED A SUPERSEDED EVIDENCE DIGEST FOR A LICENSED CAPABILITY.
f9e9cc0c (2026-07-26, "the deduction lane hashes the prose it serves")
strengthened the lane, rewrote evals/deduction_serve/report.json, and moved
the authoritative pin in scripts/verify_lane_shas.py:60 to c855d55c —
recording the superseded 0b461a5a on line 56. It did NOT regenerate
CLAIMS.md. So for two days the published claim for deduction_serve_v1
pointed at evidence that no longer existed. The claim TEXT never changed;
only the pointer was wrong — which is exactly the failure a digest exists to
prevent. Regenerated; the diff is one line, the digest.
The drift had reached a THIRD artifact: this assessment's own
M3-comprehension-reasoning.md card cited the same stale 0b461a5a as its
would-fail-if-absent evidence. Corrected.
THE MECHANISM IS G-7's, AGAIN. One commit updated the lane, its verifier and
its lane test, and missed one downstream artifact — and the pin that guards
that artifact (tests/test_claims_md_is_current.py) EXISTED AND RAN ON NO GATE.
Nothing surfaced it until this arc ran the full tree once. Both it and
test_ratification_ceremony.py are now registered in TEST_SUITES["smoke"].
The membership ratchet caught the author a second time: promoting
test_claims_md_is_current.py onto the gate without removing its baseline line
failed test_baseline_has_no_stale_entries immediately. Baseline 748 -> 747.
Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Wcw2pnMBwyvmNyQg4uPEt4
PR-6 promoted tests/test_safety_pack.py onto the gate and left it listed in
tests/full_only_baseline.txt. test_baseline_has_no_stale_entries failed the very
next smoke run, naming the file and the direction.
That is the both-directions half of the ratchet doing exactly the work it was
written for, roughly an hour after it landed, against the person who wrote it.
A baseline that silently keeps entries it no longer needs stops being a
measurement and becomes a story about one — so the failure is the feature.
Recorded in the pin's own docstring and in G-7, because a mechanism catching its
author is better evidence than a sabotage test: the sabotage was contrived, this
was not.
Baseline: 749 -> 748. The ratchet only turns down.
Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Wcw2pnMBwyvmNyQg4uPEt4
G-9's charge: "law-enforced-by-review is weaker than law-enforced-by-test." Its
stated method — verification pass first, then mechanical PRs — is what made this
worth doing, because the three prohibitions turned out to be in three different
states and only one of them was the state the register assumed.
(a) EXACT RECALL — the real gap, and worse than recorded. "Would a cosine
ranker actually fail a test?" answers NO across nearly the whole surface. The
only import ban in the repository covers exactly one file
(core/physics/wave_manifold.py, via test_third_door_cohesion.py — itself in no
curated suite). Meanwhile generate/realize/recall.py, which IS the recall path,
states the prohibition in its module docstring — "an exact, deterministic
equality scan (no cosine / HNSW / ANN)" — and nothing enforced it. Law asserted
where it executes, enforced nowhere: H-8's failure mode applied to a
prohibition rather than a record.
Now scanned across vault/, generate/realize/, generate/meaning_graph/, field/,
recognition/ for banned ANN libraries AND for hand-rolled cosine* definitions.
The second half matters: an import ban only stops the convenient version, and
ten lines of numpy is the realistic way a well-meaning contributor violates
this. A third pin ties the docstring claim to the scan, so the sentence and its
enforcement cannot drift apart.
(b) GOVERNANCE BYPASS — now exact. INV-07 proves governance works WHEN CALLED;
nothing proved it could not be skipped. Mechanism: no function in
chat/runtime.py may construct a TurnVerdicts without invoking
safety_check.check, and every such call must pass self.safety_pack — because a
bypass need not skip the call, it can substitute the pack. Measured at close:
two verdict-constructing functions, both governed.
(c) SAFETY-PACK NON-SWAPPABILITY — already well pinned; the gap was REACH, not
coverage. tests/test_safety_pack.py pins unratified-pack-refused-in-production,
missing-companion-report-refused, seal-failure-refused, path-traversal-rejected,
missing-pack-fails-closed — and ran in NO curated suite, so a fail-closed safety
contract was verified only after merge. Promoted onto the gate; the new pin
guards that placement rather than duplicating the substance.
FOUR SABOTAGES, EACH OBSERVED RED, each caught by its own pin: a banned library
imported on a recall path, a hand-rolled cosine ranker, a serving path
assembling verdicts without governing, the safety pack demoted off the gate.
Both new files registered in TEST_SUITES["smoke"] on creation (#136). Source
scans only — no runtime behavior, no flag, no serving path touched.
Registers: G-9 CLOSED with the per-item findings; PR-6 marked LANDED.
Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Wcw2pnMBwyvmNyQg4uPEt4
N-9 is the finding, and it PREVENTS work rather than re-scoping it.
WHAT THE PLAN GOT WRONG. N-3 concluded the 23-vs-13 smoke delta was unintended
drift, on two in-repo comments asserting parity, and called that "the strongest
evidence available that the drift was never intended". H-12 repeated it; R-14
built three options on it recommending "raise CI, +46s measured"; PR-4
specified a bidirectional parity pin as deliverable one.
A third artifact answers all of it, and none of them read it.
tests/test_cli_test_suites.py::test_cli_smoke_suite_covers_ci_smoke_gate has
pinned smoke.yml SUBSET-OF TEST_SUITES["smoke"] since before this assessment —
which is why every measurement finds CI-only == 0; a pin, not luck — beneath a
comment headed "DELIBERATELY ONE-DIRECTIONAL — do not 'complete' this by also
asserting local_paths <= ci_paths". Its history is 50fa287d (2026-07-25):
"revert(tests): drop the local-to-CI smoke parity assertion — CI is not a
gate." An earlier agent read the same comment, drew the same inference, made
the assertion symmetric, and reverted it hours later: "That was wrong, and
AGENTS.md says so in a line I had already read."
AGENTS.md:280 — GitHub is "a mirror only; its Actions are billing-locked and
produce dead signals — never chase them." §277 — the workflows are "secondary
observability only." So the local suite is the SOURCE and is deliberately
broader; the delta is the design.
- PR-4 pin 1: WITHDRAWN. It exists, in the direction that protects something.
- R-14: premise corrected in the packet. Option A spends 46s on a workflow
that does not gate; option B would delete real protection for a fiction;
option C's live half is a two-line comment fix, done here.
- N-3's exposure claim was too GENEROUS to CI: for a push that skipped the
local gate there is no automatic check at all. Larger than stated, and not
closable by editing smoke.yml.
WHAT LANDED. Measured: 749 of 877 test files are in no curated suite. The
plan's mechanism ("assign every orphan, or 749 exclusion reasons") is hollow at
that scale — glob topic-suites ("adr": tests/test_adr_*.py absorbs 172) satisfy
the assertion while changing nothing about what executes, which is precisely the
Third-Door objection the plan levels at G-7's own first formulation. All four
demonstrated incidents (#113, #136, negation-survives-articulation,
speculative-subject-lifecycle) were caused by a NEWLY LANDED orphan; none by a
legacy one.
So: a ratchet. tests/test_suite_membership.py + tests/full_only_baseline.txt.
New orphans fail. The baseline is enforced in BOTH directions — a promoted or
deleted file must leave it, so the list can only shrink — and its count is
pinned so bulk movement is a reviewed decision, in the discipline
test_volume_honesty.py uses.
Also: the gate-guarding pin now runs on the gate. test_cli_test_suites.py lived
in `fast`; the pre-push gate runs smoke + deductive. The pin guarding the gate
did not run on the gate.
Three sabotages, each OBSERVED RED: a new unregistered test file, a stale
baseline entry, the parity pin demoted out of smoke.
Note on the delta: now 12 (local 27, CI 15) because this session added three
pins locally. Under the corrected reading that is the design, not widening
drift — the invariant is CI-only == 0, and that is pinned. The registers now
say so; "ten files" was a measurement, never a target.
Process note: `git checkout --` during sabotage cleanup destroyed uncommitted
edits to core/cli_test.py, which were redone. Backups, not checkout, on files
with unstaged work.
Registers: G-7 CLOSED, H-12 amended, R-14 dissolved in the packet, PR-4
re-specified, N-9 added to plan §0.
Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Wcw2pnMBwyvmNyQg4uPEt4
Wave 2's PR-9, whose authority is "mechanical PR" and which is gated on no
ruling. Widened by the 2026-07-28 triage from the accrual guard alone to all
three broad `except` blocks on the turn spine.
Each guard is DELIBERATE and stays: accrual is additive and must never crash a
turn; a pack/catalog failure must not take down logos authority; a probe
failure must not touch the served surface. Narrowing them to ImportError/
OSError — as an external assessment proposed — would convert a malformed
decision object into a turn-spine crash, trading silent failure for
served-surface failure. What was wrong is not the tolerance, it is that each
wrote a value INDISTINGUISHABLE FROM ITS HEALTHY CASE:
- a raising read->realize->determine chain left the same None accrual a quiet
turn leaves;
- a logos decision that raised looked exactly like a turn where logos was
never consulted (both: logos_decision_kind == "");
- a crashed geometric probe recorded the same empty neighborhood as a probe
that ran and found nothing — in the block whose entire job is producing the
data that prices the anti-unification roadmap.
Silence, in the one layer whose constitution is "failures are typed, never
silent" (INV-34).
chat/runtime.py _accrual_swallows / _last_accrual_error, read via
accrual_swallow_telemetry(). _last_turn_accrual is
still None on failure, so _maybe_surface_determination
returns byte-identically — ADDITIVE BY CONSTRUCTION,
and pinned as such.
cognition/result.py logos_error on CognitiveTurnResult.
cognition/pipeline.py logos_error recorded; probe_error,
neighbor_scan_errors, graph_anti_unify_error on the
OOV telemetry.
tests/test_accrual_swallow_telemetry.py — 7 pins. The two behavioral ones were
OBSERVED RED against a build carrying the fields with the recording stripped,
so they discriminate rather than merely pass. Two test bugs of mine were fixed
along the way (wrong logos module path; CognitiveTurnResult has required
fields) — the tests were wrong, not the code.
Registered in BOTH TEST_SUITES["smoke"] and smoke.yml on creation, per #136.
Local/CI delta verified unchanged at exactly ten (local 25, CI 15, CI-only 0),
so this settles nothing R-14 has to rule.
Cleanup as found: three dead imports removed from chat/runtime.py
(_COGNITION_PACK_ID, _TEACHING_CORPUS_ID, engine_identity_for_config), each
confirmed to have no re-export dependence first — workbench/readers.py imports
engine_identity_for_config straight from core.engine_identity. ruff 4 -> 1 on
these paths; the survivor is a pre-existing structural E402.
Registers: H-11 CLOSED; PR-9 marked LANDED in the plan.
Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Wcw2pnMBwyvmNyQg4uPEt4
71542e61 fixed only half of H-13 and asserted a premise that is false. This
replaces it, and records why, because the failure mode is the one this arc
exists to catch.
WHAT WAS WRONG. The refcount rested on "seeding and eviction walk the same
source list, so the counts balance." They do not. The call site's arms are
MUTUALLY EXCLUSIVE (pipeline.py §10b): a proposal is seeded when SPECULATIVE or
released when COHERENT, never both. So a release never balances a claim the
same proposal made, and a COHERENT proposal ALWAYS releases claims it never
held. Refcounting therefore repaired only the sub-case where both proposals had
been taught speculatively. The ordinary case — an unreviewed `wisdom`, then a
separate correction about `practical wisdom` that passes review on its own turn
— still stripped the marker. Measured against the landed refcount:
`sibling marked? False`.
The first attempt's tests passed because their helpers split tokens themselves,
mirroring a call pattern production no longer used — tests written to the fix
instead of to the defect. That is the same "asserted past the evidence" failure
diagnosed in two external assessments earlier today, committed here one commit
later, and catchable by reading fifteen lines at the call site.
WHAT IS CORRECT. _speculative_subjects maps token -> SET OF CLAIMING SUBJECTS.
A token is evicted exactly when the subject that actually claimed it is
reviewed; discarding an absent claimant is a no-op, so a neighbour's marker
survives. Token derivation moved out of the call site into
_speculative_index_tokens, because ownership is what the call site cannot see —
doing the split out there made every token its own claimant, which was the bug.
Tests rewritten to call the cache exactly as §10b does. The lead pin
(test_coherent_proposal_does_not_unmark_an_unrelated_subject) was confirmed to
FAIL against the refcount design before this landed, so it discriminates rather
than merely passing. 13 pins green; the 8 pre-existing lifecycle pins unchanged.
SUITE MEMBERSHIP, resolved rather than deferred. The file pinning this behavior
sat in NO curated suite, reachable only through `full` — which is why the
regression was invisible (G-7's mechanism, on the pin that would have caught
it). Registered now in BOTH TEST_SUITES["smoke"] and smoke.yml, per the #136
precedent that an unregistered pin runs nowhere. Verified the local/CI delta is
unchanged at exactly ten files (local 24, CI 14, CI-only 0), so this is
compatible with all three R-14 options and settles none of them; PR-4 still
owns the standing ten.
Cleanup as found: three unambiguously dead imports removed from pipeline.py
(hashlib, RatificationOutcome, RatifiedIntent). ruff 4 -> 1; the survivor is a
pre-existing structural E402, left alone.
Ratification note unchanged: this changes served surfaces — markers appear on
turns where they previously vanished. Direction is toward disclosure and it
restores ADR-0021 §Articulation's stated intent rather than extending it.
Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Wcw2pnMBwyvmNyQg4uPEt4
The speculative-marker cache seeds a subject AND each >=4-char token split from
it, so two independent proposals can claim the same token ("wisdom" from both
`wisdom` and `practical wisdom`). `_forget_speculative_subject` popped the token
outright on COHERENT promotion — regardless of which proposal had seeded it — so
promoting one proposal stripped the marker from another that was still
unreviewed. Unreviewed material was then served WITHOUT the "(speculative, not
yet reviewed)" prefix: the expensive failure direction, on the surface whose
whole job is telling reviewed knowledge from unreviewed.
_speculative_subjects becomes OrderedDict[str, int] — the value is a reference
count. Seeding increments and refreshes LRU position; promotion decrements and
evicts at zero. Seeding and eviction walk the same source list so counts
balance; where they disagree the count stays positive and the subject keeps its
marker, which is the honest direction to fail. An unmatched release floors at
removal and never goes negative, so it cannot borrow against a later proposal's
claim. The LRU cap survives unchanged as a size valve.
Four new pins in tests/test_speculative_subject_lifecycle.py, ALL FOUR OBSERVED
RED BEFORE GREEN — sibling token evicted, marker lost at the served surface,
repeated teaching cleared by one promotion, and count underflow. The eight
pre-existing lifecycle pins pass unchanged.
Also H-8e — expand_relation_closure's docstring claimed "Cycle: if path would
revisit a node, skip", describing a check the code does not perform. The code
was right: termination is structural (monotone growth over a finite triple set),
and a path-revisit skip would REFUSE SOUND transitive derivations, since a
witness path that revisits a node still proves a true fact. The docstring now
says that, and records that `path` is the witnessing step rather than full
ancestry — extending it would move operator_invocation and therefore trace_hash,
so it is deliberately not done here.
Registers updated: H-13 marked FIXED with the ratification note (this changes
served surfaces toward disclosure, restoring ADR-0021 §Articulation's stated
intent rather than extending it); H-8e marked CORRECTED.
Found while fixing it, recorded not acted on: tests/test_speculative_subject_
lifecycle.py — the file pinning this exact behavior — is in NO curated suite,
reachable only through `full`. G-7's mechanism, caught on the pin that would
have caught the defect. Assignment left to PR-4; adding it to TEST_SUITES
["smoke"] alone would widen N-3's ten-file delta and prejudge R-14.
Pre-existing ruff findings in pipeline.py (3 unused imports, 1 E402) confirmed
present before this change and left alone — module-top, not adjacent to the edit.
Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Wcw2pnMBwyvmNyQg4uPEt4
Track A of docs/assessment/50-execution-plan.md §6, executed under its own
protocol. Criterion pre-registered at dfc394d2 (that commit carries the
thresholds as importable constants and NO results); this commit carries the run.
VERDICT: NO-GO. Full credit by ADR-0252 §5.4's own terms.
docs/research/sme-experiment-verdict-797ebad5.md
evals/structure_mapping/adr0252_s5/results/report-797ebad5.json
deterministic_digest b3d9d27592e213104c51bcace7415fd819d722a1b06f7d0a5ca65bd5a234e4ec
53 cases, 1378 pairs, 0.0% non-convergence — no exception contributed to any
number. RS-A (structure-only) separates perfectly, AUC 1.0000, margin +0.0867,
and fails structure-sensitivity. RS-B (attribute-bearing) fails all three.
The mechanism, measured rather than argued: the similarity quotient that would
deliver attribute-invariance is the same quotient that annihilates structural
contrast. An add-vs-subtract minimal pair — one entity, identical numbers, one
relation kind changed — aligns at residual exactly 0.0, its two configurations
being related by a PROPER rotation about e1. Sweeping the attribute weight
(committed as diagnostic_sweep), structure-sensitivity fails at every setting,
and every regime where the SME property survives is a regime where attributes
contribute nothing: AUC 1.00 -> 0.97 -> 0.83 -> 0.69 as they begin to matter.
Scope stated narrowly: this refutes H1 for embeddings that encode role-structure
as point positions aligned by conformal Procrustes under similarity. The
argument is about the quotient, so it generalises across that class; it does not
refute every Cl(4,1) representation, and it does not touch the symbolic
structure-mapping lane already in evals/structure_mapping/.
Two findings that were on nobody's list:
N-8 — the experiment was NOT unrun. rnd/sme-experiment-v2 @ 96e5f468 is titled
"Verdict: GO" and rnd/structure-mapping-experiment @ fc9d0c14 carries an earlier
one. Neither survives inspection: attempt 1 leaked the S1-S4 label into the
embedding; attempt 2 was blind and then took its separability from
`except ValueError: res = 1000.0` — a solver exception counted as a distance —
on a corpus 46/51 of which is outside holdout_dev/v1 with nothing marked, with
duplicate graphs, colliding ids, and an extractor that was never committed. The
register, the plan and the ADR all read the same absence and inherited the same
error for nine days. A branch tip is not a record.
G-21 — the math reader returns a selected graph for 5 of 500 holdout_dev/v1
cases (1.0%), all one skeleton; 24/150 on the public lane. That is why §5.1's
four-structure corpus is not extractable, and it is a sharper measurement of the
comprehension frontier than G-3's construction count.
Registers updated: G-1 carries the verdict, H-10 confirmed and discharged, G-21
added, plan §0 gains N-8, Track A closed, synthesis frontier 2 rewritten.
Pins: tests/test_adr_0252_s5_blindness.py, each observed red before green
(label leak and duplicate-graph sabotage both caught).
Off-serving — evals/structure_mapping/adr0252_s5/ is imported by no serving
path, emits no answers, and changes no flag.
[Verification]: 797ebad5 + this branch — `uv run core test --suite smoke -q`
641 passed in 199.76s; `uv run ruff check` clean; report regenerated after lint
with an unchanged digest.
Phase 5 item 1 asked for the overlap to be sized "by measuring the reader's
construction set against the writer's rather than by growing corpora blindly".
New lane `evals/construction_inventory` does that: it sweeps the writer's whole
parameter space (every RhetoricalMove x IntentTag x predicate x quantifier x
tense x aspect, both public entry points, 32292 cells), quotients it by the
reader's OWN function-word skeleton, and comprehends each construction under
three vocabularies.
writer constructions 1739
reader constructions 19 (mint-site AST-guarded)
overlap, faithful 6
accepts but MIS-READS 22
refuses 1711
vocabulary-dependent 0
Faithfulness, not acceptance, is the criterion — and that reverses two claims
in the plan's §6 RESULT, using a metric that was already on the page:
* g_read_rate is 1/293 but **g_args_rate is 0.0**. The one surface that "reads",
`all molecules are defined as compounds`, is comprehended as
`subset(molecule, defined_as_compound)` — the reader chunked the writer's verb
phrase into a class name. It accepted; it did not comprehend.
* "one construction wide" was wrong both ways: zero on that corpus, six over the
writer's actual output space. A corpus cannot report an inventory's size.
Three findings re-order Phase 5 item 1:
1. The reader FABRICATES on 22 constructions — neither reads nor refuses.
`every dog is a mammal` -> member(every_dog, mammal);
`furthermore, all dogs are mammals` -> asserted(furthermore).
Both are ordinary user English, not writer artefacts. Root cause: _RESERVED
lacks the function words the writer emits, and _parse_propositional accepts
any single token as a fact.
2. It reaches SERVED output. deduction_surface recites
`Given: furthermore; p implies q; p.` — a premise the user never stated — and
chat/runtime.py realizes declarative turns into the held self, so a
fabricated atom is vault-writable. Widening the inventory first would widen
the fabrication surface with it.
3. ADR-0265's defect class survives inside ADR-0265's designated owner of clause
grammar: the four aspect arms of _inflect_predicate bind `negated` to a
wildcard and never read it, so `dog has been defined as mammal` serves both
the assertion and the denial (10530/16146 points). Unreachable today — no
producer sets aspect — so a loaded gun, not a casualty. It survived because
ADR-0265's invariant is structural (is `negated` threaded?) and cannot see an
arm that receives the flag and ignores it. A behavioural sweep can.
The two fixes are written already, as the mutations that turn the defect pins
red. They are NOT applied here: they change what CORE comprehends from user
input, which is a serving change on the truth path — authorization-gated, ADR
first, on the ADR-0261 §5.1 refuse-don't-drop precedent.
Guards, so the tables cannot rot: reader constructions are pinned to an AST
count of the reader's 10 mint sites; the declared tense/aspect axes are pinned
to _inflect_predicate's match arms; the committed corpus is pinned to the
fillers in use. 14 pins, 13 mutations observed RED.
[Verification]: in-worktree, CPython 3.12.13, uv sync --locked.
deductive 517 (was 503, +14 — count moved, registration confirmed)
smoke 641 (unchanged; the file is registered in `deductive`)
lane SHAs 11/11 match, no pin edited
pyright 0 errors on all new files
mutations 13/13 red, including both fabrication fixes
CORE served the affirmative of propositions the user denied. Measured live on
main @ 536d6e55 with realizer_grounded_authority=True:
"evidence does not support truth" -> 'Evidence is verified: what supports truth.'
"evidence supports truth" -> 'Evidence is verified: what supports truth.'
Byte-identical.
Phase 4 (#136) pinned "render_semantic has no negated parameter" as a defect.
Sizing the exposure before acting on it -- the step §7 requires -- showed the
pin was correct and INCOMPLETE. The defect is two drops in series, and the
first is in the graph:
intent.negated parsed from the user's words, always has been
GraphNode NO FIELD FOR IT <-- first drop
ground_graph rebuilds nodes field-by-field; nothing to carry
depth enrichment rebuilds nodes field-by-field; nothing to carry
plan_articulation cannot carry what the node does not hold
realize_semantic never read step.negated
render_semantic no parameter for it <-- second drop
Five separate constructors, each silently defaulting to the affirmative.
Why every gate was green
------------------------
On the default config the ungrounded realizer emits "...", _is_useful_surface
rejects it, and the runtime echo wins the resolver -- and the echo contains the
user's own "does not". The truth path was correct BY ACCIDENT. ADR-0088 Phase B
grounds the graph first, which is exactly when the realizer's surface becomes
useful enough to win. So the defect sat behind a shipped flag, invisible to
every property-of-one-surface test, because nothing compared a denial to its
assertion.
Delivered
---------
- GraphNode.negated, threaded intent -> graph -> ground -> enrich -> step ->
surface. Serialized ONLY when True, so every pre-existing as_dict and every
trace_hash folded from one stays byte-identical -- which is why no lane pin
moves.
- Clause grammar delegated to its one owner. Four of the eight intent
"templates" were never frames; they were plain clauses. They now call
render_step. Writing a second negation implementation in semantic_templates
was the tempting fix and is rejected: Phase 2A spent a unit giving every
linguistic fact one owner, and that design rebuilds the disease one level up.
- Frames that keep a finite verb (VERIFICATION, PROCEDURE, COMPARISON) get an
explicit negated form. RECALL is a speech act with no proposition to deny, so
it falls back to the clause path rather than drop the denial (ADR-0261 §5.1).
Measured
--------
delegation on affirmatives 192/192 byte-identical
serving realizer, all corpora 85/347 -> 109/347 (+24 = the denials)
feature-bearing bucket 49/214 -> 73/214
CONTROL (nothing droppable) 33/33 -> 33/33 unchanged
multi-node (clause joining) 3/100 -> 3/100 unchanged, out of scope
lane pins 11/11 byte-identical, none edited
The control staying 33/33 and multi-node staying 3/100 is the evidence that
this moved the denials and nothing else.
The exhaustive control earned its keep immediately
--------------------------------------------------
It found FIVE more intents serving a denial as its own assertion --
TRANSITIVE_QUERY, FRAME_TRANSFER, NARRATIVE, EXAMPLE, DEDUCTION -- because an
intent with no frame fell back to the UNKNOWN *template* (which cannot say
"not") rather than the UNKNOWN *clause* (which can). The default is now the
capable path, so the next intent added inherits correctness. Found by the
control, not by inspection.
Mutation -- every link reverted individually
--------------------------------------------
baseline 19 pass
graph_from_intent drops intent.negated 8 FAIL
plan_articulation drops node.negated 6 FAIL
ground_graph drops it on rebuild 2 FAIL
realize_semantic stops passing step.negated 6 FAIL
unframed intents fall back to the template 1 FAIL
pipeline depth-enrichment drops it 1 FAIL
The last row was GREEN on the first run -- my fix there was unguarded. That is
what added the structural invariant: every GraphNode(...) built on the serving
path must NAME `negated` or be recorded in an allowlist with a reason. The
defect was five constructors; a per-site test must be written per site, and a
site added without one is invisible. Only recognition/connector.py is exempt
(an EpistemicNode has no polarity to carry).
Four Phase 4 pins are revised, not relaxed -- #136's M2 mutation ("render_semantic
GAINS a negated parameter -> FAIL") has now happened for real, and forced the
deliberate revision it was built to force.
Registered in `smoke` in the same PR that creates it, per the #136 finding that
an unregistered pin runs nowhere.
Still unexpressed, deliberately: quantifier, tense, aspect. NO PRODUCER sets
them anywhere on the serving path, so threading them would be machinery with no
caller. render_step already handles all three the moment a producer exists.
[Verification]: in-worktree on CPython 3.12.13 with `uv sync --locked` --
smoke 641 (was 621), deductive 503, lane pins 11/11 with no pin edited.
Phase 4 of the grammar-unification arc, resolved by option (b) of the plan.
The problem: 149 green fluency cases scored `realize_target`, and
`core/cognition/pipeline.py` never calls it -- it calls `realize_semantic`.
`english_fluency_ood` reported 117/117 + 39/39 for a function that does not
speak, for the whole life of the lane.
The plan framed this as "which realizer is better". Reading the source says it
is not a quality question. `render_semantic`'s signature is
`(intent, subject, predicate, obj, secondary, language, root)` -- no `negated`,
no `quantifier`, no `tense`, no `aspect` -- and `realize_semantic` never reads
them off the step. So the serving writer cannot express content the
ArticulationStep is carrying:
negated=False -> 'Knowledge is defined as opinion.'
negated=True -> 'Knowledge is defined as opinion.'
It serves the AFFIRMATIVE of a negated proposition. That is the ADR-0261 §5.1
family, not a fluency defect. Pinned here as a defect; NOT fixed, because
fixing it changes served output and belongs to Shay.
Measured -- identical contract, all seven scored corpora
--------------------------------------------------------
bucket n realize_target realize_semantic
feature-bearing, single node 214 207 49
no features, multi-node 100 100 3
no features, single node (CONTROL) 33 33 33
total 347 340 85
The control is what makes the rest mean anything. Every corpus hardcodes
IntentTag.UNKNOWN, so "the serving writer scores badly" could have been an
artifact of never giving it a real intent. On the 33 cases carrying nothing it
cannot express, the two are IDENTICAL -- so the gap on the other 314 is the
dropped features and the clause joining, not the intent.
Delivered
---------
- `grammatical_coverage/runner.py` reports `serving_accuracy` beside
`accuracy`; `english_fluency_ood` delegates to that run_lane and gains it
for free. The realizer is a parameter now instead of a hardcoded import.
- `tests/test_phase4_realizer_resolution.py` -- the control, the decomposition,
the negation defect pin, and the §6 evidence.
- The claim is restated at both places it was made: the lane docstring and
`grammar_roundtrip/contract.md`.
Zero served bytes change. No serving authorization needed.
§6 -- the plan's pre-commitment was wrong, and is corrected rather than
quietly edited
------------------------------------------------------------------------
§6 forks on read_rate: risen => Phase 5 diversity; near-zero => §1.8 graph-model
mismatch => ADR. It pre-committed to the second. Neither is what the
measurement says.
g_read_rate went to 1/293. The unblocked case was blocked by a one-line WRITER
defect (predicate-nominal object agreement, #135), not by §1.8. The other 292:
no_template_match 289 reader has no SUBJ-VERB-OBJ template at all
unknown_morphology 2 prepositional objects (reserved_word_in_np)
unsupported_negation 1 reader has no negated-categorical template
Every one is the reader declining a CONSTRUCTION, not a projection disagreeing
about a graph it parsed. Where a construction is in both inventories the round
trip closes exactly. So the barrier is the OVERLAP of the two construction
inventories, currently one construction wide -- Phase 5's item 1, tractable,
not an ADR-scale model decision.
§1.6 read "uniform no_template_match" as evidence FOR the type mismatch. It is
not: no_template_match is a coverage fact, and the corpus was 289/293 bare
transitives, a construction the reader never claimed to read. The measurement
was mostly reporting the corpus's composition.
And a finding in my own stack
------------------------------
`tests/test_realizer_quantifier_agreement.py` lived ONLY in the `cognition`
suite, which is not on the AGENTS.md pre-push gate. So every pin Phases 3 and 4
added to it -- including the invariant covering all twelve inflection branches
-- ran in NO gate. That is why smoke stayed at 621 across two PRs that added 13
tests between them, and it is the same silent-red shape the smoke list already
calls out for test_adr_index.py.
Registered into `deductive`, which now runs 504 instead of 406.
Mutation
--------
baseline 12 pass
serving metric computed with realize_target 4 FAIL
render_semantic GAINS a `negated` parameter 2 FAIL
realize_semantic delegated to realize_target 6 FAIL
Row 2 matters most: if someone FIXES the negation defect, the pin forces a
deliberate revision instead of passing silently.
Still open, still Shay's: option (a), promoting realize_target to the serving
path. It now has a price tag -- 340/347 over 85/347 and the ability to say
"not", against a move in every surface hash and whatever the Shadow Coherence
Gate ruling in core/cognition/surface_resolution.py was protecting.
[Verification]: in-worktree on CPython 3.12.13 with `uv sync --locked` --
smoke 621, deductive 504 (was 406; +12 new Phase 4 pins, +86 previously
ungated), lane pins 11/11 unchanged, no pin edited.
Phase 3 fixed two of the twelve branches of `_inflect_predicate` by hand and
pinned them with a hand-written oracle. The other ten kept handing whole
predicate phrases to single-verb functions, so the same root cause was still
live on **57 of 96** (branch x multi-word predicate) pairs, 9 of 12 branches:
"belongs to" --perfective--> "has belongs toed"
"belongs to" --imperfective--> "is belongs toing"
"belongs to" --past--> "belongs toed"
"is defined as" --future--> "will is defined a"
"contrasts with" --negated--> "does not contrasts with"
Every branch now routes through `morphology.inflect_phrase_head`, which applies
a single-verb inflection to the finite verb and carries tokens 2..n through
byte-identically. Two closed tables were needed because `_base_form` is a
suffix stripper and the head of every copular predicate is a form of BE:
`base_form("is")` returned "i" and `present_participle("is")` returned "iing".
Also fixes predicate-nominal object agreement. `render_step` pluralized the
subject and never the object, so it wrote "all dogs are a mammal". The object
agrees only for a predicate nominal -- a closed set of two -- because a
prepositional object carries its own number and a productive rule would write
"all claims are grounded in evidences".
Measured
--------
tail-mangling pairs 57/96 -> 0/96
g_read_rate (round-trip) 0.0 -> 0.003413 first non-zero in the arc
grammatical_coverage v1 49/49 (2 cases corrected, see below)
english_fluency_ood 117/117 + 39/39 + 13/13 unchanged
discourse_paragraph 12/12 + 6/6 + 5/5 + 1/1 unchanged
zero_code_domain_acquisition 18/18 + 30/30 + 21/21 unchanged
`gram_C14_p01` and `gram_C14_p10` expected "...defined as compound", which is
not English. Corrected, and the old string moved into `reject_surfaces` so the
case now actively rejects what it used to accept. The correction is not my
judgment: the reader independently READS "all molecules are defined as
compounds" and REFUSES the singular form, and that is precisely what took
g_read_rate off zero.
Why the invariant instead of a bigger oracle
--------------------------------------------
A per-branch oracle has to be extended by hand for each new branch, and a
branch added without one is invisible -- which is how eight survived Phase 3.
The pin is structural instead: English marks tense, number and aspect on the
finite verb, so inflection must leave tokens 2..n byte-identical. It needs no
oracle and covers branches nobody has written yet.
It is necessary, not sufficient: "does not contrasts with" preserves its tail
perfectly and is still wrong. `test_do_support_puts_the_head_in_the_bare_
infinitive` is the sufficiency half.
Mutation
--------
baseline 84 pass
inflect_phrase_head applied to the whole phrase 39 FAIL
_IRREGULAR_BASE emptied 2 FAIL
_IRREGULAR_PARTICIPLE auxiliaries removed 2 FAIL
PREDICATIVE_NOMINAL emptied 3 FAIL
PREDICATIVE_NOMINAL widened to every copular predicate 2 FAIL
The last row is the one that matters: the plausible-but-wrong fix -- "pluralize
the object under a plural subject" -- is caught by the mass-noun control.
Not fixed here, deliberately: "has not the following steps" is archaic rather
than wrong, and modernizing it to do-support is a separate judgment call.
[Verification]: smoke 621, deductive 406, lane pins 11/11 unchanged
(grammatical_coverage is not a pinned lane; no pin was edited).
_inflect_predicate applied base_form — a SINGLE-VERB function — to whole
humanized predicate phrases, stripping the final word's last character
class, and its plural branch never consulted `copular`. Nine of the 26 seed
predicates came out wrong and every multi-word one did:
is defined as -> "is defined a" want "are defined as"
has the following ... -> "...step" want "have the following steps"
belongs to -> "belongs to" want "belong to"
causes -> "caus" want "cause"
Root causes, both single-point as §1.4 predicted:
1. morphology.agree_plural_phrase inflects the HEAD (the finite verb is the
first token of every humanized predicate) and carries the rest through.
base_form stays single-verb and is PINNED as still wrong on a phrase, so
nobody "fixes" the symptom in the wrong place.
2. The plural branch now agrees the phrase. The plural NEGATED branch had the
same defect ("do not is defined a") and is fixed with it: a plural copula
takes a bare "not", everything else takes do-support.
Two further defects found while writing the eval cases, both fixed here:
base_form("causes") -> "caus". The -es sibilant rule fired on a stem
ending in a single "s"; it must require a doubled "ss" ("passes"->"pass"),
because a single "s" is nearly always a stem ending in "e" that took a
plain -s ("causes"->"cause").
pluralize("proof") -> "prooves". f/fe -> ves is NOT productive in English
(proof->proofs, chief->chiefs, roof->roofs); it is a closed set. Now
derived as lexicon.VES_PLURAL_SINGULARS from the ves-rows of
IRREGULAR_SINGULARS, so the rule cannot claim a word the table does not
know. Phase 2B had put pluralize on the SERVING path, so this one was live
— and the surface-hashing pin from #133 confirms no served surface moved,
which is that guard's first real use.
New construction C14 quantified_copular, 13 cases: the combination that was
broken was the one never tested. Measured before: 152 corpus cases carry a
quantifier and ZERO combine it with a copular predicate.
MY FIRST DRAFT OF THOSE CASES COULD NOT FAIL. must_contain/word_order listed
quantifier, subject and object but NOT THE VERB, so "all molecules is
defined a compound" passed and reverting the fix left 47/47 green. Rewritten
with the agreed verb in both constraints plus reject_surfaces carrying the
ACTUAL pre-fix output, computed by running the reverted code rather than
guessed. Mutation now:
baseline 13/13 C14
revert phrase-head agreement 4/13
revert -es stem rule 12/13
revert closed ves set 11/13
The 4 survivors of the first mutation are the mass-noun controls, which is
correct — "all evidence is grounded in truth" must NOT pluralize.
Same lesson as the lane pins in #133: a pin that cannot fail guards nothing,
and the only way to know is to break the thing on purpose.
grammar_roundtrip's graph corpus is harvested from the committed case files,
so it grew 280 -> 293. Updated exactly rather than loosened to an inequality:
a corpus that grows or shrinks should require a deliberate edit.
[Verification]: in-worktree on CPython 3.12.13, uv sync --locked —
agreement 26/26 (from 17/26); english_fluency_ood 117/117 + 39/39 + dev
13/13 unchanged; grammatical_coverage v1 49/49; smoke 621 unchanged;
deductive 405; scripts/verify_lane_shas.py 11/11, no pin edited.
The pinned artifact carried verdict counts only. The runner's docstring
justified that: prose is "presentation, not decision", so the pinned bytes
"stay stable against wording-only changes; wording is covered by
tests/test_deduction_surface.py."
Measured, that justification was false. With render._display_noun sabotaged
so every clause reads "all SABOTAGE_dogs are SABOTAGE_animals":
11 lane SHA pins -> 11/11 byte-identical, blind
test_deduction_serve_lane + _license -> 20 passed, blind
tests/test_deduction_surface.py -> 41 passed, blind
(the named wording guard)
evals/grammar_roundtrip -> RED, the only witness
So CORE's user-visible output was unguarded by its own hash pins, which is
how the ratified v1b band served "all dog are mammal" for the entire arc
with wrong=0 intact.
build_report now emits surface_sha256 + per-case surfaces from the real
deduction_grounded_surface — the same call chat serving makes, so what is
hashed is what a user reads. Note build_combined_report re-projects five
named fields per split, so a field added to build_report alone never reaches
the pinned bytes; both had to change. The payload is not the report.
Surfaces are recorded, not just digested, so a moved pin shows the exact
sentence that changed in review instead of an opaque hash to go re-derive.
2,766 -> 37,280 bytes.
deduction_serve_v1 under sabotage: BEFORE byte-identical (blind)
AFTER 52370b73 vs c855d55c (RED)
Re-pinned surgically, one line, old hash recorded beside it, never --update.
Verdicts untouched: 166/166 correct, wrong=0. The hash moved because the
payload grew.
test_surface_hash_moves_when_the_renderer_is_sabotaged makes it permanent:
it corrupts the renderer, requires a digest to move, AND asserts the
aggregate counts are unchanged — proving the digest tracks PROSE rather than
smuggling in a decision change. A pin that cannot fail guards nothing.
Accepted cost: wording-only changes now move this pin. That is the intent —
a wording change IS a user-visible change and should require a deliberate
re-pin. The other 10 lanes are untouched; deduction_serve was fixed because
it is the one demonstrably serving prose to users.
[Verification]: in-worktree on CPython 3.12.13, uv sync --locked —
smoke 621 unchanged; deductive 405 (403 + 2 new);
scripts/verify_lane_shas.py 11/11 with the new pin, and 10/11 (RED on
deduction_serve_v1) under the sabotage it previously could not see.
The ratified, flag-ON v1b categorical band served "all dog are mammal" for
every noun: the A/E/I/O templates supply all/no/some + are, which demand a
plural, and the slots were filled with singular canonical entity ids. Four
of the 47 ratified corpus cases were affected, and wrong=0 never noticed
because every verdict was correct — only the prose was wrong.
Two fixes that compose into a closed round trip:
1. generate/morphology.py becomes the single owner of the number RULES
(Phase 2A gave the TABLES one owner). Both directions live there:
pluralize + singularize, over lexicon.IRREGULAR_PLURALS /
IRREGULAR_SINGULARS, with mass-noun and compound-head handling.
templates.pluralize and reader._singularize were two independent copies
of the regular suffix rules that disagreed about which irregulars they
knew; both now delegate.
2. render._display_noun re-inflects at render time and swaps _ for a space,
so compound ids read "guard dogs" rather than "guards dog" or
"guard_dog".
Result: reader gives wolves -> wolf, renderer gives wolf -> wolves. The
round trip closes.
0 malformed of 47 (from 4)
reader-vs-reader singularization 20/20 (from 8/20), 0 silently wrong
s_surface_match_rate 0.0 -> 0.625, now EQUAL to s_renderable_rate, so the
only remaining losses are surfaces the renderer cannot express at all
NEAR-MISS — widening a reader broke soundness, caught by the lane:
Routing the reader through the full 29-entry table first produced wrong=1 on
band v6-EX. ds-ex-0012 ("No fish are mammals. Therefore some fish are
mammals.") answered invalid where gold is refuted.
_singularize returning None makes the reader REFUSE, and the serving
composer tries the categorical band FIRST, falling through to more capable
bands. "fish" previously returned None (it matches no suffix rule), so v6-EX
got the case and decided it correctly. Resolving "fish" made v1b accept a
sentence it cannot decide.
Fixed by a principle, not a patch: a number-INVARIANT form is ambiguous in
number — "fish are mammals" is plural, "a fish is a mammal" is singular, and
the token cannot tell you which — so a reader that must not guess number
declines it. lexicon.INVARIANT_NUMBER is derived from the singularizer's own
key == value rows, so the rule cannot drift from the table.
=> Coverage and correctness are different axes. Fixing a corrupted VALUE is
safe; widening ACCEPTANCE changes which band answers, and in a
first-match composer that turns a right answer into a wrong one. Same
family as ADR-0261 5.1 refuse-don't-drop.
Also fixed, found by testing the inverse law: the two number tables were not
mutual inverses. The pluralizer lacked cactus->cacti, fungus->fungi,
die->dice and the invariants aircraft/means/offspring, so CORE could read
"cacti" but wrote "cactuses", and would have written "aircrafts", "meanses",
"offsprings". No round trip can close across a non-invertible table.
THE LANE SHA PINS ARE BLIND TO SERVED ENGLISH:
2B changed 4 served surfaces and the pins came back 11/11 byte-identical.
The deduction_serve_v1 hashed report holds only n/counts/by_gold/
correct_by_gold/all_cases_correct/mismatch_examples — no prose at all.
Confirmed by sabotage: with _display_noun returning "SABOTAGE_" + ..., so
every clause reads "all SABOTAGE_dogs are SABOTAGE_animals",
11 lane SHA pins -> 11/11 byte-identical, blind
test_deduction_serve_lane + _license -> 20 passed, blind
Phase 1 grammar_roundtrip -> 3 tests RED, caught it
This corrects #129: byte-identity of the pins is the arbiter for values and
verdicts, NOT for surface text, so the 2A/2B split is not a partition of
"changes users can see." Phase 2A's 11/11 conclusion still stands on its own
independent evidence (24/24 table-equality checks vs the pre-migration
literals), but the justification was thinner than stated.
Before Phase 1 no test in the tree would have noticed CORE's served prose
turning into word salad. That is exactly how "all dog are mammal" survived
on a ratified flag-ON band with wrong=0 intact.
Deliberately not fixed: mass-noun verb agreement ("all evidence ARE truth"
should be "is"). The copula is fixed text inside the templates, so agreeing
it changes the template shape rather than a slot, and no mass noun reaches a
categorical clause in any serve corpus. Recorded in render.py.
No lane pin edited — none moved.
[Verification]: in-worktree on CPython 3.12.13, uv sync --locked —
smoke 621 unchanged; deductive 403 passed (383 + 20 new/rewritten);
scripts/verify_lane_shas.py 11/11 (see blindness note above — gate on
grammar_roundtrip for surface work, not on these pins).
Collapses the duplicated closed English tables into generate/lexicon.py.
Measured on main @ 0948f7cd: the same facts were encoded up to five times
across the reading and writing paths, each copy written independently as a
successive deduction band landed.
The 2A/2B split is decided empirically per #129 — make the change, run the
11 pinned lanes, let byte-identity rule. Result: 11/11 byte-identical, so
this entire unit is 2A and nothing spilled into the authorization-gated 2B.
Collapsed to one object:
- connectives 3 identical copies (member, verb, cond_member)
- predicate display 2 identical 26-entry copies (semantic_templates,
templates)
- be-form inventory 5 copies, not the 2 §1.3 counted. The structural
test found realizer_guard._BE_AUX and
chat/runtime._BE_FORMS, which a COPULA-shaped grep
could never see.
Derived rather than duplicated, so they can no longer drift:
- STRUCTURAL = CONNECTIVES | {therefore}
- QUANTIFIER_TOKENS = QUANTIFIER_LEAD | QUANTIFIER_NON_LEAD
- NEGATION_BEARING_WITH_NOT = NEGATION_BEARING | {not}
- READER_IRREGULAR_SINGULARS = IRREGULAR_SINGULARS restricted to 8 keys
§1.3 re-measured while doing this, and it overstated the disease (plan
§1.9). Almost nothing contradicts:
- the "3 divergent plural tables" are 1 pluralizer + 2 singularizers,
i.e. inverse directions; comparing them as copies is a category error
- reader's 8 singulars are a STRICT SUBSET of member's 29 with the
difference in reader's favour empty, so its values are all correct and
only its coverage is short
- the "3 divergent quantifier sets" are 3 distinct facts; every/each lead
a clause but take singular nouns, so their absence from
PLURAL_QUANTIFIERS is correct
- english's negation set is a subset of member's, and the difference is
principled: v2-EN normalizes "<copula> not", v3-MEM refuses it
Exactly one genuine contradiction exists in the whole inventory:
discourse_planner maps is_defined_as -> "is" where the others map it to
"is defined as". Preserved as a register choice, pinned by test.
Found while reading, pinned as current behaviour for 2B to flip:
reader._singularize does NOT refuse uncovered plurals despite a comment
claiming it does — it falls through to a bare -s strip, so news -> new and
species -> specy, exactly the corruptions member.py's table comment says
its table exists to prevent.
Two 2A exit criteria were unmeasurable as written and are replaced:
- "Jaccard -> 1.00" cannot work. measure_grammar_seam.py scans source
literals, so unifying a fact removes it from those files and Jaccard
FELL, 0.083 -> 0.023 — the success direction reported by a metric
shaped to read like failure. Replaced with an object-identity count:
3 distinct objects behind 8 names, from 8-behind-8. Value comparison
cannot tell a shared object from two equal copies.
- "one table per fact" presumed §1.3's inventory was right. The report now
separates must-be-one-object (all UNIFIED) from related-but-distinct
(6 relations, all HOLD).
ADR-0258 §5 already decided morphology's destination is a ratified pack,
triggered by a grc/he member band. No such band exists, so the promotion is
correctly deferred and lexicon.py is a staging area, not a rival home — no
new ADR. Destination measured for whoever picks it up: packs/en/morphology
.jsonl has 9 records and zero irregular plurals, features.number has no
consumer, and _pack_morph_roots_for reads a top-level "root" key that 0 of
31 records across all four packs define, so it returns {} every call.
[Verification]: in-worktree on CPython 3.12.13, uv sync --locked —
smoke 621 unchanged; deductive 383 (364 + 19 new);
scripts/verify_lane_shas.py 11/11 byte-identical.
The measurement foundation for docs/plans/grammar-unification-2026-07-26.md.
WHY: evals/deterministic_fluency reports 1.00 on all six predicates and
still passes "banana does the.", "wet ground rains the is." and
"is is is is." — it checks terminal punctuation, presence of a verb-shaped
token, and two anti-shape regexes. Heuristic predicates will always have
that failure mode, because grammaticality cannot be measured without a
grammar. So this lane measures agreement between the two halves of CORE
that already encode grammar, and requires the measurement to FAIL on salad.
Two directions, reported separately because they fail for different
reasons and have different remedies:
G-round-trip graph -> realize_target -> surface -> comprehend -> graph
S-round-trip surface -> comprehend -> graph -> categorical renderer -> surface
v1 baseline on main @ 9696443a:
graph_cases 280 surface_cases 8
g_write_rate 1.000 s_read_rate 1.000
g_read_rate 0.000 s_renderable_rate 0.625
g_exact_rate 0.000 s_surface_match_rate 0.000
negative_cases 16
reject_rate 1.000
g_read_rate and s_surface_match_rate are pins on measured DEFECTS, not
goals; they may be revised upward only. s_surface_match_rate = 0 is the
§1.7 categorical render defect caught by construction — the lane found it
without being told to look.
g_args_rate and g_predicates_rate are deliberately separate: high argument
agreement with low predicate agreement would mean the grammars align and
only the vocabulary is split, a materially different remedy from both
being low. That distinction decides the arc's direction (plan §6).
Design notes:
- The committed cases.jsonl is the SINGLE source for authored surfaces —
no in-module duplicate, since a second copy of a corpus is the defect
this arc exists to remove. Negative shuffles are DERIVED at run time so
they cannot drift from the positives.
- The shuffles are lexically identical to positives (same vocabulary, same
length, order destroyed) so the lane cannot pass by vocabulary-checking.
- Fixed rotation, not a PRNG, so reject_rate is byte-reproducible.
- The lane keeps a local copy of the reader's quantifier map ON PURPOSE so
it never becomes a consumer of what it measures;
test_quantifier_map_matches_reader fails loudly if the reader changes.
- _render_categorical deliberately reaches a private serving function: a
lane measuring a private copy would measure what users never see.
Every guarantee is paired with a mutation test. The load-bearing one is
test_reject_rate_goes_red_when_the_reader_accepts_everything: an
accept-everything reader must drive reject_rate to 0.0. Without it,
reject_rate == 1.0 would be unfalsifiable — precisely the defect that
makes the existing fluency lane decoration.
Also documents plainly what round-trip does NOT prove: it measures mutual
intelligibility, not English quality. english_fluency_ood accepts "river
flows valley" and round-trip would be happy with it. No metric here may be
cited as evidence of prose quality.
scripts/measure_grammar_seam.py reproduces every number in the plan's §1
so a reader can check them instead of trusting them.
[Verification]: in-worktree on CPython 3.12.13, uv sync --locked —
smoke 621 (unchanged), deductive 364 (349 + 15 new). Lane SHA pins
verified separately. No serving code touched; new files plus one suite
registration line.
Both units append to `deductive` at the same anchor, immediately after
`tests/test_curriculum_practice.py`. Resolution keeps BOTH entries; nothing
else in either hunk overlapped, and `core/cli_proposal_queue.py` auto-merged
(polarity adds `--polarity` to the ratify parser, this branch adds the reseal
subparser and ratify's "next:" line).
Resolved here rather than at merge time so the resolution is visible on the PR
that caused it.
`polarity` was read by NOTHING. `CurriculumChain.sentence` was unconditionally
affirmative, so a row authored to REFUTE an atom compiled to a premise
ASSERTING it and the question came back `entailed`. The independent oracle
ignored the field too — so gold agreed and wrong=0 stayed green. That is the
shape of defect no gate in this repo can catch, which is why ADR-0264 settled
the epistemology before any code moved.
R1 — row-level `polarity`, absent => affirmative. Not an `intent` value and not
a new `operator_family`: bands key on the connective-derived family, so
`modal_negative` would open a fresh band at n=0 instead of adding refuted
volume to the band it belongs to. An UNRECOGNIZED token is dropped at the
admission boundary rather than read as affirmative — that is the one direction
of error the rule exists to prevent. `as_row` OMITS the field when affirmative,
so committed corpora stay byte-identical and no lane hash moves.
R2 — a negative row compiles under the sentential-negation prefix the argument
reader already parses, so `(¬p, therefore p)` is a tautological refutation.
R3 — `atom_sentence` is polarity-independent, which IS the rule: a negative row
must mint the SAME propositional atom. Demonstrated, not just asserted —
`test_r3_a_different_connective_would_silently_fail_to_refute` shows a negated
paraphrase returning UNKNOWN, i.e. a taught refutation that quietly fails to
refute with nothing going red.
R4 — one atom cannot hold both polarities; rejected at ratification, naming
both polarities and the consequence. Left to serve time it would surface as
INCONSISTENT_PREMISES: the band silently going dark across the whole family
instead of the authoring mistake being named. Duplicate-edge and contradiction
stay distinguishable failures.
R8 — the oracle learns polarity INDEPENDENTLY (its own constants, its own row
reader; asserted at source level that it imports nothing from
teaching.curriculum_premises) and EXCLUDES negatives from the reachability
adjacency, because a denial supplies no step from subject to object.
NO negative row is committed — authoring curriculum is Phase F and Shay's. So
the shipped gold mix is still entailed 72 / unknown 4968, zero refuted, and
Phase C's `test_gold_mix_has_no_refuted_class` stays green with its message
retargeted from "unimplemented" to "no content yet".
Also adds `--polarity` to `core proposal-queue ratify`, so the mechanism is
reachable from the operator surface rather than only from Python.
The end-to-end tests repoint BOTH loaders at a temp corpus instead of writing
the committed one. An earlier draft appended to physics_chains_v1.jsonl and
restored it, which is safe under smoke/deductive (serial) but NOT under `full`,
which injects `-n auto`: a parallel worker reading mid-mutation would see an
uncommitted row and the failure would look like a corpus problem rather than
test isolation. Each loader is still redirected separately at its own module
constant and still parses with its own code, so R8's independence is untouched.
[Verification]: in-worktree on canonical CPython 3.12.13 with `uv sync
--locked`: smoke 593 (571 + 22), deductive 332 (310 + 22). curriculum_serve
lane unchanged at n=32 correct=32 wrong=0 anti_recall=5; all 11 lane SHA pins
match; committed corpora untouched (git status clean under
teaching/domain_chains/). Mutation-checked all three load-bearing rules: R2
sentence ignoring polarity -> 5 red; R8 negatives back in the adjacency -> the
exclusion test red; R4 check removed -> both R4 tests red. The R8 test has an
affirmative control proving the 2-hop path is real when the row is positive,
so `depth == 0` cannot pass vacuously.
Plan §6A, report-only per DIVISION-OF-WORK §4. Both halves ran. The cargo half
is real evidence; the LANE half is vacuous, and saying so is most of the point.
cargo test (PYO3_PYTHON=/opt/homebrew/bin/python3.12, --offline): 43 passed,
0 failed across 7 suites including test_crdt_hash_parity.rs. The plan's read
was right — cargo 1.97.1 and the registry cache are present locally and the
cloud session was blocked only by sandbox network policy.
Lane parity: 11/11 pinned SHAs match under both backends. That number means
nothing. Sabotaging EVERY core_rs callable to raise and re-running under
CORE_BACKEND=rust still gives `lanes: 11/11 match pinned SHAs`, exit 0. Three
sufficient reasons:
- The hot path bypasses the dispatch. algebra.cga.cga_inner is NOT
algebra.backend.cga_inner; 69 call sites import the hot functions from
algebra.cga/algebra.cl41 directly vs 24 through algebra.backend. The ~34k
cga_inner->geometric_product calls/turn that plan §6C calls ~73% of turn
time never reach the dispatch.
- The dtype gates reject the runtime workload. versor_condition IS imported
from algebra.backend by pipeline.py:24, and still falls through: the Rust
arm requires f32 and the wave-field workload is f64.
- Every dispatch arm ends in `except (..., Exception): pass`, so a broken,
garbage-returning, or absent kernel falls back to Python SILENTLY. That is
what makes the first two unfalsifiable by lane hashes.
Instrumented under CORE_BACKEND=rust: zero Rust calls in the whole
deductive_logic lane (pure ROBDD, no CGA), and exactly ONE
(versor_apply_with_closure_f64) in a full pipeline turn. CORE_BACKEND=rust is
today a one-call switch.
So §6B ("if parity holds, consider Rust default") rests on a measurement that
cannot fail, and §6C is the real blocker and is upstream of the backend. NOT
acted on — §6.4 says a parity/determinism finding is escalated, not fixed.
Recorded three follow-ups in dependency order: make the dispatch fail loudly,
add a parity test that fails when the Rust path is not TAKEN, only then route
the hot path.
The one thing fixed here is a plain defect blocking the measurement.
`core rust build` — the activation path print_rust_status() itself recommends —
raised `TypeError: _run() got an unexpected keyword argument 'env'` on every
invocation. core/cli.py::_run is injected into cli_rust.cmd_rust_build, which
passes env=rust_build_env() to carry PYO3_USE_ABI3_FORWARD_COMPATIBILITY, and
_run lacked the keyword the CommandRunner Protocol declares. One keyword, plus
tests/test_cli_runner_contract.py deriving the expected keywords FROM the
Protocol so the two cannot drift apart again. Verified end-to-end: `core rust
build` now builds and installs.
[Verification]: in-worktree on canonical CPython 3.12.13 with `uv sync
--locked`: smoke 575 (569 + 6), deductive 291 (unchanged). New tests fail
under mutation (revert the env keyword -> 3 red). Registered in `fast` AND
`smoke`. Installing core_rs does not activate it (_ALLOW_RUST requires
CORE_BACKEND), verified: extension present + no env var -> "rust backend:
inactive", so no default behaviour moves.
Audit-ledger R7, implementing docs/handoff/curriculum-license-loop-2026-07/
E1-R7-ASSERTION-SPEC.md. `ChatRuntime.chat` seals a TurnEvent before the
surface authority is known; `CognitiveTurnPipeline` back-stamps the resolved
surface and canonical trace_hash onto `turn_log[-1]` after `chat` returns. So
the in-memory log ended up correct and the DURABLE stream carried the
pre-override record — the thing that persists was the wrong one.
Mechanism, per the spec's M1-M3:
- M1 — stage an INDEX into `turn_log`, never the TurnEvent. The flush re-reads
the log, so every back-stamp is picked up by construction and the staged
record cannot go stale. Capturing the event object would be the original
defect with extra steps.
- M2 — `attach_telemetry_sink` does NOT flush a staged turn event, the
opposite of what it correctly does for `_pending_reboot_payload`. The reboot
payload is final when buffered; a staged turn event is mid-flight. Flushing
on attach would silently restore the defect.
- M3 — deferral is opt-in. The pipeline wraps the runtime rather than the
reverse, so the runtime cannot detect whether a pipeline will run. A runtime
used directly (eval and demo paths) emits inline, byte-identically and at
the same moment as before.
The `finally` at the serve boundary is load-bearing (I5): a pipeline that
raises after `chat` returns must still flush, or an error path silently
swallows a telemetry record — worse than the staleness being fixed.
TWO CORRECTIONS TO MY OWN SPEC, reported rather than quietly absorbed:
1. I2 bullet 3 and I3 assert on an emitted `trace_hash`. There is no such
field — `chat/telemetry.py` never serializes `TurnEvent.trace_hash`, so the
sink never carried a trace_hash, stale or otherwise. R7 repairs the SURFACE
only. Pinned by `test_trace_hash_is_not_in_the_wire_format_at_all` so the
claim cannot be re-asserted; the replay-not-recomputation ruling is kept on
the object that does carry the hash.
2. The surface divergence is unreachable under the DEFAULT config:
`finalize_turn_surface` is a no-op on all 36 turns probed, because the
runtime's own surface already wins the resolver. It fires under
`realizer_grounded_authority` — 24 divergences over 84 turns. A test on the
default config would have passed while proving nothing, exactly as the spec
warned, so the fixture pins that flag and asserts the pre-fix divergence
exists before asserting it is gone.
[Verification]: in-worktree on canonical CPython 3.12.13 with `uv sync
--locked`: smoke 591 (569 + 22), deductive 291 (unchanged). All seven existing
sink test files pass UNCHANGED (80 tests). warmed_session_consistency
telemetry_consistency_rate = 1.0. Mutation-checked all three mechanisms:
staging the event instead of the index -> 3 I2 tests red; flushing on attach ->
the M2 test red; dropping the `finally` -> the I5 test red. Registered in
`smoke` and `cognition`.
`RatificationReceipt.pending_stages` has named `ledger_reseal` since the
ratification ceremony was written and nothing performed it, so a sealed ledger
went stale the moment curriculum changed. This is that stage, plus the CLI
verb a human runs, and it is deliberately NOT reachable from `ratify_chain`
(doctrine forbids ratification automation, and a per-row reseal would churn
the ledger's content_sha256).
The load-bearing property is not "the ledger can be rewritten" — it is that
rewriting it cannot grant a serving license by accident. Reliability is
commitment precision, so a curriculum band clears theta_SERVE on correct
NON-COMMITMENTS alone (Phase C, PR #121). A plain regenerate verb would flip
four bands from disclosed to authoritative as a side effect of housekeeping,
on evidence that is ~99% non-entailed, which ADR-0262 §5.1 rules
unacceptable. So:
- `plan_reseal` computes the license DELTA before writing anything, reading
the current ledger through `load_capability_ledger` — the production entry
point — so a tampered ledger refuses here instead of being overwritten and
losing the evidence, and the stage inherits the capability's DECLARED
absence policy rather than choosing its own (bridge rule 5).
- `apply_reseal` REFUSES by default when any band would become newly
licensed, naming the bands and their `entailed` counts. `--allow-new-
licenses` is the ratification. Revocations are never gated: losing a
license is the gate becoming more conservative and needs no authorization.
- `deduction_serve` is deliberately absent from `_RESEALABLE`. Its ledger is
SHA-sealed, ratified, and gating a live flag, and 21 of its 25 bands do not
clear theta_SERVE on distinct evidence. A verb that regenerated it would
erase a measured exposure as housekeeping.
- `ratify`'s output now names the command that performs its pending stage.
Naming a stage without naming its command is how this one stayed unperformed.
Exit codes: 0 dry-run/applied, 1 refused, 2 unknown capability.
[Verification]: in-worktree on canonical CPython 3.12.13 with `uv sync
--locked`: smoke 571 (unchanged), deductive 327 (310 + 17).
`chat/data/curriculum_serve_ledger.json` still absent after the full suite —
no test writes it. Caught and fixed one real gate failure:
`test_no_production_adapter_passes_missing_ok` correctly rejected an explicit
`missing_ok=True`; the fix routes through the manifest instead.
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.
compile_premises emitted every chain in a family, and read_verb_argument
refuses past MAX_PREMISE_SENTENCES=16 — a 17-chain family answered nothing
at all, which meant no curriculum band could ever earn a license regardless
of corpus size (ADR-0264 §4.1).
compile_premises now accepts an optional query=(subject, connective, obj)
and scopes compilation to it: the default is term incidence (chains whose
subject or object is one of the query's two terms), narrowing further to
the exact query-atom rows only if term incidence would still exceed the
cap. Both scopes are verdict-identical to full-family compilation, because
every compiled premise mints one independent propositional atom that no
other atom in the argument can constrain (ADR-0264 R5) — verified over all
8,520 routable questions in the four served domains: 0 mismatches, 4,494
empty scopes all confirmed `unknown` under full-family compilation.
decide_curriculum_question (the only production caller) now distinguishes
an empty SCOPE (open-world UNKNOWN, R6) from an empty FAMILY (the existing
`empty_curriculum` refusal) — getting this wrong would have flipped most
of the question space from unknown to declined. premise_count keeps
reporting family size for the user-visible surface (R7); the compiled
scope size is carried separately on a new `scope_size` field.
Physics gold is unchanged: entailed 14 / unknown 12 / declined 6, lane
report byte-identical, SHA pin unchanged.
[Verification]: smoke suite 569 passed (140.5s), deductive suite 291
passed (50.5s, +6 new tests in tests/test_curriculum_serve.py), warmed_session
lane 10 passed (146.0s), curriculum_serve lane wrong=0 (n=32), all on
canonical Python 3.12.13 with uv sync --locked.
Two unrelated functions shared the name assert_corpus_sound: the
practice-corpus soundness check in evals/deduction_serve/practice/gold.py
(no args) and the curriculum lane's case-contract check in
evals/curriculum_serve/runner.py (domain, cases). Renamed to
assert_practice_gold_sound and assert_lane_cases_sound respectively so
each name states what it validates, and updated all call sites, the
gold.py __all__ export, and two prose mentions that named the function.
[Verification]: uv run --python 3.12.13 core test --suite smoke -q → 569
passed; --suite deductive -q → 285 passed.
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.
Nine ADRs (0254, 0256-0263) were merged into main and left stamped Proposed.
Two carried an explicit ratify-on-merge predicate their own merge had already
discharged, and ADR-0256 governs deduction_serving_enabled, which was ratified
True on 2026-07-24 and is serving live traffic. So the governance record
asserted "not yet decided" about a decision already in force.
That is the asymmetry the assessment arc (#113) found one file over in
workbench/api.py: the honest path degrades, the stale record lies. An unwritten
ADR is a visible gap; a Proposed one that is actually in force is a false
statement.
Each stamp records the ACTUAL ratifying act -- "Accepted, ratified by Joshua
Shay via <merge> (<sha>, <date>)" -- derived from the commit that added the file
and verified an ancestor of main, not assumed. Merge authority is Shay's alone
(AGENTS.md: no merge automation), so the merge IS the ratifying act.
No ADR content changed. No flag changed. ADR-0262's stamp says so explicitly:
accepting it does NOT enable curriculum_serving_enabled, which stays False
pending ratified volume -- eleven bands re-measured today, still 24x-73x short.
tests/test_adr_status_governance.py pins two independent invariants in the
smoke (pre-push) suite:
1. A default-ON flag is not governed by a Proposed ADR. The flag -> ADR mapping
is DERIVED by walking core/config.py for `<name>: bool = True` and reading
ADR refs from the preceding comment block -- not a hand-written table, which
would be the same second-copy-of-a-closed-set defect ADR-0256's arc fixed.
2. A ratify-on-merge predicate cannot coexist with Proposed. Self-discharging:
the file being on main IS the merge having happened.
Plus a vacuity guard, because a derivation that parses zero flags would make
every other assertion pass on an empty set.
Registered two orphans found in passing: test_adr_index.py (5) and
test_ratification_ceremony.py (14) landed in #113 in NO curated suite, so 19
tests -- including the one mechanism that can move curriculum volume -- ran
only under `full`, which gates nothing. Fifth instance of this shape.
Deliberately NOT fixed, recorded in the research doc: the 312-file corpus has
27 unparseable status lines and draft/ratified/active variants. A closed-vocab
assertion would fail on ~35 pre-existing files and get muted, and a muted gate
reads as coverage.
[Verification]: smoke 555 passed in 137.73s (236 baseline + 314 + 5, +1.2s);
governance pin 314 passed in 1.18s standalone and MUTATION-CHECKED -- reverting
ADR-0256 to Proposed fails both invariants independently (2 failed/312 passed);
orphans 19 passed; ruff clean. Canonical Python 3.12.13, uv sync --locked.
Two of main's remaining failures, diagnosed rather than silenced. Neither was a
broken invariant.
## test_adr_0119_1_sealed_holdout (2) — hardcoded developer-machine path
IDENTITY_PATH = Path("/Users/kaizenpro/.config/core/holdout_keys/repo_holdout.txt")
An absolute path to one laptop, in a test. It passed there and failed
unconditionally on every other machine — a second developer, a fresh clone, any
agent session. Now read from CORE_HOLDOUT_KEY (the same env var
evals/holdout_runner.py already uses) with ~/.config/... as the fallback, which
still resolves to the original location on the authoring machine, so those tests
keep running for real there.
Absent the key the tests now SKIP with the reason stated. A machine without the
holdout key cannot run them; that is not the same as the sealed-holdout contract
being violated, and the two must not report identically.
evals/holdout_runner.py is untouched — it still refuses any plaintext fallback
once an identity is supplied. Only the tests' ability to execute is conditional.
## test_adr_0244_gamma_calibration (1) — genuine drift, re-calibrated
The first TEN pinned leakage values are byte-identical to the 2026-07-17 pin.
Divergence starts at probe turn 12 ("ice is cold") and persists — the signature
of a state trajectory splitting at one turn, not a numeric wobble. Deterministic
across runs and IDENTICAL on pristine main, so it is engine drift accumulated
across the generalization arc, not a change from this branch.
Re-pinned, because this is the guard's own documented action ("regenerate with
collect_live_benign_leakages()"; the test says a hit means "re-calibrate and
reconsider the flag flip"), and because the conclusion it protects is unchanged:
benign leakage still overlaps the attack distribution, so flag_flip_authorized
stays False and identity_wave_gate stays OFF. Reconsidered; still not authorized.
These are measurements of an EXPERIMENTAL off-by-default path where every probe
turn is a boundary-violated refusal — that over-triggering on benign traffic is
the finding the calibration exists to record.
Recorded honestly in the constant's comment: the turn where drift begins is
known, the arc commit that caused it is NOT isolated, and the comment says so.
## Not fixed here, because they were already fixed
test_issue_300_versor_margin (3) and test_adr_0172_w0_1_trace_replay_equivalence
(1) were subprocesses shelling out to `uv run core chat` and dying on
"No interpreter found for Python 3.12.13" — the exact-pin problem. Both pass on
this branch already, fixed by 3b2c760. Eight such failures in the main baseline
traced to that one cause.
## Observation, not changed
tests/test_adr_0119_1_sealed_holdout.py carries the full decrypted holdout cases
as EXPECTED_PLAINTEXT, in cleartext, next to the .age file they are encrypted
into. That may be deliberate; flagging rather than touching it.
[Verification]: the six formerly-red files — versor_margin, trace_replay,
gamma_calibration, sealed_holdout, lane_sha_verifier, derived_close_proposals —
44 passed, 2 skipped (holdout key absent, reason stated). Full-tree run in
progress.
NON-CANONICAL: Python 3.12.11.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FduW6Krm3PPQv3P5iwBYtx
*** VERIFICATION PENDING — see the bottom of this message. ***
99 of the 107 full-suite failures are one test, red on clean main since
2026-07-23, and it is guarding the wrong field.
WHAT IS NOT WRONG. test_register_matrix_trace_hash_invariant — the load-bearing
ADR-0072 truth-path-isolation claim — passes on all 99 registers on pristine
main. trace_hash does not vary across the register axis. The invariant holds;
the seals and wrong=0 are untouched.
WHAT IS. test_register_matrix_canonical_surface_byte_identical asserts on
CognitiveTurnResult.surface, on its docstring's claim that this is "the
truth-path field compute_trace_hash consumes". That stopped being true when
Phase 0 flipped resolve_surface to response-first precedence and introduced
SurfaceResolution.hash_surface as the register-invariant capture. Since then
pipeline.py folds hash_surface into compute_trace_hash, while result.py
documents surface as "final voiced surface (what the user sees)".
So the test demanded that registers NOT differ on the served surface — the
opposite of what the register axis exists for, and a direct contradiction of
test_register_substantive_consumption.py, which pins that they DO differ and is
green in smoke. Two tests on main asserting opposite things about the same
bytes; the red one is in neither gate, so it stayed red.
ROOT CAUSE, and why this is a code fix rather than a test edit: hash_surface was
pipeline-local. The invariant was unobservable from a turn result, which is why
the test reached for the nearest visible field and drifted. Exposing it makes
the contract testable at the seam that owns it:
- CognitiveTurnResult.hash_surface — new field, populated from the value the
pipeline already computed
- evals/cognition CaseResult carries it through
- the test asserts byte-identity on hash_surface, with the history recorded so
it cannot drift back
No stored baseline to regenerate — the fixture computes it live.
[Verification]: NOT YET RUN. The register matrix is ~16 min and was still
executing when the working tree had to be committed; its output buffers through
tail, so no partial signal was available. Committed to preserve the work with an
honest label rather than to claim it passes. DO NOT MERGE until
`pytest tests/test_cognition_eval_register_matrix.py` is green — expected 0
failed where main has 99, with test_register_matrix_trace_hash_invariant still
passing. Result to follow.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FduW6Krm3PPQv3P5iwBYtx
Reverts the bidirectional half of the smoke parity pin added earlier today,
and records why at the assertion site so it is not "completed" again.
I read the suite list's comment ("so the local-first pre-push gate equals the
CI gate") as promising equality, made the assertion symmetric, and reported
tests/test_pack_draft_serve_boundary.py as drift because it sits in the local
gate and not in .github/workflows/smoke.yml.
That was wrong, and AGENTS.md says so in a line I had already read: GitHub
Actions are billing-locked and produce dead signals, never chase them. The
Forgejo host (e2-micro) cannot run workflows either. CI here is run MANUALLY
in-worktree — usually faster anyway on UMA hardware with real RAM — which is
what scripts/ci/local-ci.sh now front-ends. smoke.yml is secondary
observability, not a gate.
So the local suite is the SOURCE and is free to be broader. Asserting the
reverse converts a file nobody executes into a maintenance obligation, and one
agents cannot discharge at that — pushing workflow changes needs an OAuth scope
the push credential lacks. The three "owed" lines were never owed.
The surviving direction (CI subset-of local) still protects something real: if
smoke.yml ever names a path the local gate lacks, the local gate is narrower
than a published claim about it.
Arc-close brief updated to match: the "three files owed to smoke.yml" open item
is replaced by a recorded non-finding, so the next reader does not re-derive it.
Its shipped-work table now lists what actually landed in its place — the local
CI runner, the lane-roster fix, and the proposal-sink test-isolation fix.
[Verification]: 21 passed across test_cli_test_suites, test_lane_sha_verifier,
test_adr_index; ruff clean on the changed test.
NON-CANONICAL: Python 3.12.11, not the pinned 3.12.13 (unavailable for
linux-x86_64). Hash-pinned tests reproduce committed values regardless.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FduW6Krm3PPQv3P5iwBYtx
Found by running the full tree: every full-suite run left an untracked
speculative proposal artifact in teaching/proposals/derived_close_facts/.
test_runtime_flag_on_emits_after_consolidation passes engine_state_path=tmp_path,
but that does not redirect the proposal sink — the runtime calls
emit_derived_close_proposals(ctx) with no sink argument, so it resolves
DEFAULT_SINK: the real, repo-relative directory. Fixed by monkeypatching the
module-level default the runtime resolves through.
Its neighbour was defending against exactly this with:
assert not any(sink.glob("*.json")) or True # may have pre-existing from other tests
which is vacuously true, and whose comment was describing the leak it was
papering over rather than the contract it claimed to check. With the flag-ON
test no longer writing there, an honest assertion is possible: snapshot the
live sink BEFORE the tick and require it unchanged after. (First attempt at
this took the snapshot after idle_tick(), which compares the sink to itself —
corrected here.)
Only visible now that the parents[3] DEFAULT_SINK bug is fixed and the sink
resolves inside the repo; before 2026-07-24 these artifacts were landing
outside the tree entirely, which is why nobody saw them.
[Verification]: 28 passed across test_derived_close_proposals,
test_proposal_queue, test_idle_proposal_review; working tree clean afterward,
no residue.
NON-CANONICAL: Python 3.12.11, not the pinned 3.12.13.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FduW6Krm3PPQv3P5iwBYtx
Two things, both surfaced by finally being able to run the tree.
## A pre-existing silent red
tests/test_lane_sha_verifier.py::TestExpectedLaneCoverage was RED on clean
main — confirmed by running it in a pristine worktree at origin/main, and the
file is untouched by any of this branch's work. deduction_serve_v1 (ADR-0256)
and curriculum_serve_v1 (ADR-0262) shipped during the generalization arc and
neither was added to EXPECTED_LANES.
The roster's `extra` assertion is a deliberate tripwire: a new lane is supposed
to fail it once so an author acknowledges the addition. Nobody did, and nothing
caught it, because that file is in neither `smoke` nor `deductive` — no local
gate and no CI job runs it. Same silent-red family as the exact-tuple pin S5
found in passing, and as the register-axis e2e tests that sat red for two days.
This commit is the acknowledgement the tripwire was asking for.
That it took running the full tree to find is the argument for the rest of this
commit.
## scripts/ci/local-ci.sh
AGENTS.md says the merge bar IS the local run. There was no single entry point
for it, and on any host without Python 3.12.13 exactly there was no local run
at all: pyproject pins requires-python == "3.12.13", so uv sync --locked fails
outright and every gate becomes unrunnable.
sh scripts/ci/local-ci.sh --tier smoke|gate|full
`gate` is the three pre-push steps; `full` is the whole tree in parallel. Suite
membership is read from core/cli_test.py::TEST_SUITES through the CLI and never
restated, so this runner cannot drift from the hook the way smoke.yml drifted
from the local list.
Interpreter contract, deliberately fail-closed: the pinned interpreter is
CANONICAL, and the runner refuses on anything else with instructions.
--allow-interpreter-fallback opts into any 3.12.x and stamps every run
NON-CANONICAL. Same discipline as ratified_ledger's missing_ok and the MLX
skip path: a degraded mode is legitimate, a degraded mode reporting itself as
the real thing is not.
Evidence recorded in the script header so the flag is not cargo-culted: the
committed-SHA / trace_hash / content_sha256 / lane-SHA pin set (119 tests) was
run on 3.12.11 and 118 passed, the one failure being the stale roster above,
which fails on 3.12.13 too. That is evidence the exact pin is not load-bearing
for bit-exactness — it is NOT a ruling that it should be relaxed. Relaxing it
is a reproducibility decision and belongs to a human.
## Observed, not fixed
Something in the suite writes a real proposal artifact into
teaching/proposals/derived_close_facts/ instead of a tmp_path. The three test
files that name that sink all override it correctly, so the leak is elsewhere —
likely an idle_tick/contemplation path exercising DEFAULT_SINK. Untracked
residue, removed here, flagged for follow-up. Notable because it is only
visible now that the parents[3] bug is fixed and the sink resolves inside the
repo.
[Verification]: gate tier via the new runner — smoke 236, warmed_session 10,
deductive 285, all passed, correctly stamped NON-CANONICAL. Fail-closed path
verified (exit 3 with instructions). lane_sha_verifier 6 passed. Full tests/
tree in progress at 63% with 0 failures at time of commit.
NON-CANONICAL: Python 3.12.11, not the pinned 3.12.13. Not merge evidence.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FduW6Krm3PPQv3P5iwBYtx
Squashes the arc's work into one commit; the workflow-file edit it originally
carried is excluded (see the end of this message).
## Lane 1 — Workbench recorded a proved answer as ungrounded
With deduction_serving_enabled ratified ON (ADR-0256), workbench/api.py's live
chat route builds a bare ChatRuntime(), so the deduction composer decides
Workbench turns and stamps grounding_source="deduction" — but
_coerce_grounding_source carried a hand-copied whitelist of the six pre-arc
labels and silently rewrote anything else to "none". The runtime comment
reasoned this was inert because "REPL turns do not flow through Workbench's
CognitivePipelineRecord path". True, and irrelevant: the traffic flows the
other way. Stale since 2026-07-24.
Scope is one field. workbench/api.py:818 prefers TurnEvent.epistemic_state,
which read epistemic_state_needed — honest. So the UNregistered path degraded
honestly while the hand-copied whitelist asserted a falsehood; a second copy of
a closed enum was worse than no copy. Hence registration AND derivation:
GROUNDING_SOURCES exposes the Literal's members, and the coercion reads it.
workbench-ui badges/tokens/snapshot follow; enumCoverage.test.ts forces atomicity.
## Lane 2 — the ratification ceremony
The discovery loop was instrumented but not closed. teaching/ratification.py
turns a reviewed decision into a chain record, a corpus commit, and a receipt.
Its design turns on one observation: _ratified_rows DROPS unadmissible rows
silently — correct when serving, a trap when ratifying, because the file grows,
the commit lands, and the band count does not move. So the ceremony refuses to
call an append a ratification until it has re-read the curriculum through the
real loader and seen the chain arrive; a non-admitted append is rolled back.
Validation is a pre-flight courtesy, admission is the proof.
Arena queue entry and ledger reseal are deliberately NOT performed (bridge rule
1); the receipt names them. Front door: `core proposal-queue ratify`, a sibling
of `review` rather than a flag on it.
## Lane 3 — structural closures
- ADR-0263 gains rule 5: absence policy is DECLARED in CAPABILITY_LEDGERS, not
passed at the call site. An AST-matched test fails if a serving path passes
missing_ok again.
- Deductive suite added WHOLE to the pre-push gate: 285 tests in 29s against
smoke's 216 in 62s, so no coverage trade was needed.
- Smoke/CI parity assertion made bidirectional. It was one-directional, and had
drifted.
- test_prior_surface_deduction_binding.py pins correction binding on the
deduction path. The review's diagnosis did NOT reproduce — hash_surface moves
in lockstep — so it pins what is there. Mutation-checked.
- Domain-keyed ADR index over 312 flat-numbered files, explicitly partial.
- Arc-close brief template, plus this arc's own brief filled in against it.
## Lanes 4 and 5 — two premises falsified by measurement, one of them mine
Math 4.2: baseline reproduced (correct=5 wrong=0 refused=495); all four named
cases traced to one seam with each gap isolated by one-variable probes. Then the
number that changes the recommendation: the gap blocking case 0000 affects 1
case in 500, the 'than' gap blocking 0001 affects 2. ADR-0251's prohibition on
per-case growth now rests on a count. No reader change made.
CGA: versor_condition is 0.22% of a turn, not the "~10x proof latency" I claimed
— that multiplied an isolated microbenchmark by a call count and compared it to
a single verdict's latency. The real cost is geometric_product at 33,986
calls/turn (~73%) via cga_inner in search paths. The obvious closed form is NOT
bit-exact (954/4000 in f32); backend.vault_recall's serial fold IS (3000/3000,
worst-rel 0) and is the correct target. cargo test could not run —
static.crates.io is denied by the sandbox network policy — so the Rust parity
question stays open and the typestate lane is carried forward, not shipped
uncompiled.
## Not landed: three lines owed to .github/workflows/smoke.yml
The CI smoke gate is narrower than the local one —
test_pack_draft_serve_boundary.py (ADR-0253 INV-33) has been local-only, unseen
because the parity pin checked one direction. The edit was authored and rejected
at push for lacking the `workflow` OAuth scope, so it is recorded as a named,
dated PENDING_IN_CI exception rather than dropped: the assertion still fires on
any new divergence, and a second guard fires once the three land.
[Verification]: pre-push gates all green — smoke 236 passed, warmed_session 10
passed, deductive 285 passed. Ratification 14, ADR index 5, CLI suites 10.
Grounding/epistemic sweep 741 passed 1 skipped. workbench-ui 598 passed across
73 files, tsc -b clean. capability index 11 passed, digest unchanged. Math
holdout correct=5 wrong=0 refused=495. Committed chain corpora byte-unchanged
after the tests that write to them.
Environment caveat: the repo pins requires-python ==3.12.13, which uv cannot
fetch for linux-x86_64, so `uv sync --locked` fails. All Python runs used a
scratch venv on 3.12.11 with declared deps — not the locked universe, not the
full ~12k suite. The pin was left untouched. Re-run on a 3.12.13 host before
treating this as merge evidence.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FduW6Krm3PPQv3P5iwBYtx
- Promote tests/test_register_substantive_consumption.py into the smoke
suite (local core/cli_test.py + mirrored .github/workflows/smoke.yml):
the falsifiable register-axis (ADR-0069/0071/0077) contract was red on
main outside every gate for the 2026-07-22..24 window, masked by an
unrelated flake label. 36 tests, ~4.6s.
- Found in passing: test_core_test_deductive_suite_expands_to_entailment_lane
was already red on clean main (an exact-tuple pin from when the
`deductive` suite had one file; it has since grown to 12). Fixed to a
contains-style assertion so it can't go stale the same way again.
- Two new capability-index adapters: deduction_serve_existential_result
(Band v6-EX alone — the one band among six that changes the decision
procedure itself) and curriculum_serve_result. Breadth 11 -> 13; baseline
re-frozen (a deliberate re-freeze per its own docstring).
- Promotion sweep: confirmed via the existing wrong=0 lane gate (run
repeatedly this session across 166/166 deduction-serve cases) that no
split beyond the already-applied ds-mem-0020 has a stale declined-gold
case a current band now decides differently.
`core proposal-queue {status,list,show,review}`: a read + review-state-only
CLI over the two proposal sinks that had a Python API but no operator
surface — teaching/proposals/comprehension_failures/ (N5 contemplation)
and teaching/proposals/derived_close_facts/ (idle_tick PR-2 bridge).
"review" appends to a sidecar log (teaching/proposals/review_log.jsonl)
recording that a human looked at an artifact; it never mutates the
artifact, never ratifies, never mounts, never flips a flag. The existing
ratification corridor (teaching/proposals.py, `core teaching
proposals`/`review`, a different sink) is untouched.
Found and fixed a real bug while building the second sink's reader:
generate/determine/derived_close_proposals.py's DEFAULT_SINK used
parents[3] instead of parents[2], resolving one directory ABOVE the repo
instead of inside it. Silent because the feature is default-off and no
test asserted the path — confirmed live via a stray artifact sitting
outside the repository since 2026-06-16. Pinned by
test_derived_close_default_sink_resolves_inside_the_repo.
Implements the measurable test from COMPREHENSION-READER-AUDIT.md's Q5,
generalized to the deduction and curriculum composers: a deterministic
refusal histogram split mechanism/coverage/engine_refused, plus an
admissions-per-batch counter for measuring a future lexicon or curriculum
expansion's effect as a before/after delta instead of asserting it.
Found and fixed a real classification bug while building it: an argument
successfully read by a band but declined by the ROBDD engine for
inconsistent premises was being misattributed to whichever reader band
happened to run last in the fallback cascade. That case is neither a
shape gap (mechanism) nor a vocabulary gap (coverage), so it gets its own
"engine_refused" bucket rather than being folded into either.
Baseline measurement: deduction refusals are 100% mechanism-class today
(the bands read a closed connective grammar, not open vocabulary) —
curriculum refusals are a real mechanism/coverage mix, confirming the
volume ceiling S6 quantifies independently from the ratified-chain side.
Ratified by Shay 2026-07-24. Default flipped False -> True.
Evidence at ratification:
- 25 shape-bands, each holding SERVE at 720/720 wrong=0 on the SHA-sealed
ratified ledger (theta_SERVE=0.99, n >= 657 committed per band);
- deduction-serve lane 166/166 wrong=0 across six hand-authored splits;
- commit gate narrow by construction — a sentence-initial "therefore" IS an
argument, so the composer cannot claim turns that are not one;
- an unearned shape is still served DISCLOSED, never asserted.
Blast radius verified, not assumed: core test --suite deductive 268 passed
with the flag ON by default; a live default-config ChatRuntime decides real
English and existential arguments through grounding_source="deduction" while
ordinary turns still fall through to pack grounding; lane pins 10/11 match —
unchanged by the flip (the single miss is the pre-existing public_demo drift
documented for Tier S).
Rollback is flipping this back: flag-off remains byte-identical to pre-arc
dispatch and leaves no residue, pinned by
test_flag_off_preserves_pack_token_gloss_byte_identity.
curriculum_serving_enabled stays OFF — every curriculum band is unearned, and
the Sonnet brief now records the explicit bar for revisiting it.
The curriculum composer committed on question SHAPE — any `Does …?` text —
and then, being fail-closed, always returned a surface. With the flag on,
"Does the build pass?" would have been taken away from the rest of dispatch
and answered "I haven't been taught the or pass"; "Does anyone know the
time?" likewise. Found by checking the flag's readiness before recommending
ratification; never live, since the flag is default-off.
The asymmetry is the lesson: the deduction composer may commit on shape
because a sentence-initial "therefore" IS a signal of intent — text shaped
that way is an argument. `Does …?` is one of the most common ways to open
any English question and signals nothing. A fail-closed composer is only as
safe as its commit gate is narrow.
The gate is now `is_curriculum_question`: claim the turn only when the
question parses to three tokens AND its terms are vocabulary a served
subject actually teaches. Everything past the gate stays fail-closed —
including `ambiguous_reading` (both terms taught, two subjects claim them)
and `out_of_curriculum` (terms taught, relation unknown), which are real
curriculum questions with honest answers. The DECIDER is unchanged, so the
lane still records untaught-vocabulary probes as declined: the curriculum
path declines them AND does not speak for them.
[Verification]: tests/test_curriculum_serve.py 28 passed (+8: five ordinary
`Does …?` questions pass through untouched, three routable ones still
answered); core test --suite deductive 268 passed; curriculum lane 32/32
wrong=0 with its pinned report SHA unchanged.
Phase 3.3 of the generalization arc. Estimation (ADR-0175), deduction
serving (ADR-0256) and curriculum serving (ADR-0262) had each written the
same seal -> ratify -> SHA-verify -> serve-gate machinery. core/ratified_ledger.py
now owns it and states the four rules once: only sealed practice writes;
tamper-evidence is structural (a load that cannot reproduce content_sha256
REFUSES); ceilings are not negotiable at the call site; absent evidence is
never a license.
Each capability keeps a thin adapter that names its artifact and preserves
its public API. One real difference is now declared rather than implied:
missing_ok distinguishes a ledger a capability SHIPS with (absence = broken
deployment, refuse) from one whose practice volume is still being built
(absence = nothing earned yet, serve disclosed) — curriculum serving is the
second kind today.
Safety property is byte-identity, asserted not assumed: re-sealing the
committed 25-band deduction ledger through the bridge reproduces it
byte-for-byte, so no artifact and no lane pin moves.
Effect: a new subject arena now needs a gold corpus and a band key, not a
re-implementation of ratification — which is what §3 meant by sequencing
the bridge ahead of the second subject.
[Verification]: tests/test_ratified_ledger_bridge.py 8 passed; core test
--suite deductive 252 passed; estimation/license test set 355 passed;
committed deduction ledger byte-identical after reseal.
Phase 2 of the generalization arc, implementing the plan's §4
curriculum-entailment gold contract. "Does force cause acceleration?" is
answered from the ratified physics chain corpus and nothing else; "Does
gravity cause acceleration?" is declined because gravity is in no pack
CORE has been taught; "Does force cause motion?" is unsettled even though
the curriculum contains both links, because nothing ratified says
causation composes.
Path (flag-gated, default off): closed question grammar -> subject routing
by ratified vocabulary -> family-scoped premise compilation from reviewed,
pack-resident chains -> the SAME argument bands (ADR-0260/0261) -> the
ROBDD engine. Zero subject-specific decision code: physics differs from
philosophy only in which rows load.
Epistemology enforced mechanically, not by convention:
- gold is a function of (curriculum, question); cases pin chain ids and
the runner FAILS a case whose pinned chain is absent or unratified;
- untaught => UNKNOWN, never "no" (open-world; silence is not denial);
- an independent oracle (own loader, ratification predicate, family table,
agreement rules, verdict rule) re-derives every gold — it disagreed once,
on "entropy reveals energy" vs "entropy causes energy", and the ORACLE
was the side that was wrong;
- anti-recall probes are a lane GUARD: a split without >=3 true-but-untaught
probes refuses to run.
Findings recorded rather than worked around (ADR-0262 §5):
- every curriculum band is UNEARNED and every answer is hedged. A band needs
n>=657 with a real outcome mix; physics teaches 7 causal + 9 modal
relations, so at most 16 questions in the subject can ever be ENTAILED.
A balanced band needs ~219 taught relations per subject×family — a ~25x
gap that only ratified curriculum content closes. Phase 2's blocker is
curriculum volume, not machinery.
- REFUTED is unreachable from present corpora (every chain is positive).
- there is NO biology domain-chain corpus; the biology OOD lane measures
fluency, not truth. The four subjects with ratified chains and mounted
packs are physics, mathematics_logic, systems_software,
philosophy_theology — the composer serves all four.
[Verification]: curriculum lane 32/32 wrong=0 with 5 anti-recall probes and
all three contract guards passing; tests/test_curriculum_serve.py 20 passed;
core test --suite deductive 252 passed; lane pinned as curriculum_serve_v1.
Completes the all/no/some square. `some` is read for the first time:
"All mammals are vertebrates. Some whales are mammals. Therefore some
whales are vertebrates." is decided entailed; "All unicorns are horned.
Therefore some unicorns are horned." is decided unsettled, with the
no-existential-import reading disclosed in the surface.
Mechanism (generate/proof_chain/exist.py): v5-VP's per-individual
lowering over a domain widened by one Skolem witness per existential
premise and — the load-bearing part — one ARBITRARY element per
existential conclusion, at which every universal still instantiates and
about which no premise asserts anything. Refuting an existential means
deriving a universal, and the arbitrary element is what makes that
derivation genuine instead of an assumption of domain closure: without
it "Rex is a wolf. Rex is not tame. Therefore some wolves are tame."
reads as REFUTED from a domain of one wolf. With it, UNKNOWN.
Four `en_exist_*` bands earned SERVE at 720x wrong=0 on the first arena
seal (25-band ledger); the 32-case hand-authored v2_exist split passed
32/32 first run; full lane 166/166 wrong=0 across six splits.
Also fixes a wrong-answer path this work uncovered in Band v1b
(ADR-0261 §5.1): `to_syllogism` FILTERED premises it could not express
out of the projection and answered from the remainder, so "Aristotle is
a philosopher. All philosophers are scholars. Therefore some scholars
are philosophers." lost its only witness and was served "that doesn't
follow". It now refuses — matching its propositional sibling
`to_deductive_logic`, which always has — and the argument falls through
to the bands that can hold a singular fact. The categorical band
re-earned SERVE 720/720 after the change; both examples are pinned as
regression cases.
Promotion: `ds-mem-0020` declined -> unknown (ADR-0258's existential
scope-out, now read; an anonymous witness never transfers to a named
individual, so UNKNOWN is its honest verdict, not entailed).
[Verification]: core test --suite deductive 232 passed; reader tests
33/33; register+surface tests 52 passed; arena 25 bands x 720 wrong=0
(all SERVE); lane 166/166 wrong=0; lane SHAs 9/10 match (public_demo
drift is pre-existing on the base commit and unrelated — evidence in
the Tier-S brief).
Generalization arc Phase 1.1 (Tier F, docs/plans/generalization-arc-2026-07-24.md):
"All philosophers teach. Socrates is a philosopher. Therefore Socrates
teaches." now reads and decides — the ADR-0258 §6.3 verb-predicate
scope-out, and the reading gate for Phase 2's subject serving.
generate/proof_chain/verb.py extends v3-MEM's per-individual lowering with
a second atom family in one shared space: membership atoms (v3-MEM's own
parsers reused verbatim) + verb atoms (individual, verb-lemma-group,
object-term). Verb universals instantiate as mem(i,C) -> [~]verb(i), so a
copula-minted membership fact discharges a verb rule. The ONE new semantic
identification is closed 3sg agreement (verb-specific irregular table +
the three regular suffix rules — deliberately NOT the noun table, which
would misroute "lives"/"leaves"). Scope-tight segmentation: single-token
name/class/verb/object shapes only; everything longer refuses typed
(tense_out_of_band joins the closed reason vocabulary). Rendered by
render_entailment_verb (member surface, UNKNOWN scoping extended to the
verb reading). Rides deduction_serving_enabled (default off); pure
widening — every previously-served argument byte-identical, no existing
lane case changed outcome.
Earned: 4 en_verb_* bands x 720 arena wrong=0 (first seal), ledger resealed
at 21 bands; corpus soundness asserted against the independent truth-table
oracle over each template's intended lowering (15,120 cases, INV-25).
Lane: v2_verb 28 hand-authored real cases 28/28 first run; full lane
134/134 wrong=0 across five splits. Surgical single-line pin update +
CLAIMS regen.
Stacked on feat/generalization-phase0 (merge that first — shared
verify_lane_shas.py / CLAIMS.md / generate_claims.py lane-ADR mapping).
[Verification]: verb reader tests 34/34; deduction battery 86/86
(surface +5, e2e +1, lane +1) in-worktree; arena seal + oracle
cross-check green; smoke + warmed_session via pre-push gate.
Plan of record for the generalization arc lands here
(docs/plans/generalization-arc-2026-07-24.md) with the risk-tiered
handoff pack (docs/handoff/generalization-2026-07/); Phase 0.1 executed.
Root-caused the three stale lane pins (docs/research/lane-drift-
investigation-2026-07-24.md):
- miner/curriculum: 5c69b741 (ADR-0244 §2.7) widened content digests
16→64 hex; intentional, pins were simply never updated. Re-pinned
surgically; SHAs triple-run deterministic.
- demo_composition: a REAL serving regression. #76's canonical-first
base precedence in resolve_surface, activated when #96 routed pipeline
turns through the resolver, served the pre-R6 canonical — the register
axis (ADR-0077 R6 substantive + ADR-0071 R4 decoration) was silently
stripped from every pipeline-served turn. T13's back-stamp then made
telemetry mirror the stripped bytes; register-tour claims went red and
the e2e tests pinning the axis (outside smoke) were failing on main.
Fix: response-first base precedence (served bytes = post-guard, post-R6,
post-R4) + new SurfaceResolution.hash_surface carrying the truth-path
canonical-precedence bytes through substrate overrides, folds, hedge,
logos-morph, and the speculative marker. compute_trace_hash folds
hash_surface, so the register axis varies the served surface ONLY and
trace_hash stays register-invariant (ADR-0069 inv C). _prior_surface
(correction binding) stays on the truth path. Dead FailureClass import
and _assessment_residual removed.
Also: generate_claims.py had raised on every run since the
deduction_serve_v1 pin landed without a _LANE_ADR entry — mapping added
(ADR-0256), CLAIMS.md regenerated, --check green.
[Verification]: register tour 6/6 claims (substantive-differs restored,
all_trace_hashes_identical held); 88 tests green in-worktree
(test_surface_resolution incl. 2 new fallback tests,
test_register_substantive_consumption incl. previously-red e2e tests,
test_grounded_open_hedge_arm, test_cognitive_turn_pipeline,
test_warmed_session_lane 10/10 T13 consistency); miner/curriculum lane
SHAs byte-identical across three runs; deduction_serve + deductive_logic
lane reports byte-identical to their unchanged pins; smoke suite via
pre-push gate.
Claude
