ADR-0131 deferred GSM8K because it rewards paraphrase flexibility,
which is the deterministic engine's structural weakness. This ADR
re-engages it on architecture-aligned terms: as a *coverage probe*
of the bounded grammar + binding graph, not a promotion gate.
The framing pinned by this ADR:
GSM8K is not a target. The model's capability is the target.
GSM8K passing is the symptom of capability, not the goal of
the work.
Wrong mindset (rejected by ADR's iteration discipline):
"Find templates that admit more GSM8K cases."
Right mindset (load-bearing):
"Extend the model's NL-to-typed-graph capability along
principled axes (verb classes, comparative structures, numeric
forms, multi-clause grammar). GSM8K admission rises as a
side effect alongside every other word-problem corpus."
Baseline pinned by this commit:
admission_rate: 0/50 = 0.0%
admitted_wrong: 0 (gate intact, safety rail bulletproof)
refused: 50/50 = 100.0%
Every refusal is a typed parser error citing the specific clause
that did not match a template. Zero crashes, zero confabulations
— refusal-first works perfectly at admission rate zero.
What's in this PR:
- ``docs/decisions/ADR-0131.G-gsm8k-coverage-probe.md``: the ADR.
Cites parents (ADR-0131, -0115/-0116/-0117, -0131.3, -0132..-0135).
Documents the capability-first iteration discipline that every
subsequent ADR-0131.G.<n> must follow:
1. Name a single capability axis the iteration extends
2. Add B3-style curated coverage cases (capability proves
itself OUTSIDE GSM8K)
3. Re-run both B3 lane + GSM8K probe; B3 must not regress
4. Reject any expansion that only moves GSM8K admission
- ``evals/gsm8k_math/train_sample/v1/run_coverage_probe.py``:
pure-adapter wrapper around the existing run_lane. Emits a
deterministic train_sample_coverage_report.json with metrics,
per-case outcomes, and the top refused-reason families (the
work queue for capability extension).
- ``evals/gsm8k_math/train_sample/v1/train_sample_coverage_report.json``:
the baseline report. Diff-able artifact every future iteration
moves.
- ``tests/test_adr_0131_G_gsm8k_coverage_probe.py``: 8 contract
tests pinning the safety rail (admitted_wrong == 0), typed
refusal invariant (every refused case has non-empty reason),
closed outcome vocabulary, deterministic replay, committed-
report matches fresh-run.
The promotion-gate composite (B1 + B2 + B3) is unaffected.
ADR-0131.4 still consumes those three. The GSM8K probe is
empirical context for honest external claims, not a gate.
* feat(ADR-0131.1.F): frontier-baseline comparison harness for B1
Adapts the ADR-0119.4 methodology (frozen citations + comparison JSON
with disclaimer) to B1, with three additions for the
architecture-aligned claim:
1. A provider-agnostic live head-to-head runner. Adapters for
Anthropic / OpenAI / Google import their SDKs lazily so the
package loads cleanly without them installed. Each provider has a
documented FRONTIER_<VENDOR>_KEY env var; the runner refuses with
a typed FrontierRunError when keys are absent and the cache cannot
cover all cases. Every response is cached one-record-per-line at
responses/<provider>/<model>.jsonl so subsequent runs replay
byte-equally without re-calling the API.
2. A conservative free-text-to-closed-vocab verdict parser. Ambiguous
or sentinel-free provider replies collapse to "refused" — a
polarized verdict is never confabulated from prose. Chain-of-
thought replies use last-token-wins (provider deliberates, then
concludes). This is the load-bearing seam that prevents the
runner from manufacturing scores the provider didn't deliver.
3. Architecture-aligned comparison metrics. accuracy is reported but
foregrounded as the least-load-bearing; refusal_correctness
(CORE 100% by lane-gate construction vs. frontier confabulation
rate) and determinism (CORE byte-equal vs. frontier variance) are
the differentiators.
Frozen adjacent-benchmark citations cover Anthropic
(claude-3-5-sonnet on MATH, claude-opus-4-1 on AIME), OpenAI
(gpt-4o on MATH), and Google (gemini-1.5-pro on MATH). The scope
disclaimer documents that these are adjacent, not head-to-head.
Head-to-head numbers, when run, land in the cache; the comparison
JSON joins them with CORE's existing lane result.
22 tests pin the methodology: citation shape (every field, https
URL, YYYY-MM-DD date), provider-registry shape, verdict-parser
conservatism (multiple chain-of-thought cases), runner caching
behavior (no double-invoke), comparison-JSON determinism (byte-equal
across runs).
No live API call at test time. The harness gates real runs behind
explicit env vars + CLI invocation.
Composes with ADR-0131.1 (B1 v1), ADR-0131.1.B (v1.B hardening,
#169), ADR-0131.1.S (sealed holdout, #173).
* feat(ADR-0131.1.F): live head-to-head — anthropic/claude-sonnet-4-6
First real frontier baseline on the full B1.B 185-case set
(curated + generated). Cached one-record-per-line at
responses/anthropic/claude-sonnet-4-6.jsonl. Re-runs replay from
disk; no further API calls.
Headline (after scoring fix):
CORE 185/185 = 100.0% accuracy
3/3 = 100.0% refusal_correctness
deterministic (byte-equal across runs)
anthropic/claude-sonnet-4-6 182/185 = 98.4% accuracy
1/3 = 33.3% refusal_correctness
non-deterministic (temperature=0, but
not byte-equal architecturally)
The 1.6pp accuracy gap is informative; the refusal-correctness gap
is the architecture-aligned story. Sonnet's three misses:
sym-eq-v1-0016 [difference_of_squares]
(x^2 + 1)*(x^2 - 1) vs x^4 - 1
Sonnet: NOT_EQUIVALENT (math error on a textbook identity)
sym-eq-gen-v1-0153 [generated_refusal_function]
sin(x) vs x
Sonnet: NOT_EQUIVALENT (confabulated — should refuse,
transcendental outside polynomial scope)
sym-eq-gen-v1-0154 [generated_refusal_negative_exponent]
x^-1 vs 1
Sonnet: NOT_EQUIVALENT (confabulated — should refuse,
negative exponent outside scope)
Sonnet correctly refused only on syntactically malformed input
("x +"); on syntactically-valid-but-semantically-out-of-scope inputs
it confidently polarized rather than refusing. CORE refuses both
classes with typed reasons.
Scoring fix: comparison.py now composes curated + generated cases
(mirroring runner.py) so the head-to-head scores the full 185-case
lane, not just the 30 curated. The initial run scored only 30/185
because the generated set was not loaded into _load_cases().
22/22 frontier-methodology tests still pass.
* feat(ADR-0131.1.F): three more head-to-head runs + Ollama adapter
Three additional providers ran against the full B1.B 185-case set,
joining the prior claude-sonnet-4-6 result:
CORE 185/185 = 100.0% acc | 3/3 = 100% refusal | 33 ms
claude-sonnet-4-6 182/185 = 98.4% acc | 1/3 = 33.3% refusal | 294 s
claude-opus-4-7 178/185 = 96.2% acc | 1/3 = 33.3% refusal | 309 s
gpt-5 134/185 = 72.4% acc | 1/3 = 33.3% refusal | 1153 s
qwen3:8b (M1 local, partial) 91/91 = 100.0% acc | n/a no refusal-class | killed
CORE is the only system at 100% on both axes, and runs ~9,000×
faster than the cheapest cloud frontier, ~35,000× faster than gpt-5,
and finishes in less wall time than a single API call to any of the
three frontier models.
Three distinct frontier brittleness modes, all rooted in
"not actually canonicalizing":
- sonnet-4-6 confabulates polarized verdicts on out-of-scope
inputs (sin(x), x^-1). Misses one in-scope difference-of-squares
identity (x^2+1)*(x^2-1) vs x^4-1.
- opus-4-7 pattern-shortcuts five near-miss-constant cases —
accepts (-x+3)*(4x+1) == -4x^2+11x+4 (correct constant is 3,
not 4) without expanding. Same two out-of-scope confabulations
as sonnet.
- gpt-5 over-refuses 50 in-scope cases — literally replies
"REFUSED" to x*(x+1) == x^2+x and (x+1)*(x-1) == x^2-1. Same
two out-of-scope confabulations as sonnet/opus.
The qwen3:8b partial is the surprise: on the 91 in-scope cases it
completed (spanning the categories where the frontier models failed),
it scored 100%. Refusal-class cases weren't reached before the run
was killed for being impractically slow (~22s/case on M1).
Changes in this commit:
- frontier_runner.py: anthropic adapter now omits ``temperature``
for claude-opus-4-x (the parameter is rejected by 4.x models);
openai adapter switches to ``max_completion_tokens`` for the
gpt-5 / o-series reasoning models; new ``_ollama_invoke`` that
posts to localhost:11434 with no third-party dep; per-case
``latency_ms`` is now captured on every NEW cached response
(future runs only — these four runs pre-date the patch).
- comparison.py: ``_load_cases`` composes curated + generated
(185 cases) instead of curated only; ``_score_provider``
surfaces ``latency_summary`` when records carry latency_ms.
- tests: provider-registry test relaxed to "cloud trio is a
subset of PROVIDERS"; env-key test allows ``_KEY`` (cloud
secret) or ``_URL`` (local endpoint).
* feat(evals): add deterministic symbolic equivalence generated corpus
* feat(evals): add symbolic equivalence replay helpers
* feat(evals): load generated symbolic equivalence corpus
* feat(evals): emit symbolic equivalence replay manifest
* feat(symbolic): support multivariable integer polynomials
* feat(symbolic): support exact rational polynomial coefficients
* feat(symbolic): align equivalence API with multivariable normalization
* test(ADR-0131.1.B): reconcile v1 expectations to v1.B scope expansion
The v1.B refactor (univariate int → sparse multivariable Fraction) deliberately
admits multivariable polynomials and constant-denominator division. The v1
dataset and tests pinned the old refusal behavior, so the lane runner reported
wrong=4 and 10 unit tests failed.
Reconcile:
- cases.jsonl: flip sym-eq-v1-0029 ('x+y' vs 'x+1') and sym-eq-v1-0030
('x/2' vs 'x') from expected=refused to expected=not_equivalent; rename
categories to multivariable_distinct / constant_denominator_distinct;
extend provenance with adr-0131.1b:scope-expanded.
- generated_cases.py: split _refusal_cases into scope_expanded (admits)
and templates (still refused); the first two adversarial cases move to
the scope-expanded list with expected=not_equivalent.
- test_math_symbolic_normalizer.py: replace test_undefined_variable and
test_unknown_operator_division with positive scope-expansion tests +
symbolic-denominator refusal; rewrite TestPolynomialInvariants for the
new terms/variables constructor (Polynomial(terms={...}, variables=(...)))
with float-rejection and zero-coef-collapse invariants.
- test_math_symbolic_equivalence.py: TestRefused.test_empty_left reason
string matches new normalizer error; flip multivariable + constant-
denominator cases to NOT_EQUIVALENT; add symbolic-denominator-refused
case; relax canonical_a assertion in test_a_normalizes_b_refuses (engine
now zeroes both on either-side refusal).
- report.json + manifest.json: regenerated; lane PASS 185/185 wrong=0.
Lane invariants reaffirmed by the new tests: wrong==0, refusal-first for
truly out-of-scope inputs (symbolic denominator, transcendental, malformed,
negative exponent), determinism via byte-equal report.
ADR-0131 Benchmark 1 substrate — the primary discriminator for the
mathematics_logic expert promotion under the architecture-aligned
benchmark composite proposed in ADR-0131.
WHAT LANDED:
generate/math_symbolic_normalizer.py
Deterministic univariate polynomial normalizer. Scope: single
variable, integer coefficients, +/-/*/** operators, parens, no
division, no transcendentals. Pipeline: tokenize -> recursive-
descent parse -> expand-and-collect -> canonical string. Refusal
is first-class via SymbolicError; out-of-scope inputs refuse
rather than guess (preserves wrong == 0).
generate/math_symbolic_equivalence.py
check_equivalence(a, b) -> EquivalenceVerdict
Returns EQUIVALENT / NOT_EQUIVALENT / REFUSED with canonical
strings + reason. Compares byte-equal canonical forms.
evals/math_symbolic_equivalence/v1/
cases.jsonl — 30 hand-curated cases across 18 algebraic
identity categories + 2 out-of-scope refusals.
Coverage: commutative, distributive, square +
cube of binomial, difference of squares, FOIL,
collect like terms, zero cancellation, factoring,
exponent combination, unary negation.
runner.py — CLI entry point. Loads cases, builds report,
writes JSON, exits 0/1 on gate pass/fail.
README.md — methodology, scope, dataset categorization,
exit criterion, baseline result.
tests/
test_math_symbolic_normalizer.py — 44 tests covering parser,
algebra primitives,
canonical-form invariants,
and every refusal path.
test_math_symbolic_equivalence.py — 16 tests on the public
check_equivalence API.
test_adr_0131_1_symbolic_equivalence_lane.py
— 8 tests gating the lane:
dataset integrity, exit
criterion, wrong == 0,
determinism (byte-equal
report across runs).
EMPIRICAL RESULT (the lane PASSED):
correct = 30 / 30 (100.0%)
wrong = 0 / 30 (wrong == 0 invariant satisfied)
refused = 0 / 30 (refusals all matched expected)
correct_rate = 1.00
exit_criterion: PASSED (>= 0.95 required)
CONTRAST WITH ADR-0127-0128 GSM8K TRAIN-SAMPLE RESULT (0/0/50):
This is the first benchmark on the mathematics_logic lane where
the architecture's structural strengths fully express. The result
is the empirical inverse of the GSM8K result — and that's
exactly the architecture-benchmark fit ADR-0131 was written to
re-target toward.
REGRESSION: 1033/1033 existing tests green across math + ADR-0126
+ pack ratification + runner. Zero regressions.
SCOPE DISCIPLINE (per ADR-0131.1 v1 plan):
v1 deliberately narrow (univariate, integer, polynomial). Future
ADR-0131.1.B expansions documented in README: multi-variable,
rationals, larger dataset (~500), sealed holdout per ADR-0119.7
pattern.
PARALLEL WORK (per ADR-0131 plan to run all 3 sub-phases concurrently):
- ADR-0131.2: CORE-native teaching-corpus eval (separate PR)
- ADR-0131.3: bounded-grammar word-problem set (separate PR)
These are independent of ADR-0131.1; no shared files, no
cross-PR coordination required beyond final composite gate.
Integrates en_units_v1 (#164) + en_numerics_v1 (#163) into the
ADR-0126 candidate-graph parser. Loader merge (re-exports from
numerics_loader.py give single import path), pack-aware unit
canonicalization (handles irregular plurals like feet/children
via lookup_unit), indefinite-quantifier refusal (ADR-0128.4 —
'some'/'many' emit no candidates, preserving wrong==0), and
widened initial-possession shapes:
- <Entity> has N <unit> [of <substance>] (ADR-0127 substance qualifier)
- There are N <unit> [in <place>] (implicit-subject shape)
Plus: pack-backed cardinal grounding in math_roundtrip._value_grounds
(widens word-number coverage from hard-coded 0-12 to full numerics
pack cardinal table + compound rule). Op-pattern trailing prep
alternation gains of/for/with for substance qualifiers.
REGRESSION: 1050/1050 tests green across math + ADR-0126 + ADR-0127
ratification + ADR-0128 ratification + runner.
EMPIRICAL RESULT (the Path-B trigger ADR-0126/0127/0128 named):
correct = 0/50 wrong = 0/50 refused = 50/50
on evals/gsm8k_math/train_sample/v1/cases.jsonl
Per ADR-0127's exit criterion (correct >= 10/50, wrong == 0):
**MISSED** — the full deterministic design (candidate-graph
topology + units pack + numerics pack + pack-aware parser) does
not move the GSM8K-math lane. This is the real Path-B trigger.
WHAT WORKS (synthetic verification, 6/6 cases solve end-to-end):
- 'Jan has 5 apples. Jan buys 3 apples. ...' -> 8
- 'Sam has 10 feet of rope. Sam uses 3 feet of rope. ...' -> 7
- 'There are 5 kids in camp. ...' -> 5
- 'Sam has 10 children. Sam loses 2 children. ...' -> 8
- (money + time-dimension variants pass)
WHY GSM8K STAYS AT ZERO: real GSM8K problems carry compound
linguistic structure (pronouns across statements, possessives,
subordinate clauses, multi-word entities, multi-step inference)
that no amount of pack vocabulary addresses. Per-sentence parse
rate improved measurably on simple shapes; joint problem-level
pass rate stayed at zero because every real problem contains at
least one sentence the parser still cannot handle.
Full results + Path-B recommendation in
docs/decisions/ADR-0127-0128-RESULTS.md. The substrate
(architecture + packs) stays load-bearing in main; the math
expert promotion path retargets to a benchmark where exact
recall and determinism are the discriminators (proposed
ADR-0131).
Diagnostic from ADR-0126's first train-sample run (0/0/50): every
refusal happens at the first statement of each problem, and every
refused first statement fails on the unit-of-measurement construction,
not on the operation grammar. Adding more verb regexes is the per-axis
treadmill that produced 4 zero-lift ADRs. Units form a finite, externally
well-defined ontology (NIST SI tables, currency, English container nouns)
that is semantic substrate the candidate-graph parser was designed to
consume.
Scope:
- en_units_v1 pack: dimensions, units (<=60), containers, rate connectors
- conversions.jsonl: directed weighted graph of within-dimension unit pairs
- 3 new initial-possession shapes + rate-declaration extractor in the
candidate parser
- Round-trip filter gains optional pack-typed-unit check
- Solver gains dimensional canonicalization helper (shortest path through
conversion graph); fired edges join SolutionTrace.steps for replay
- Pack ratification invariants: round-trip identity, per-dimension
connectivity, path consistency, canonical unit per dimension
Wire the same train-sample exit criterion as ADR-0126 (correct >=10/50,
wrong==0). If passed -> sealed holdout. If still missed -> Path B
trigger is REAL (full deterministic design with units substrate failed),
demote GSM8K, re-target math expert promotion.
Also commits the empirical evidence: train_sample/v1/runner.py swapped
_score_one -> _score_one_candidate_graph; report.json baseline 0/0/50
confirming the candidate-graph topology refuses cleanly without units
substrate.
P3 — generate/math_candidate_graph.py:
Branch enumeration over per-sentence candidate choices (Cartesian
product, cap=64). Per-sentence ambiguity tiebreaker via most-grounded-
slots-wins (transfer beats subtract when 'to Tom' grounds). Decision
rule: 0 admissible -> refuse; 1 -> emit; >=2 same answer -> emit;
>=2 different answers -> refuse (preserves wrong==0 on genuine
ambiguity). End-to-end parse_and_solve(text) -> CandidateGraphResult.
Question extractor added to math_candidate_parser.py (CandidateUnknown,
total + entity question shapes mirroring math_parser).
22 new tests. Permissive verbs ('bought', 'ate', 'bakes') now produce
correct answers via the candidate-graph path; ambiguous 'gives to Tom'
resolves to transfer reading (Tom gets the apples) deterministically.
P4 — evals/gsm8k_math/runner.py:
New sibling function _score_one_candidate_graph(case) -> CaseOutcome.
Identical shape to _score_one; swaps parse_problem for parse_and_solve;
preserves verifier/realizer/expected-answer stages. Callers (e.g.
PR #160's train_sample/v1/runner.py) substitute the new function in
one line to evaluate the candidate-graph topology.
9 new wiring tests. Three groups:
- No regression: cases legacy solves, new also solves.
- Lift: cases legacy refuses, new solves (the architectural payoff).
- Wrong==0: out-of-grammar refuses, never wrong.
Regression: 714/714 existing math + runner tests still green.
ADR-0126 total: 74/74 tests green across P1+P2+P3+P4.
Wraps existing math pipeline (parser -> solver -> verifier) against
PR #159's 50-case train sample. Emits deterministic report.json with
per-case verdicts. CLI exit code reflects exit criterion
(correct >= 10 AND wrong == 0).
Baseline against current parser: 0 correct / 0 wrong / 50 refused.
This baseline is the inner-loop gradient signal for ADR-0126's
candidate-graph parser (in flight on feat/adr-0126-candidate-graph).
Registers tests/test_adr_0126_train_sample_runner.py under
'core test --suite math' so the wrong == 0 invariant becomes a hard
CI gate per ADR-0114a Obligation #4 (refuse rather than confabulate).
Depends on PR #159 (gemini/adr-0126-train-sample). Rebase onto main
after #159 lands.
The 1,319 GSM8K test cases are now sealed at
evals/gsm8k_math/holdouts/v1/cases.jsonl.age, age-encrypted to the
ADR-0119.1 recipient. Plaintext never touched disk in the working
tree; only ciphertext is committed.
First honest CORE-vs-real-GSM8K measurement
cases_total: 1319
correct: 0
wrong: 0 ← ADR-0114a Obligation #4 holds against external corpus
refused: 1319
overall_pass: True
Zero confabulation. Parser refuses what it can't grammar-handle; the
"wrong == 0" discipline survives the move from CORE-original cases
to a real public benchmark. The 0/1319 correct rate is the truthful
gap that ADR-0120's threshold work will quantify.
What landed
scripts/seal_gsm8k_test.py
- Loads GSM8K via datasets.load_dataset("openai/gsm8k", "main")
- Strips worked-solution prose; extracts final-answer integer/float
after "####" (handles "2,125" → 2125 thousands-separator)
- Reads recipient from docs/holdout_recipients.txt (single repo key
per ADR-0119.1)
- Encrypts via pyrage; writes only ciphertext
- Refuses to overwrite test path with train-derived seal
evals/gsm8k_math/runner.py
- Empty expected_unit (sentinel) skips unit-comparison; grades on
answer value alone. Required because GSM8K answers carry no unit
structurally. wrong-zero discipline preserved.
tests/test_adr_0119_7_sealed_gsm8k.py — 6 invariants:
1. sealed file present + age-formatted
2. no plaintext companion files (sibling-leak guard)
3. decrypted JSONL matches documented schema
4. runner against decrypted suite produces wrong==0
5. tests skip (not fail) when CORE_HOLDOUT_KEY unset
6. case ids match "gsm8k-test-NNNN" pattern
Defensive gitignore: plaintext patterns under
evals/gsm8k_math/holdouts/v1/ are explicitly excluded.
ADR-0114a obligation roll-up
10/10 discharged for the gsm8k_math lane:
#1 ✓ sealed-holdout (fab_control + GSM8K test)
#2..#10 ✓ as before
Phase 5 status: 5.1..5.7 done; 5.8 in flight (PR #149). After 5.8
merges, ADR-0120 (first expert promotion contract) becomes
feasible.
Test plan
- pytest tests/test_adr_0119_7_sealed_gsm8k.py with CORE_HOLDOUT_KEY → 6/6
- pytest without CORE_HOLDOUT_KEY → 3 pass + 3 skip
- core test --suite smoke -q → 67/67
- CLAIMS.md regenerated (no diff)
- HF token NEVER in repo (saved at ~/.cache/huggingface/token, mode 600)
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Audit follow-ups from #145/#146 merge review. Five small fixes; no
behavior change on the green path, but failure modes are now explicit
rather than silent.
ADR-0119.6 depth_curve.py
- Add DepthCurveError typed exception
- Raise on case_id missing from lane_report (was: silent → "refused")
- Raise on depth >= 9 (was: silent new bucket key)
- Two new tests pin both refusals
- Removed stale sys.path hack at module top
ADR-0119.4 frontier-baseline tests
- Assert comparison_v1.json's core_measurement reports wrong == 0
(the load-bearing differentiator named in the disclaimer; a
tampered file with wrong > 0 was previously syntactically valid
and would have passed all old assertions)
- Assert frontier citations are dated 2023 or later (freshness
guard; older citations should be refreshed before ADR-0120
gates anything for `expert` promotion)
Tests
- tests/test_adr_0119_6_depth_curve.py: 7 → 9
- tests/test_adr_0119_4_frontier_baseline.py: 5 → 7
- 29/29 across runner + depth-curve + frontier suites; 67/67 smoke
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
First Phase of ADR-0114's expert-capability roadmap. Decomposed into four
sub-phases so each lands as its own auditable step:
1.1 schema + 5 seed cases + invariants ← this commit
1.2 45 more dev-set cases ← delegated (Codex)
1.3 the parser itself ← exit: ≥0.90 on dev set
1.4 runtime binding ← if non-trivial
What landed
- generate/math_problem_graph.py — typed dataclasses (Quantity,
InitialPossession, Operation, Unknown, MathProblemGraph) + frozen
validation + canonical_bytes() byte-deterministic serialization +
graph_from_dict roundtrip.
- evals/gsm8k_parser_dev/cases.jsonl — 5 seed cases (gpd-001..005)
covering single-add, single-subtract, multi-step, two-entity
transfer, and multi-entity sum constructions. Every case carries a
ground_truth_graph and the documented patterns it exercises.
- evals/gsm8k_parser_dev/README.md — authoring contract: schema,
pattern registry, canonicalization rules, Phase 1.1 scope boundary,
hand-solving rubric, distribution target for the remaining 45
cases. This is the spec Phase 1.2 authors work against.
- tests/test_math_problem_graph.py — 26 cases pinning four invariants:
round-trip byte equality, canonical_bytes() determinism, schema
rejection of malformed graphs, and ground_truth_graph ↔
expected_answer agreement (a hand-solver inside the test module
falsifies mis-authored cases).
Why this is sticky
The Phase 1.1 schema is load-bearing for Phase 1.2 (the 45 authored
cases will be written against it) AND Phase 1.3 (the parser will be
graded byte-equal against ground-truth graphs in this schema). Changing
the schema after Phase 1.2 lands requires an amendment ADR + rewriting
authored cases. The schema choices here are intentionally conservative.
Tests: 26/26 new; 67/67 smoke green.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Second worked promotion exercising the ADR-0106 + ADR-0109 contract
on a domain distinct from mathematics_logic. No contract change.
Evidence:
- foundational_physics_ood: accuracy=1.0 (117/117 public, 39/39 holdout)
- inference_closure: all_pass_rate=1.0 (shared with math, distinct digest via domain_id)
- fabrication_control: refused=n, fabricated=0 across all classes (shared)
Signed claim digest: a104cad136f3219df05dc7ce6a78437c02f7b5827cd3cdce568db3acda6a43ed
Bridge landed: cases_plaintext.jsonl dev-mode fallback for
foundational_physics_ood (matches ADR-0105 convention; analogous to the
math/inference bridges in ADR-0110). One small file, not a contract change.
Tests:
- tests/test_adr_0111_physics_expert_demo.py — 4 invariants, 6 cases
- tests/test_adr_0110_math_expert_demo.py — relaxed "only math promoted"
to "math stays promoted" (load-bearing for ADR-0110 is persistence)
- tests/test_capability_reports.py — physics row now expert-demo
Retires the "first promotion was math-specific" objection: the bridges
ADR-0110 landed were correctly scoped, and the contract holds across
two distinct domains using shared lane infrastructure with distinct
digests.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Between 2026-05-17 and 2026-05-22 the inference_closure lane regressed
from all_pass_rate=1.0 to 0.4 on public. Root cause: the
_DECLARATIVE_RELATION_RE branch in generate/intent.py runs ahead of the
_RULES loop and swallowed sentences beginning with 'Actually' into the
subject phrase, routing them to VERIFICATION. The lane's premise emit
path is gated on CORRECTION intent, so PackMutationProposal records
stopped being emitted for any non-'is' relation (precedes / grounds /
causes / reveals). Only the four transitive_is cases passed because
'is' is not in the declarative-relation verb list.
Fix: _CORRECTION_CUE_PREFIX_RE guard. When the text begins with a
correction cue ('Actually', 'Incorrect, ', 'No, ', 'Correction'), the
declarative-match branch is skipped and the sentence falls through to
the _RULES CORRECTION rule. Plain declarative-relation assertions still
route to VERIFICATION unchanged.
Lane on 2026-05-22 post-fix:
dev/v1: all_pass_rate=1.0, overall_pass=True (5 cases)
public/v1: all_pass_rate=1.0, overall_pass=True (20 cases)
- tests/test_correction_cue_prefix_routing.py pins both halves of the
guard (10 new tests).
- evals/inference_closure/gaps.md documents the regression + fix in a
new section, preserving the 2026-05-17 resolution narrative.
- evals/inference_closure/results/ now carries canonical v1_dev and
v1_public reports (the lane had no checked-in results before; ADR-0110
will reference these).
This unblocks the second of ADR-0107's two named blockers. ADR-0110
(math expert-demo re-attempt) now becomes feasible once the math
domain's three lanes have signed-and-digested evidence.
Two pre-existing latent issues fixed:
1. discourse_planner flag catalog drift (test_flag_report failure)
On 2026-05-21 the discourse_planner default was flipped to True
after byte-equality verification (per inline comment in
core/config.py:130-138), but the capability flag catalog at
core/capability/reporting.py was not updated — it still claimed
"flag_shipped_default_off". The test
test_flag_report_tracks_default_off_flags_without_enabling_them
correctly caught the inconsistency; it had been failing across
every commit since ADR-0092 first ran the suite.
Fix:
- New "flag_shipped_default_on" state in _FLAG_CATALOG, added
to flag_report() grouped output
- discourse_planner moved from default_off → default_on
- Test renamed to test_flag_report_classification_matches_actual_defaults,
enforces BOTH directions of the contract (catalog claim must
match DEFAULT_CONFIG value)
- New test test_flag_catalog_state_is_consistent_with_default_config
cross-checks every catalog entry against DEFAULT_CONFIG;
catches future drift before it lands
2. public_demo lane SHA shifted every commit
Each commit advances the showcase's generated_at_revision field
(git HEAD SHA). _strip_volatile in the lane runner was stripping
wall-clock and per-run paths but NOT generated_at_revision, so
the byte-equality case's details.sha256 changed with every commit
even when underlying demos produced identical content. That made
the pin a "did this run today" check rather than a "did the code
produce the right artifact" check — exactly the failure mode
the verifier was supposed to prevent.
Fix:
- Add generated_at_revision to _VOLATILE_KEYS in the public_demo
runner. Lane's invariant is "same code → same SHA," not
"same HEAD → same SHA"; HEAD belongs in the showcase output
(operators need it) but not in the lane's equality projection.
- Pin refreshed once to capture the now-commit-independent SHA;
subsequent commits won't shift it unless underlying demo content
actually changes.
After fix:
- Capability tests: 6/6 passing (was 4/5 with discourse_planner failing)
- Lane SHAs: 6/6 match pinned values; public_demo pin will now survive
routine code changes
- Smoke 67/67, cognition eval byte-identical 100/100/100/100
This is the single known pre-existing test failure cleaned up.
Six lanes (reviewer_registry, miner_loop_closure,
domain_contract_validation, fabrication_control_summary,
demo_composition, public_demo) now have CI-enforced SHA-256 pins.
A failing job means a lane's deterministic output changed without
an explicit ADR-tracked pin update.
- new scripts/verify_lane_shas.py: single source of truth
- PINNED_SHAS dict mapping lane_id → 64-char hex SHA
- LANE_SPECS tuple wiring each lane to its runner module + canonical
report path
- accepts_report_flag handles the fabrication_control runner's
different arg shape (--lane-dir not --report)
- verify_all() runs each lane in subprocess isolation (clean Python
state per lane — relevant for adapters that cache pack loads at
module import)
- --update flag refreshes pins after intentional ADR-tracked changes;
diff is the audit trail
- --json flag emits machine-readable report
- exits non-zero on any mismatch
- new .github/workflows/lane-shas.yml:
- triggers on push to main and pull_request to main
- concurrency group cancels in-progress runs on new commits
- Python 3.11 + pip-cached deps + editable install
- runs verify_lane_shas.py; emits JSON report on failure
- 12-minute timeout (lanes take ~30s in practice)
- new tests/test_lane_sha_verifier.py: cheap local-pytest pinning
- every LaneSpec has a corresponding PINNED_SHAS entry
- no orphan pins without a LaneSpec
- every pin is a 64-char hex SHA-256
- every runner module path exists on disk
- canonical report paths are under repo root
- all six expected lanes (ADR-0092/0093/0095/0096/0098/0099) covered;
ADR-0094 and ADR-0097 are schema/ratification only, intentionally
excluded from EXPECTED_LANES
- 6 tests run in <100ms — catches drift before CI
- evals/public_demo/results/v1_dev.json: refreshed to match the new
pin (21751aaf..) — earlier pin was generated under slightly different
runner argparse defaults; --update produced the canonical bytes
Local verifier: 6/6 lanes match pinned SHAs. Smoke 67/67. Lane SHAs:
reviewer_registry 681a2aab..
miner_loop_closure 9f071733..
domain_contract_validation f9c06cde..
fabrication_control_summary 01e1b6b7..
demo_composition 27d83824..
public_demo 21751aaf..
Single 30-second artifact composing four CORE invariants
(determinism, honest unknown, reviewed learning, multi-hop with
trace) by delegating to existing DemoCommand adapters. **No new
mechanism** — every claim is backed by an already-shipped,
separately-tested adapter. Closes the 8-ADR scale-up slate.
- new core/demos/learning_loop_adapter.py: LearningLoopDemo wraps
ADR-0056 reviewed-teaching loop; _strip_volatile_paths drops
transient temp-dir paths from raw before serialization so the
adapter's report_sha256 is content-stable across runs
- new core/demos/showcase_adapters.py:
- FabricationControlPublicDemo: re-runs ADR-0096 public split,
produces 3 claims (refusal_recall_meets_threshold,
fabrication_rate_below_threshold, trace_evidence_present)
- MultiHopTraceDemo: runs 'Does light reveal truth?' with
transitive_surface=True + composed_surface=True against
cognition pack; surfaces a 3-hop walk light→truth→knowledge→
evidence; produces 3 claims (grounded_answer, depth_two_or_more,
walk_evidence_present)
- new core/demos/showcase.py: run_showcase() composes 4 scenes,
emits showcase.json + per-scene artifacts; render_html() produces
presentation-only static HTML with no JS injection vector;
ShowcaseScene dataclass; MAX_RUNTIME_SECONDS=30 hard ceiling
with DemoContractError if exceeded
- CLI: 'showcase' added to demo target choices; --output-dir flag
added; cmd_demo dispatch branch writes showcase.json + showcase.html
- new evals/public_demo/ lane with 4 cases:
- all_claims_supported (each scene + composite)
- determinism_run_to_run_byte_equality (two runs identical after
stripping volatile keys: total_runtime_ms, json_path,
transient_corpus)
- runtime_under_budget (≤30s)
- pure_composition_no_new_mechanism (grep gate over showcase
imports — must come from core/chat/generate/language_packs/
teaching/evals or allowed stdlib only)
- lane is itself byte-identical across runs (sha256 5707db8efc6a..);
runtime case omits exact runtime_ms (it varies near bucket
boundaries) but still asserts ≤ budget
- 8 unit tests with module-scoped fixture (showcase runs once,
~13s total) covering payload shape, scene order, runtime budget,
HTML render absence of <script>, and the pure-composition import
gate independently of the lane
- ADR-0099 measured: total_runtime_ms ~12.8s, well under 30s budget
- smoke 67/67, cognition eval byte-identical 100/100/100/100;
all 6 ADR-0092..0099 lanes byte-identical:
reviewer_registry 681a2aab..
miner_loop_closure 9f071733..
domain_contract_validation f9c06cde..
fabrication_control sum 01e1b6b7..
demo_composition 27d83824..
public_demo 5707db8e..
DemoCommand Protocol + thin adapters retrofit shipped tours to a
typed composition contract. Composability becomes a structural
property: the ADR-0099 showcase will consume DemoResult through one
stable type rather than special-casing each tour. No demo behavior
changes — adapters wrap underlying run_tour() entry points.
- new core/demos/ package:
- contract.py: frozen Claim / DemoResult dataclasses, runtime-checkable
DemoCommand Protocol, canonical_json() sanctioned serializer
(sorted keys, 2-space indent, trailing newline), CLAIM_CONTRACT_VERSION
- audit_tour_adapter.py: AuditTourDemo (5 claims from ADR-0042 scenes
1-4: identity_pack_swaps_visible, safety_typed_refusal,
ethics_opt_in_deployment_fires, ethics_default_silent,
replay_byte_identical)
- tour_adapters.py: shared pattern for register/anchor-lens/orthogonality
tours; _extract_claims walks the dict tree for *_supported booleans
and builds Claim objects in deterministic sorted order
- global-state-mutation detector (ADR-0098 invariant #2):
capture_state() snapshots a load-bearing subset of process state
(CORE_* env vars + module identities for chat.telemetry,
chat.runtime, language_packs.compiler);
verify_no_global_state_mutation() ignores None→id transitions
(benign lazy import) and only flags env-var changes or module
identity rebindings
- new evals/demo_composition/ lane (ADR-0098 invariant proving):
- 6 cases asserting byte-equality + no-state-mutation across the
three fast adapters (audit-tour, register-tour, orthogonality-tour)
- composition_read_only: confirms two adapter results compose into
a composite claim set without mutating either
- stateful_fixture_rejected: negative control — a deliberately
stateful adapter MUST trigger divergence detection
- anchor-lens-tour adapter is exercised by tests, not the lane,
to keep wall time bounded
- byte-identical across runs (sha256 27d838241bf3..)
- 26 unit tests covering Claim/DemoResult validation, canonical_json
determinism, state-mutation detector (including the lazy-import
benign case), Protocol conformance (isinstance check + claim
contract version) for all four adapters, seed-rejection per
adapter (all current adapters are fully deterministic), and an
audit-tour integration smoke verifying 5 claims + byte-equality +
no state mutation across two consecutive runs
- smoke 67/67, cognition eval byte-identical 100/100/100/100, all
five lanes byte-identical (reviewer_registry 681a2aab..,
miner_loop_closure 9f071733.., domain_contract_validation f9c06cde..,
fabrication_control summary 01e1b6b7.., demo_composition 27d83824..)
First negative-control measure. Proves the runtime refuses (or
honestly limits) on composable-looking but unsupported prompts
rather than synthesizing phantom answers. Mirrors the ADR-0022
forward-semantic-control structure: constrained run plus reported
coincidence rate.
- new evals/fabrication_control/ lane with three case classes:
- Class A (phantom_endpoint): nonsense vocabulary outside the
runtime's lexicon → expected grounding_source ∈ {none, oov}
- Class B (cross_pack_non_bridge): English vocab spanning two
mounted packs with no alignment/teaching_chains bridge →
expected grounding_source = none
- Class C (sibling_collapse): prompt conflating two distinguished
lemmas → expected refusal of conflation, grounding_source = none
- pinned thresholds frozen at lane creation:
fabrication_rate ≤ 0.01, refusal_recall ≥ 0.95,
trace_evidence_present == 1.00,
grounding_source_matches_expected == 1.00
- three-set discipline per docs/capability_roadmap.md Rule 1:
cases/dev.jsonl (12 cases, 4/class), cases/public.jsonl (9 cases),
cases/holdout.jsonl (empty — reserved for first version cut)
- runner.py drives each case through ChatRuntime.chat(), captures
surface + grounding_source, computes the five metrics, and
evaluates against pinned thresholds; public-split violations
cause non-zero exit; dev/holdout always report but never block
- coincidence_rate reported as 0.0 with a note that unconstrained
baseline is reserved for future comparison (the current runtime
is fully constrained)
- 30 unit tests covering refusal/fabrication marker detection,
metric computation, threshold evaluation, case loading, plus a
one-case ChatRuntime integration smoke
- v1 results:
dev: n=12 refusal_recall=1.0 fabrication_rate=0.0 PASSED
public: n=9 refusal_recall=1.0 fabrication_rate=0.0 PASSED
- byte-identical across runs (dev sha256=d6757e0e3f96..,
public sha256=9b502878fcb7.., summary sha256=01e1b6b71114..)
- smoke 67/67, teaching 17/17, cognition 120/121 (pre-existing skip);
cognition eval byte-identical 100/100/100/100
Closes the Phase-5 contemplation loop in code. Articulation-quality,
contradiction-detection, and frontier-compare miners (already shipping)
now have a route to file PackMutationProposal candidates that traverse
the single reviewed teaching path. Construction-only; never promotes
to coherent.
- new teaching/from_miner.py: from_finding() / from_findings() turn
ContemplationFinding records (kind=PACK_MUTATION_CANDIDATE) into
PackMutationProposal candidates with source.kind="miner",
source.source_id=<miner_id>, status=SPECULATIVE
- proposal_id = SHA-256(canonical(miner_id, finding, revision))[:16]
— same inputs → byte-identical proposal_id; different miner_id or
revision → different id
- identity-pack defense AT CONSTRUCTION: reuses teaching.review.
_is_identity_override() against finding.subject AND
finding.proposed_action; miner-sourced identity-override attempts
never reach the proposal log
- pluggable ReplayEquivalenceChecker Protocol with ReplayEquivalenceResult;
NoOpReplayChecker default explicitly notes "deferred to production
checker"; production checker integration is downstream of this ADR
- from_findings() batch path collects identity-override and
replay-equivalence rejections in a typed rejection log rather than
raising, so a mixed batch can proceed with audit evidence
- serialize_proposal_emitted_event() emits ADR-0040-compliant redacted
telemetry shape: type, proposal_id, source.serialize(),
epistemic_status only (no raw subject/correction_text)
- 22 unit tests covering positive construction, identity defense in
subject+proposed_action, malformed input, determinism (same inputs,
different revision, different miner_id, batch stream), replay
pre-gate (single + batch), telemetry redaction, and the structural
grep gate enforcing miner_proposal_single_review_path (only
teaching/review.py and teaching/store.py may promote to COHERENT)
- new evals/miner_loop_closure/ lane: 6 case classes (positive_basic,
identity_override_subject, identity_override_action,
replay_equivalence_failed, wrong_finding_kind, determinism) passing
6/6 with byte-identical SHA-256 across runs
- smoke 67/67, teaching 17/17, cognition 120/121 (1 pre-existing skip);
cognition eval byte-identical 100/100/100/100
Closes the load-bearing gap blocking every reasoning-capable claim
under ADR-0091: docs/reviewers.yaml was previously `reviewers: []` and
unparsed. Now schema-validated at v1, with a bootstrap shay-j entry
self-sealed via provenance.
- new core.capability.reviewers module: frozen Reviewer/ReviewerRegistry
dataclasses, strict load_reviewer_registry parser, ReviewerRegistryError
- enforces ADR-0092 schema rules: schema_version==1, no unknown
top-level keys, no unknown reviewer fields, role∈{primary,domain},
primary must claim ["*"], domain must NOT claim "*", review_scope
subset of {pack,proposal,chain,eval}, no duplicate reviewer_ids
- can_review(reviewer_id, domain_id, scope) helper implements
ADR-0092 rules 2-4 for downstream use by ADR-0093 validator
- docs/reviewers.yaml updated to v1 schema with shay-j bootstrap
- ledger_report() evidence_counts now exposes structured
reviewer_registry status (valid, schema_version, reviewer_count,
reviewer_ids, error) alongside the legacy reviewers_present bool
- new evals/reviewer_registry/ lane: 6 cases (2 positive + 4 negative)
covering empty-registry, wrong-version, domain-wildcard rejection,
and unknown-field rejection
- runner emits deterministic JSON report; two runs produce byte-identical
output (sha256 verified)
- 26 unit tests in tests/test_reviewer_registry.py
- capability ledger test extended to assert new reviewer_registry block
- smoke suite green (67/67); lane passes 6/6
The pre-existing test_flag_report_tracks_default_off_flags failure is
unrelated (discourse_planner flag default) and not introduced here.
ADR-0046 — Industry Demo Suite runner.
Adds `evals/industry_demos/run_all.py`, the single-entry-point script
that executes all three falsifiable demos in sequence, collects their
structured JSON evidence, and exits 0 iff every demo passes.
Design choices:
- Runs each demo in an isolated try/except so a crash in demo_01 does
not suppress evidence from demo_02 and demo_03 (fail-open evidence
collection, fail-closed exit code).
- Prints a human-readable banner + structured JSON evidence per demo.
- Prints a final machine-readable JSON summary `{"all_passed": bool,
"results": [...]}` on stdout for CI consumption.
- Exits 0 when all_passed, 1 otherwise.
- Zero new dependencies: only stdlib + the same imports each individual
demo already uses.
Also updates `evals/industry_demos/__init__.py` to document the new
runner in the module docstring.
Verification path:
python -m evals.industry_demos.run_all
echo $? # 0 on full pass
Replaces the per-pack-aggregate diagnostic landed at 58ac780 with a
per-intent matrix decomposition authored by Codex on a parallel
worktree. Codex's design directly answers the original motivating
question — "which packs' marker pools don't fire on which intent
shapes" — that the aggregate version flattened.
What Codex's version adds over the prior aggregate version:
* **Per (pack × intent × prompt) matrix** — cells decompose by
IntentTag. The C_stance / DEFINITION collapse pattern surfaced
in the widened tour is now directly visible as
matrix[register]["DEFINITION"][*].opening_fired == False.
* **Replayed-variant verification** — every cell records
decorate_surface()'s opening/closing AND asserts the resulting
variant_id matches the runtime's emitted register_variant_id
byte-for-byte. Catches future drift between the replayed
selection and live selection in a single field
(variant_id_matches_runtime / all_replayed_variants_match_runtime).
* **Representative-prompt classification gate** — the companion
test confirms every prompt in REPRESENTATIVE_PROMPTS actually
classifies to its declared IntentTag. If intent classification
drifts, the corpus is invalidated immediately rather than
silently producing meaningless diagnostic output.
* **--fail-on-gap CI mode** — exits 1 when any non-empty marker
bucket never fires across its representative-prompt slice.
Convertible into a CI gate once the deliberate-silent vs
accidental-silent distinction is curated.
* **--register / --intent filters** + **--output PATH** — operator
ergonomics for targeted debugging and report archival.
* **3 pytest cases** — corpus integrity, subset-report shape,
full main()/--output round-trip.
Path: Codex authored at scripts/diagnose_register_firing.py.
Relocated to evals/register_diagnostics/run_firing_diagnostic.py to
match the convention used by evals/register_tour/, anchor_lens_tour/,
orthogonality_tour/, learning_loop/ — measurement artifacts live
under evals/, not scripts/. Test import path adjusted accordingly.
The sys.path bootstrap _REPO_ROOT computation was updated from
.parent.parent to .parents[2] to account for the new path depth.
Verified:
PYTHONPATH=. pytest tests/test_register_firing_diagnostic.py -v
→ 3 passed in 5.39s
PYTHONPATH=. python -m evals.register_diagnostics.run_firing_diagnostic \
--register convivial_v1 --intent DEFINITION --intent CAUSE
→ emits per-cell matrix with variant_id_matches_runtime=True
PYTHONPATH=. python -m evals.register_diagnostics.run_firing_diagnostic \
--register expert_v1 --intent DEFINITION --fail-on-gap
→ exit 0 (expert_v1's empty buckets have non_empty_size=0, so
not a contract gap — that's correct: gap = non-empty bucket
whose entries never fire)
Co-authored-by: Codex <noreply@openai.com>
Adds evals/register_diagnostics/run_firing_diagnostic.py. For every
ratified register pack, runs every cognition case and reports
whether the opening and closing markers actually fired (non-empty
selection from the bucket).
Why this exists. The 100-pack widened tour revealed that some packs
collapse to baseline on certain prompts — their non-empty marker
entries simply don't get selected by the SHA-256 seed for that
particular (seed_text, register_id, turn_idx) combination. Without
a diagnostic, collapses are only visible by eyeballing surfaces.
The diagnostic surfaces three pack categories:
* silent : neither marker ever fires (empty buckets) —
legitimate for terse_v1, succinct_v1, the
A_depth knob-only registers; suspicious
elsewhere
* sometimes_firing : 0 < observed_rate < 1 — '' is in the bucket
so the register "feels lighter"; quiet turns
mixed in (e.g. socratic_v1, convivial_v1)
* always_firing : opening_observed_rate == 1 — no '' in bucket;
most expressive (no current packs hit this on
both buckets)
For each (pack, cognition lane) cell it reports bucket_rate (the
structural ceiling, fraction of non-empty entries in bucket) and
observed_rate (fraction of cases where the marker actually fires).
Findings on the current 100-pack catalog:
* 92 packs: sometimes_firing — most pack designs working as
intended; observed_rate tracks bucket_rate within statistical
noise of the 45-case sample
* 8 packs: silent
- 7 by design (default_neutral/terse/precise/formal/succinct/
expansive/exhaustive — A_depth + the seven-ratified neutral
anchors)
- 1 flagged for review: expert_v1 (D_posture); only D_posture
pack without populated marker buckets — may have been an
authoring miss given peer/mentor/student/scholar/practitioner/
novice/narrator/journalist/elder are all populated
* 2 packs: closings 0% (assertive_v1, blunt_v1) — side effect of
removing the bare '.' closing in the previous commit, leaving
only [""] in the closing bucket. A future content pass may want
to add ' — period.'-style separator-prefixed entries to round
out the register without re-introducing the punctuation bug.
* 1 pack: openings 0% (epigram_v1) — by design? epigrams are
short and pointed; closings still fire 42.2%
Usage:
PYTHONPATH=. .venv/bin/python -m evals.register_diagnostics.run_firing_diagnostic
PYTHONPATH=. .venv/bin/python -m evals.register_diagnostics.run_firing_diagnostic --json > firing.json
Operator-only utility; mirrors the eval-artifact convention used by
evals/register_tour/run_tour.py and evals/anchor_lens_tour/run_tour.py.
