dbeb1b2f00
* fix(tests): replace brittle git-status partition assertion with behavioral invariant * docs(ADR-0167): record closure of brittle partition git-status assertion * fixup: restore FOLLOWUPS §6 (holonomy ablation) — unresolved, just shipped in #360
243 lines
10 KiB
Markdown
243 lines
10 KiB
Markdown
# ADR-0167 — Follow-ups Queue
|
|
|
|
**Date opened:** 2026-05-27 (end of Wave 3)
|
|
**Parent:** [ADR-0167](../decisions/ADR-0167-audit-as-teaching-evidence.md)
|
|
**Companion:** [SESSION-2026-05-27](../decisions/SESSION-2026-05-27-adr-0167-parallel-dispatch.md)
|
|
|
|
The LexicalClaim slice landed clean (W1-A → W2-A/B/C/D → W3-A merged
|
|
2026-05-27). This file captures the named follow-ups that surfaced
|
|
during the wave and were deliberately deferred so the slice could
|
|
converge. The next operator picking up ADR-0167 work should walk this
|
|
queue top-down and decide which item the project's current capability
|
|
gate actually needs.
|
|
|
|
Each item: scope, why deferred, where the breadcrumbs live, and the
|
|
acceptance criterion that would close it.
|
|
|
|
---
|
|
|
|
## 1. Frame-opener sub-types (FrameClaim / CompositionClaim / ReferenceClaim / SlotClaim)
|
|
|
|
**Scope.** Four additional ratification handlers, one per remaining
|
|
sub-type in `SUB_TYPE_FOR_OPERATOR`:
|
|
|
|
| Sub-type | Maps from | Ratification primitive |
|
|
|---|---|---|
|
|
| `FrameClaim` | `pre_frame_filler_sentence`, `multi_subject_sentence` | Verb-category reclassification |
|
|
| `CompositionClaim` | `multi_quantity_composition`, `quantity_extraction` | Frame-split rule |
|
|
| `ReferenceClaim` | `pronoun_resolution` | Anaphora-resolution entry |
|
|
| `SlotClaim` | `question_frame_slot`, `unit_binding`, `question_target_slot`, `descriptive_frame_question` | Slot-completion table entry |
|
|
|
|
**Why deferred.** ADR-0167 §"Proposed sub-type set" explicitly chose
|
|
LexicalClaim-first because it is the lowest-risk surface (drain_token
|
|
additions cannot create wrong admissions without also passing graph
|
|
completeness). The other four touch frame-opener decisions, anaphora,
|
|
or slot bindings — each is a new admission path multiplying the
|
|
`wrong=0` surface area, and each needs its own scoping ADR with the
|
|
six open questions from ADR-0167 §"Open questions" answered for that
|
|
sub-type's mechanics.
|
|
|
|
**Where breadcrumbs live.**
|
|
- `teaching/math_evidence.py::SUB_TYPE_FOR_OPERATOR` — already maps
|
|
the operator labels; no schema change needed
|
|
- `teaching/math_lexical_ratification.py` — the template for what a
|
|
sub-type handler looks like (preconditions, receipt, idempotency,
|
|
hazard pins)
|
|
- `tests/test_math_lexical_ratification.py::test_rejects_non_lexical_sub_type`
|
|
— pins that non-lexical claims currently raise `WrongClaimSubType`;
|
|
each new handler retires its corresponding rejection
|
|
|
|
**Acceptance.** Each new sub-type ships as an ADR (likely ADR-0168,
|
|
ADR-0169, ...) followed by a wave of PRs analogous to ADR-0167's W2-D.
|
|
Each handler must:
|
|
- declare its own `SAFE_CATEGORIES` allowlist analogous to W2-D's
|
|
`{"drain_token"}`
|
|
- preserve the case 0050 hazard pin
|
|
- carry the same idempotency / evidence-tampering / unknown-category
|
|
guards W2-D established
|
|
- pass an e2e ratification → row-movement test analogous to
|
|
`test_lexical_ratification_advances_unknown_word_row`
|
|
|
|
**Priority hint.** FrameClaim is the highest-leverage next sub-type
|
|
(9 cases in the current taxonomy under `pre_frame_filler_sentence`),
|
|
but also the riskiest — frame-opener miscategorisation is exactly the
|
|
case 0050 hazard. CompositionClaim (8+11 cases) is the next-highest
|
|
count, also high-risk. ReferenceClaim (3 cases) and SlotClaim (smaller
|
|
buckets) are lower-leverage but structurally simpler.
|
|
|
|
---
|
|
|
|
## 2. Partition test architectural fix
|
|
|
|
**Status.** Closed by `fix/adr-0167-partition-test-invariant`.
|
|
|
|
The original `git status --porcelain` runtime assertion was retired.
|
|
Partition invariants are now expressed behaviorally through:
|
|
|
|
- serialization discrimination (`domain` omitted for cognition,
|
|
explicit for math)
|
|
- deterministic canonical-byte divergence between cognition and math
|
|
candidates
|
|
- existing cognition regression suites already exercised in CI
|
|
|
|
This removes the structurally brittle requirement that every future
|
|
ADR-0167 PR edit a filename allowlist merely to add a new evidence test.
|
|
The partition guarantee now lives at the protocol surface instead of
|
|
repository working-tree state.
|
|
|
|
---
|
|
|
|
## 3. Pre-existing main test failures (unrelated to ADR-0167)
|
|
|
|
**Status update.** The two failures originally flagged during W3-A were
|
|
later traced to a real wrong=0 hazard and fixed in #359 rather than
|
|
being treated as unrelated noise.
|
|
|
|
The relevant fix path:
|
|
- recognized-but-uninjectable statements now refuse instead of silently
|
|
admitting partial graphs
|
|
- audit taxonomy updated accordingly
|
|
- regression coverage added for recognizer skip-only fallback behavior
|
|
|
|
Retain this section as historical context only; do not reopen unless a
|
|
fresh regression appears.
|
|
|
|
---
|
|
|
|
## 4. Workbench v1 — math candidate rendering
|
|
|
|
**Scope.** Make the read-only operator UI (ADR-0160/0162) render
|
|
`MathReaderRefusalEvidence` candidates alongside cognition
|
|
`DiscoveryCandidate` records.
|
|
|
|
**Why deferred.** ADR-0167 §"Open questions Q4" explicitly out-of-scope.
|
|
The LexicalClaim slice ships the ratification handler but no UI to
|
|
trigger ratification through — today, an operator would call
|
|
`apply_lexical_claim()` from a Python REPL.
|
|
|
|
**Where breadcrumbs live.**
|
|
- ADR-0160 (Core Workbench v1)
|
|
- ADR-0162 (Workbench Design System)
|
|
- W-029 (proposal queue) — closest existing surface
|
|
- W-031 (replay theater) — replay primitive that math evidence
|
|
records inherit through ADR-0057
|
|
|
|
**Acceptance.** A workbench panel that lists pending
|
|
`MathReaderRefusalEvidence` candidates with sub-type, claim signature,
|
|
recognized terms, refusal context, and a ratify-action that calls
|
|
`apply_lexical_claim()` (or its sub-type-specific successor) with an
|
|
operator-supplied reviewer tag and category.
|
|
|
|
---
|
|
|
|
## 5. Cross-domain partition risks (from Gemini's W2-C audit)
|
|
|
|
Two specific code paths Gemini flagged in
|
|
`docs/handoff/ADR-0167-W2C-cross-domain-audit.md` as needing partition
|
|
discrimination:
|
|
|
|
### 5a. Contemplation pack indexing
|
|
|
|
**Scope.** `teaching/contemplation.py::contemplate()` uses hardcoded
|
|
cognition pack and corpus indexes (`_pack_index` and `_corpus_index`).
|
|
Future math-domain candidates would silently get cognition-domain
|
|
lookups.
|
|
|
|
**Acceptance.** Pack and corpus indexes parameterised by
|
|
`candidate.domain` — cognition candidates look up cognition packs,
|
|
math candidates look up math packs (currently `en_core_math_v1`).
|
|
Tests must exercise both paths.
|
|
|
|
### 5b. Replay gate default
|
|
|
|
**Scope.** `teaching/proposals.py` defaults its replay gate to
|
|
cognition's. Proposing math/admissibility candidates requires passing
|
|
`run_admissibility_replay_gate` explicitly to prevent false rejections.
|
|
|
|
**Acceptance.** Either the replay gate is selected by
|
|
`proposal.domain`, or the cognition default is made explicit and math
|
|
proposals are required to declare their gate. Decision goes in
|
|
ADR-0168 (or wherever the first non-lexical sub-type ADR lands —
|
|
that handler will be the first real exerciser of the proposals path
|
|
for math).
|
|
|
|
---
|
|
|
|
## 6. HolonomyAlignmentCase — structural-vs-blend convergence isolation
|
|
|
|
**Scope.** Determine whether the existing
|
|
`tests/test_alignment_graph.py::test_holonomy_alignment_case_positive_closer_than_negative`
|
|
proves *structurally-derived* cross-language convergence or only proves
|
|
*endpoint similarity under the mount-time blend*.
|
|
|
|
**Why deferred.** The proof obligation is executed today — the test
|
|
asserts that an aligned Logos clause produces nearer holonomies across
|
|
English/Hebrew/Greek than a misaligned negative triple. That clears the
|
|
schema's nominal claim. But the test does not distinguish two possible
|
|
explanations for the convergence:
|
|
|
|
1. **Structural.** The Hebrew tri-consonantal root rotors and Greek
|
|
case-last orientation rotations produce versors that genuinely
|
|
land in the same regions of the manifold because the morphology
|
|
operators encode equivalent semantic structure.
|
|
2. **Blend-induced.** `_apply_mounted_primary_domain_resonance`
|
|
(`language_packs/compiler.py:558`) nudges Hebrew/Greek versors
|
|
toward an English prototype at 40% blend, and the test passes
|
|
because both packs have been pulled close to the English anchor
|
|
regardless of structural derivation.
|
|
|
|
If (2) is doing the work, the three-language architecture is a *claim*
|
|
that English-anchored geometric averaging produces the right endpoints,
|
|
not a *proof* that the depth packs are structurally independent
|
|
operators converging coherently with the articulation surface.
|
|
|
|
**Where breadcrumbs live.**
|
|
- `language_packs/compiler.py::_apply_mounted_primary_domain_resonance`
|
|
— the architectural-invariant comment names this gap explicitly and
|
|
references this section
|
|
- `tests/test_alignment_graph.py:73` — the existing positive-closer-
|
|
than-negative assertion
|
|
- `language_packs/schema.py::HolonomyAlignmentCase` — the schema type
|
|
whose nominal contract is "proves structural divergence with
|
|
coherent convergence"
|
|
|
|
**Acceptance.** One of:
|
|
- **(a) Ablation test.** A test that runs the holonomy proof with
|
|
`_apply_mounted_primary_domain_resonance` disabled (or with the
|
|
blend factor set to 0.0) and asserts that the positive-closer-than-
|
|
negative relation still holds. This would prove (1) and retire the
|
|
concern.
|
|
- **(b) Reframe the claim.** If the ablation fails, document
|
|
explicitly that cross-language convergence depends on the
|
|
mount-time blend, and update `HolonomyAlignmentCase`'s contract to
|
|
reflect what it actually proves (endpoint similarity under blend,
|
|
not structural-derivation equivalence). Honest documentation of a
|
|
weaker property beats a stronger claim that the test can't support.
|
|
|
|
**Priority.** Low-urgency, high-information. Not blocking any current
|
|
capability gate. Worth picking up whenever someone next touches the
|
|
language-pack architecture — the comment at the convergence-decision
|
|
site is the trip-wire.
|
|
|
|
Per CLAUDE.md §"Schema-Defined Proof Obligations" — this is the
|
|
prototypical example of a schema-defined obligation that is executed
|
|
but where the test may not meaningfully fail under the violation it is
|
|
written to catch.
|
|
|
|
---
|
|
|
|
## Sequencing recommendation
|
|
|
|
For the operator picking this up next:
|
|
|
|
1. **First**, decide which of the four frame-opener sub-types (§1) the
|
|
next capability gate actually demands. ADR-0166 still gates this —
|
|
the three-question test must pass for whichever sub-type is chosen.
|
|
2. **Next**, parameterise contemplation pack indexing and replay-gate
|
|
selection by `candidate.domain` before the second sub-type lands.
|
|
3. **Then**, begin ADR-0168 (likely FrameClaim-first) with explicit
|
|
wrong=0 hazard pins carried forward from case 0050.
|
|
4. **Finally**, workbench rendering can follow once a second sub-type
|
|
actually exists.
|
|
|
|
No timelines. Order is by leverage, not calendar.
|