core/docs/analysis/pr87-postmerge-integrity-audit-2026-07-19.md

PR #87 Post-Merge Integrity Audit

Date: 2026-07-19
Audit branch: chore/postmerge-pr87-integrity-audit
Worktree: ../core-postmerge-pr87-integrity-audit
Merged subject: Forgejo PR #87 — feat(eval): deterministic relational operator ablation + HE/GRC Logos roadmap
Merge commit: 5fae5a67f27fdc5c58b7e11fd71a18af99403a56
Merge-base (pre-PR main tip): d8d62b8e
PR tip commit (last feature commit): 08881ab6
Authority: code, fixtures, git diff, call-chain traces, and local reproduction — not prior dossier trust.

Scope discipline: Audit-only. No production-code edits. No redesign. No HE/GRC implementation.

---

Verdict

VERIFIED: NO CORRECTIVE PR NEEDED

PR #87 adds an isolated eval/ablation surface, sealed fixtures, tests, and planning docs. It does not alter serving, refusal gates, pack validation, canonical fraction-decrease organ code, geometric closure thresholds, feature flags, or existing benchmark score logic. Residual A2k scale>1 admission is pre-existing, documented, isolated outside the n=8 sealed set, and not masked by this merge.

---

Seven-item matrix

#	Claim	Verdict	Exact evidence (path / symbol / line)	Command / test	Risk	Required action
1	Canonical semantics — ablation invokes production fraction-decrease / A2k path for operator; baseline is intentional non-geometric contrast, not a second serving organ	PASS (with noted harness duplication)	Operator: generate/relational_operator_ablation.py run_operator → resolve_promotable_fraction_decrease (L361, import L40–45). Production organ: generate/derivation/fraction_decrease.py resolve_promotable_fraction_decrease L221–225 → compose_fraction_decrease L197–218 → build_fraction_decrease L103–138 (CGA versor_apply). Baseline: run_baseline L232–331 uses build_problem_frame + assess_fraction_decrease + shared _asks_decrease_delta / _has_hazard_surface (imports, not reimplemented) + harness-local _extract_base_and_scale L195–229 and pure rational base * (1 - scale) L307–308. No re-parse of “decrease to N/M of” prose in harness.	PYTHONPATH=. python3 -c "from generate.derivation.fraction_decrease import resolve_promotable_fraction_decrease; …"; tests/test_relational_operator_ablation.py::test_operator_commits_geometric_binding_with_versor_integrity	Low. Harness baseline duplicates role extraction + rational formula + scale∈(0,1) check for scientific contrast. Divergence risk if frame roles and organ base-picking disagree on multi-base prose; sealed gold has single base. Private _asks_* / _has_hazard_* import couples baseline refuse gates to organ internals (intentional shared gate).	None for merge integrity. Future organ work should keep baseline as non-serving contrast only.
2	Fixture independence — n=8 expected labels are static, hand-justified, not solver-generated	PASS	Static oracle: evals/relational_operator_ablation/v1/cases.jsonl (8 lines). Gold: roa-v1-0001 84×(1−3/4)=21; roa-v1-0002 60×(1−2/3)=20 (notes field states formula). Adversarial 0003–0008: expected_answer: null, expected_outcome: "refused" with per-case hazard notes (final-value; affine confuser; percent partition; distractor qty; decrease-by language; multi-fraction). Runner loads JSONL only (runner.py _load_cases L45–63); does not write cases. Committed report.json is measurement artifact, not expectation source.	Independent hand math + live organ: all 8 organ outcomes match fixture labels without reading expected_answer into the solver.	Low. Tests assert against same static labels (normal sealed-eval pattern). Not a circular oracle: gold numbers are grade-school arithmetic from problem text; refuse labels match organ refuse, independently re-checked via resolve_promotable_fraction_decrease is None.	None.
3	Honest condition labels — baseline/operator equality is agreement on this lane, not improvement; depth inactive for EN-only; no global HE/GRC negation	PASS	Condition semantics in module docstring L16–25 and runners: baseline L232–331; operator L334–402; depth L405–437 (inactive_no_mapping, equals operator answer); metadata_only L440–522 (assess_contracts(..., depth=) decorate then same resolve_promotable_*). Runner findings L155–168: “should agree when both commit”; scientific_note records depth inactive on English-only. Dossier L120: metadata “does not improve arithmetic on this English slice”; HE/GRC roadmap L12–13, L28–29: do not generalize to “Hebrew/Greek never help.” Sealed counts 2/0/6 identical across baseline/operator/depth/metadata_only — equality, not uplift.	Runner stdout counts; test_depth_condition_records_inactive_mapping_and_matches_operator; test_baseline_and_operator_agree_when_both_commit	Low. Misread risk if a reader equates “operator” with “better”; text counters that.	None.
4	Scale > 1 — pre-existing A2k admit; PR neither masks nor normalizes; outside n=8 interpretation	PASS (gap real; isolated)	Obligation path: assess_fraction_decrease L471–473 adds scale_out_of_range when not (0 < scale.value < 1). Geometric binding: _versor_binding_from_scale_value L937–938 admits any finite k > 0 (no upper bound). Fallback: _resolve_fraction_decrease_contract L179–194 prefers obligation-complete runnable, else geometric with bindings. Runtime (this audit): text …decrease to 5/4…40 feet…decrease by? → obligation runnable=False missing includes scale_out_of_range; geom runnable=True binding 5/4; resolve_promotable_fraction_decrease → ≈ −10.0. Baseline refuses (assessment not runnable). Sealed cases.jsonl contains no scale>1 surfaces. Test: test_preexisting_scale_out_of_range_gap_isolated L239–254 documents split; soft allow {refused, wrong} so a future refuse-fix does not break. Diff: fraction_decrease.py byte-identical merge-base↔merge (shasum equal).	Minimal repro (below). Unit test above.	Medium product risk (pre-existing organ), zero PR-#87 introduction risk. Soft test does not hard-pin current wrong-commit forever (by design). n=8 scores unaffected.	Not a PR #87 corrective. Optional later organ PR: tighten geometric scale to match obligation (0,1) — separate scope.
5	Production / gate isolation — PR does not alter serving, gates, packs, organ semantics, closure, flags, benchmark scores	PASS	Diff d8d62b8e..5fae5a67 file set (10 paths only): 3 docs under docs/analysis/; evals/relational_operator_ablation/**; generate/relational_operator_ablation.py (new); tests/test_relational_operator_ablation.py. Zero line diff to generate/derivation/, generate/problem_frame*.py, generate/construction_affordances.py, packs/, chat/, algebra/. relational_operator_ablation not imported by math_candidate_graph, chat/, or core/ serving paths. serving_allowed=True for fraction_decrease remains pre-existing in construction_affordances.py (unchanged).	git diff d8d62b8e..5fae5a67 --name-only; git diff … -- generate/derivation/ … | wc -l → 0; rg relational_operator_ablation generate/math_candidate_graph.py chat/ core/ → none	None for isolation. New generate/ module is production-adjacent placement but eval-only consumers.	None.
6	HE/GRC roadmap authority — documentation only; draft/disconnected honesty; observed-morph rule	PASS	Single new doc: docs/analysis/hebrew-koine-greek-logos-pack-capability-roadmap-2026-07-19.md. No pack JSONL/data/compiler changes in diff. Verdict L18: READY ONLY FOR PREREQUISITE WIRING. Source packs draft/disconnected L22, L49, L68, Phase 0 L122. No-go list L155–167: no English→morph, no bulk pack build, no holonomy crown claim. Next unit L199–203 is future branch name only. Ablation boundary L26–29: EN depth inert; do not reopen fixtures without defect.	Diff name-only + doc read	None if future implementers follow no-go list.	None on this merge.
7	Regression / reproducibility — recorded commands green; report digest stable	PASS	Ablation tests 20 passed; runner passed report_sha256=53df6d17df934f0c44f5d1063b16259f6d9db0b59834b7ca20f895eb5e5719c7 matches committed report.json; dual build_report identical. Pre-existing: sprint8 affine/fraction tests 37 passed; problem_frame_contracts + construction_affordances + proportional_decrease green. Smoke suite run recorded in Reproduction evidence.	See Reproduction evidence	Low. Runner rewrites report.json on disk when executed (measurement overwrite); digest stable.	None.

---

Diff classification

Merge range: d8d62b8e..5fae5a67 (includes merge commit + 5 feature commits).

Path	Classification	Notes
docs/analysis/relational-operator-ablation-cartography-2026-07-19.md	documentation only	Phase 1 ledger
docs/analysis/relational-operator-ablation-dossier-2026-07-19.md	documentation only	Measurement dossier
docs/analysis/hebrew-koine-greek-logos-pack-capability-roadmap-2026-07-19.md	documentation only	Planning; no pack authority
evals/relational_operator_ablation/__init__.py	test/eval harness only	Package marker
evals/relational_operator_ablation/v1/__init__.py	test/eval harness only	Package marker
evals/relational_operator_ablation/v1/cases.jsonl	test/eval harness only	Static sealed fixtures
evals/relational_operator_ablation/v1/runner.py	test/eval harness only	Measurement runner
evals/relational_operator_ablation/v1/report.json	test/eval harness only	Committed measurement artifact
tests/test_relational_operator_ablation.py	test/eval harness only	20 unit/integration/adversarial tests
generate/relational_operator_ablation.py	production-adjacent but behavior-preserving	New module; not on serving path; imports production organ for operator condition; baseline is harness-local rational contrast

Absent from diff (verified): any edit to fraction-decrease organ, problem-frame contracts, construction affordances / serving_allowed, pack manifests/data/compiler, chat runtime, algebra closure, existing GSM8K score fixtures, feature flags.

Not present: production semantic change; gate/config change.

---

Canonical semantics (detail)

Operator path (production)

case.problem
  → resolve_promotable_fraction_decrease(text)     # fraction_decrease.py:221
      → compose_fraction_decrease
          → _resolve_fraction_decrease_contract    # build_problem_frame + assess_contracts / assess_geometric_proposals
          → build_fraction_decrease                # base extract + versor_apply dilation
          → self-verify
  → AblationResult.answer = resolution.answer

This is the live Gate A2k entry also used by generate/math_candidate_graph.py (pre-existing import of resolve_promotable_fraction_decrease). Ablation does not replace that graph; it calls the same function with raw text.

Baseline path (harness contrast)

case.problem
  → build_problem_frame
  → assess_fraction_decrease          # obligation completeness + scale_out_of_range
  → _asks_decrease_delta(question)    # shared private organ helper
  → _has_hazard_surface(...)          # shared private organ helper
  → _extract_base_and_scale(frame)    # HARVEST: unique decrease_to_fraction roles → Fraction values; require 0<scale<1
  → answer = float(base * (1 - scale))  # pure rational; no VersorBinding

Harmless orchestration: condition dispatch, scoring, readback assembly, depth provenance records, metadata decoration check.

Semantic duplication (harness-only, intentional):

Symbol	Lines	What is duplicated	Production counterpart
_extract_base_and_scale	195–229	Role→value map for base/scale; scale∈(0,1)	assess_fraction_decrease roles + L471–473; organ _current_base_quantity (different base heuristic)
rational formula	307–308	base*(1-scale)	Geometric dilation then subtract in build_fraction_decrease L120–138 (math-equivalent for k∈(0,1))

Not duplicated: slash-fraction prose regex; multi-fraction hazard logic body (imported); serving dispatch; pack load.

Depth / metadata

• run_depth: answer ≡ run_operator; DepthContribution.validity = "inactive_no_mapping"; empty pack/record ids.
• run_metadata_only: synthetic or override he/grc root labels into assess_contracts(..., depth=); answer still from resolve_promotable_fraction_decrease only.

---

Fixture independence (detail)

case_id	Independent justification	Organ live outcome
roa-v1-0001	84×(1−3/4)=21; train_sample_0005 shape; asks decrease-by	answer≈21
roa-v1-0002	60×(1−2/3)=20; sibling	answer≈20
roa-v1-0003	Final-value Q (“be” not “decrease by”)	refuse
roa-v1-0004	“1/4 more than” affine confuser	refuse (hazard)
roa-v1-0005	Percent partition out of family	refuse (hazard)
roa-v1-0006	Two temperatures (84 + cabin 40); ambiguous base	refuse
roa-v1-0007	“decrease by 3/4 of” not “decrease to … of”	refuse
roa-v1-0008	Two slash-fractions 3/4 and 1/2	refuse (hazard)

No generator script writes cases.jsonl. Expectations are static data fields.

---

Scale > 1 (detail)

Root cause (pre-merge, pre-PR, still present): geometric scale binding allows any k > 0; obligation assessment requires 0 < k < 1; contract resolver falls back to geometric bindings when obligation is not runnable.

Minimal repro (audit machine):

cd ../core-postmerge-pr87-integrity-audit
PYTHONPATH=. python3 - <<'PY'
from generate.derivation.fraction_decrease import resolve_promotable_fraction_decrease
from generate.problem_frame_builder import build_problem_frame
from generate.problem_frame_contracts import assess_fraction_decrease, assess_geometric_proposals
text = (
  "In one hour, the river will decrease to 5/4 of its level. "
  "If the current level is 40 feet, what will the level decrease by?"
)
a = assess_fraction_decrease(build_problem_frame(text))
print(a.runnable, a.missing_bindings)
g = next(x for x in assess_geometric_proposals(build_problem_frame(text)) if x.candidate_organ=="fraction_decrease")
print(g.runnable, [b.semantic_identity for b in g.bindings])
print(resolve_promotable_fraction_decrease(text).answer)
PY

Observed (SHA 5fae5a67): obligation not runnable (scale_out_of_range, …); geom runnable with 5/4; promotable answer ≈ -10.0.

PR #87 effect: none on this path (organ file checksum unchanged). Documents via unit test + dossier remaining-gaps row. Sealed n=8 does not include scale>1 → no change to sealed score interpretation.

---

HE/GRC roadmap authority boundary (detail)

• Diff adds only the roadmap Markdown under docs/analysis/.
• No packs/data/*, packs/he/*, packs/grc/*, packs/el/*, compiler, manifest, or DEPTH_PACK_IDS edits.
• Roadmap states source packs/{he,grc,el} draft/disconnected vs LIVE packs/data/*.
• Explicit rule: only observed, validated morphology may produce a future constraint; English→pseudo-depth forbidden.

---

Reproduction evidence

Field	Value
Commit SHA	5fae5a67f27fdc5c58b7e11fd71a18af99403a56
Branch	chore/postmerge-pr87-integrity-audit @ forgejo/main
Python	3.14.6 (/opt/homebrew/bin/python3)
uv	0.8.22
OS	macOS (host of local-first CI doctrine)

Commands and outcomes

# Ablation unit/integration (PR-recorded)
PYTHONPATH=. python3 -m pytest -q tests/test_relational_operator_ablation.py
→ 20 passed in 0.66s

# Ablation runner (PR-recorded)
PYTHONPATH=. python3 -m evals.relational_operator_ablation.v1.runner
→ passed: true
→ operator/baseline/depth/metadata_only: correct=2 wrong=0 refused=6 n=8
→ invalid (adversarial n=6): correct=0 wrong=0 refused=6
→ report_sha256: 53df6d17df934f0c44f5d1063b16259f6d9db0b59834b7ca20f895eb5e5719c7
→ live build_report digest == committed report.json digest

# Pre-existing fraction-decrease / affine organ tests
PYTHONPATH=. python3 -m pytest -q \
  tests/test_math_candidate_graph_sprint8_r6_affine_lift.py \
  tests/test_math_candidate_graph_sprint8_review_patch.py
→ 37 passed in 1.60s

# Problem-frame / construction / proportional decrease
PYTHONPATH=. python3 -m pytest -q \
  tests/test_problem_frame_contracts.py \
  tests/test_construction_affordances.py \
  tests/test_proportional_decrease_proposal.py
→ all passed (36 tests in this invocation)

# Production path identity across PR
shasum of generate/derivation/fraction_decrease.py at d8d62b8e == 5fae5a67
git diff d8d62b8e..5fae5a67 -- generate/derivation/ generate/problem_frame*.py \
  generate/construction_affordances.py packs/ chat/ algebra/ → empty

# Smoke (PR-recorded merge-bar parity)
uv run core test --suite smoke -q
→ 176 passed in 132.61s (CPython 3.12.13 via uv .venv)

Targeted pre-existing + ablation lanes verify PR #87 surface; smoke confirms no ambient merge-bar breakage at audit HEAD.

---

If corrective work is required

Not required for PR #87 merge integrity.

Optional separate future organ PR (out of scope here; pre-existing gap only):

Field	Recommendation
Branch	fix/a2k-fraction-decrease-scale-range
Worktree	../core-a2k-fraction-decrease-scale-range
Minimal files	generate/problem_frame_contracts.py (_versor_binding_from_scale_value and/or _fraction_decrease_scale_binding); possibly generate/derivation/fraction_decrease.py self-verify; tests under existing sprint8 / new pin
Regression tests	Hard-pin: scale 5/4 text → resolve_promotable_fraction_decrease is None; baseline and operator both refuse; sealed ROA n=8 still 2/0/6
Acceptance	wrong=0 on organ confusers; no serving flag churn; smoke + targeted math tests green

No rollback of PR #87 is indicated.

---

Audit method notes (anti-trust of prior dossier)

1. Diff inventory and empty-diff proof on production paths.
2. Full read of generate/relational_operator_ablation.py, runner, cases, tests.
3. Trace into fraction_decrease.py and problem_frame_contracts.py scale/binding gates.
4. Independent gold arithmetic and live organ refuse matrix for all 8 fixtures.
5. Runtime scale>1 admit reproduction.
6. Roadmap read for authority language; packs absent from diff.
7. Re-run ablation + pre-existing organ tests; digest equality with committed report.

Deliberately not used as authority: prior session narrative; cartography/dossier claims without re-verification (claims re-checked against code).