core/docs/handoff/core-logos-W-Holonomy-brief-2026-06-14.md

W-Holonomy — author real holonomy cases + the Holonomy proof-card tab

Date: 2026-06-14 Status: forward brief (post read-only Studio, L1–L5 complete: LG-1 #737 / LG-2 #738 / LG-3+LG-4 #739 all merged). Dispatch when prioritized. Predecessor doctrine: docs/workbench/core-logos-studio-plan.md (Holonomy tab §) — which the read-only wave deliberately rendered as missing_evidence because no pack-level holonomy.jsonl exists yet. This wave fills that.

Why this is the hard one (read before scoping)

Holonomy is the crown proof of the tri-language design: aligned canonical clauses (Hebrew path, Greek path, English path) must produce nearby holonomies without flattening their distinctions. The HolonomyAlignmentCase schema already exists (language_packs/schema.py) and test-level proof cases exist (tests/test_alignment_graph.py), but no logos pack carries holonomy.jsonl and there is no serving-path holonomy proof evaluator.

Two hard constraints from CLAUDE.md bind this wave:

1. No success state without proof. A Holonomy proof card may render a "holds" verdict ONLY when a real proof was computed. Absent/failed proof renders as missing_evidence / failed, never green. (This is why the read-only wave showed count-0, not a fake card.)
2. Schema-Defined Proof Obligations. HolonomyAlignmentCase only becomes a verified property when an executing test meaningfully fails under a violated case — i.e. a case whose paths flatten a distinction, or whose holonomy exceeds tolerance, must make the evaluator return failed and a test must assert that. A proof evaluator that passes everything is decoration.

This is genuine field-algebra work, not data entry. Scope it as a multi-PR wave.

Sequenced PRs

WH-1 (engine) — deterministic holonomy proof evaluator

• New evaluator (e.g. alignment/holonomy.py) that, given a HolonomyAlignmentCase (source_refs across pack_ids, expected_relation, negative_source_refs, tolerance), composes the field-path holonomy over the aligned canonical clauses using the owned algebra (algebra / versor_apply / cga_inner) — never a re-implemented metric — and returns a typed verdict: holds / flattened / exceeds_tolerance / missing_refs.
• The negative_source_refs are the falsifier: a real case must distinguish the positive alignment from the negatives (no flattening). Non-vacuous tests: a known-good case returns holds; a case with a negative ref swapped in returns a non-holds verdict; a missing ref returns missing_refs. Seed the fixtures from the existing tests/test_alignment_graph.py proof shape.
• Pure/deterministic; respects versor_condition < 1e-6 by construction (no hot-path repair). Off the serving path (no generate.derivation import).

WH-2 (pack content) — author holonomy.jsonl

• Author real HolonomyAlignmentCase rows into the logos packs (start with he_logos_micro_v1 + grc_logos_micro_v1, the curated micro packs), each with ≥2 source_refs across ≥2 pack_ids, an expected_relation, negative_source_refs, and a tolerance.
• Add a holonomy_checksum to each manifest pinned to the bytes written (sha256(holonomy_path.read_bytes())); compiler verifies it like the glosses dual-checksum. Deterministic row order.
• Depends on WH-1 so authored cases are validated against a real evaluator (don't author cases no evaluator can check — that recreates the decoration risk). Reconcile collapse anchors first if a case touches them (logos-collapse-anchor-reconciliation-2026-06-14.md).

WH-3 (reader) — surface cases + proof status

• Extend workbench/logos.py: populate LogosPackContents.holonomy_cases (currently []) from holonomy.jsonl, and add a LogosHolonomyCase read model carrying the case fields plus the WH-1 proof verdict (computed read-only via the evaluator, or read from a persisted proof if WH-1 persists one — prefer persist-first if the computation is non-trivial, mirroring C3).
• Overview holonomy_case_count becomes real; Safety missing_holonomy_refs resolves from unknown once cases exist. Schema mirror + drift gate; the read model SHA-pinned if it asserts a proof metric.

WH-4 (UI) — the Holonomy tab (proof cards)

• Add a Holonomy tab to /logos (between Alignment and Safety). Render each case as a proof card: the three paths (Hebrew ↘ / Greek ↗ / English), source_refs, expected_relation, tolerance, and the proof verdict.
• Honesty gate (load-bearing): render a "holds" affordance ONLY for holds; render flattened / exceeds_tolerance / missing_refs as their honest failed/absent state with no success styling. A test must assert that a non-holds case renders no success element.
• New logos_holonomy_case evidence subject (logos_holonomy_case:<packId>/<caseId>) + inspector projection + chain-rail derivation (the replay stage = the proof verdict). Mirror the LG-3/LG-4 subject pattern.
• Full pnpm exec vitest run (pass AND exit) + golden-file layout test if the proof card uses the DAG primitive for the path diagram.

Non-negotiables

• No success without computed proof; the evaluator must meaningfully fail.
• Engine math owned by algebra — never re-implemented in reader or UI.
• Read-only; no mutation. Off the serving path.
• Deterministic; versor_condition preserved by construction (no hot-path repair).

Out of scope

• Patch Forge / proposal envelope (W-Forge).
• Cross-turn or session-level field-coherence trends.