feat(corpus): mark the《古蘭經》thematic cue interpretation-only, never [Text] by F-e-u-e-r · Pull Request #51 · F-e-u-e-r/religion-council

F-e-u-e-r · 2026-07-06T19:53:37Z

What

Follow-up to the Islam ADR 0008 Phase 2 work (#50). The sixth《古蘭經》record — a cross-locus thematic paraphrase (多處(如 2:25、103 章); 信道而行善者必得回報之意), not a 馬堅 verbatim excerpt — is now curated interpretation_only: true so it can never be presented as an authoritative Qur'an quotation. It routes to Surface B [Interpretation]; any [Text] claim citing it fails closed at the renderer.

This is the safety line (Option 1). Splitting the theme into concrete verse records is a separate later corpus-bump PR (Option 2, see below) — the two are sequential, not alternatives.

Mechanism

A new carried-not-trusted interpretation_only flag threads through the evidence pipeline:

presentation.json sidecar (interpretation_only: true)
  → retrieve.py PRESENTATION_FIELD_TYPES   (whitelist; type-checked only, both copies)
  → EvidenceSeed → CatalogSeed
  → render_finalizer._build_authority_unit  ← guard (the only ADR-0004 authority mint point)
       seed.interpretation_only → (None, trace-interpretation-only)
       → finalize() fails atomically; no Surface A

New trace-interpretation-only renderer-bypass reason (ADR 0004 §5, fail-closed). Cited correctly as [Interpretation] it flows to Surface B; cited as [Text] it is refused atomically.

Acceptance criteria

✅ Cannot be treated as an authoritative Qur'an quote — structural renderer guard, fail-closed.
✅ Clearly marked interpretation-only / thematic paraphrase — interpretation_only: true + sidecar _note + tests.
✅ Not published-translation (no exact wording/edition supplied) and not generated-rendering (not a rendering of specific verses) — no representation_kind.
✅ No edition-backed assurance; no BM25 default flip.
✅ No retrieval behavior change — see below.

Two deliberate design choices

Reference corpus file left untouched. occurrence_id is minted from source_file + source_line (ADR 0005) and source_line is a ranking tie-breaker, so editing the bullet would shift occurrence-ids / ranking. The classification therefore lives entirely in the sidecar + code → retrieval ranking, occurrence-identity, and committed reports are unchanged (4 test_committed_*_reproducible pass).
Minimal sidecar entry (interpretation_only only; no rights/representation_kind). Reports count only representation_metadata_records / rights_metadata_records, so a bare classification flag moves no counted field — and a thematic cue is not a quotable excerpt, so this is also the honest shape.

Verification

pytest: 405 passed, 1 skipped (pre-existing conditional skip); 4 new tests (1 curation + 3 renderer-guard).
ruff: clean.
Both presentation.json and both retrieve.py copies byte-identical.

Relationship / follow-up

Independent of feat(corpus): tag Quran records with corpus family #50 (corpus_family) — both branch off main; trivial CHANGELOG/test conflicts on merge (same pattern as feat(corpus): ADR 0008 Phase 1 backfill — Hinduism canon-scope metadata #48/feat(corpus): ADR 0008 Phase 1 backfill — Christianity translation/canon metadata #49).
Option 2 (deferred, separate corpus-bump PR): split the theme into concrete 2:25 / 103:1-3 evidence records (exact 馬堅 wording + edition/provenance + rights + representation metadata), remove the 多處 cue, regenerate reports, and list the queries needing new relevance judgments. It requires the supplied verbatim text and accepts benchmark/report/judgment changes.

🤖 Generated with Claude Code

…ext] Follow-up to ADR 0008 Phase 2 (Islam). The sixth《古蘭經》record — a cross-locus thematic paraphrase (多處/如 2:25、103 章; 信道而行善者必得回報之意), not a 馬堅 verbatim excerpt — is curated interpretation_only via the A1 presentation.json sidecar. A new carried-not-trusted interpretation_only flag (whitelisted + type-checked only in the portable retriever) threads through the evidence seed -> catalog to the finalizer, which refuses to mint a Surface-A [Text] authority unit from such a seed (new trace-interpretation-only bypass; fail-closed, ADR 0004 §5). The cue routes to Surface B and can never be presented as an authoritative Qur'an quotation. Not published-translation (no exact published wording/edition) and not generated-rendering (not a rendering of specific verses); no new evidence record, no representation_kind, no edition-backed assurance, no BM25 default flip. The reference corpus file is left untouched on purpose: occurrence_id is minted from source_file+source_line (ADR 0005) and source_line is a ranking tie-breaker, so editing the bullet would shift ids/ranking. Classification lives entirely in the sidecar + code, so retrieval ranking / occurrence-identity / committed reports are unchanged (4 reproducibility tests pass). Splitting into concrete 2:25 / Sūrah 103 evidence records is deferred (option 2). Both presentation.json copies stay byte-identical; 405 passed, 1 skipped, ruff clean. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…G + test adjacency with #51) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Cross-model review (gpt-5.6-sol ultra) finding, owner-adjudicated: ADR 0008 defines Phase 1 = metadata-only/no new records and Phase 2 = new witness records + tracked corpus bump, so calling these metadata-only passes 'Phase 2' contradicted the ADR. Renames the #48 entry, the already-merged #51 entry in the Unreleased section, and the test comment. PR titles updated to match. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…ta (#48) * feat(corpus): ADR 0008 Phase 2 -- Hinduism canon-scope metadata Conservative canon-scope backfill for the three existing Hinduism generated- rendering records -- no new records, no textual-witness / edition claim. - 廣林奧義書 1.4.10 -> canon_scope: sruti, corpus_family: upanishads - 薄伽梵歌 2:48 / 4:7 -> canon_scope: smriti, corpus_family: bhagavad_gita - Existing representation_kind (generated-rendering) / rendering_mode (meaning-rendering) / provenance / rights kept intact. - test_corpus_curation: pins the Hinduism canon_scope/corpus_family (enum membership from policies/corpus-metadata.v1.json); no witness/edition claim. Metadata only: representation_kind/rights were already present, so no report metadata count changes and no report regeneration -- rankings/metrics unchanged. Both presentation.json copies byte-identical. 402 passed, ruff clean. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VHay815trwtRFTwkdCWpbM * test(corpus): pin raw sidecars + allowlist against span_assurance_tier smuggling Cross-model review (gpt-5.6-sol ultra) finding: the parsed-record assertNotIn checks are vacuous for span_assurance_tier because the portable retriever's PRESENTATION_FIELD_TYPES allowlist (deliberately) omits the field and silently drops it at parse time. Pin both layers with a raw-JSON guard over both sidecar copies. Negative-control verified: a smuggled tier fails the test. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * docs(corpus): rename the metadata backfill to ADR 0008 Phase 1 backfill Cross-model review (gpt-5.6-sol ultra) finding, owner-adjudicated: ADR 0008 defines Phase 1 = metadata-only/no new records and Phase 2 = new witness records + tracked corpus bump, so calling these metadata-only passes 'Phase 2' contradicted the ADR. Renames the #48 entry, the already-merged #51 entry in the Unreleased section, and the test comment. PR titles updated to match. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Feuer <feuer@shsl.world> Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

…cency with #51/#48; adopt Phase 1 backfill naming) Conflict resolution keeps both sides (main first, PR second) in CHANGELOG.md and tests/test_corpus_curation.py, and renames this PR's entry + test comment per the owner-adjudicated ADR 0008 phase-label finding. Verified on the merged tree: 442 passed / 1 skipped, ruff clean, fixtures OK, sidecars byte-identical, and all 8 committed benchmark reports regenerate byte-identically (JSON) / identically sans machine-specific timing lines (MD). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…fix stale prose per review) Keeps both sides (main first) in CHANGELOG.md and tests/test_corpus_curation.py. Review-driven fixes folded in per the round-A dispositions: Phase 1 backfill naming; the CHANGELOG count wording now states 12 lexical / 13 bm25 (set by the Christianity backfill; corpus_family is not report-counted); the stale 'no sidecar disclosure' prose about the 6th Qur'an record now points at the #51 interpretation_only decision, and the test additionally pins that the untagged record IS that cue (assertIs interpretation_only True). Verified on the merged tree: 443 passed / 1 skipped, ruff clean, fixtures OK, sidecars byte-identical, baseline + bm25 reports regenerate byte-identically. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The Unreleased/Added section skipped from #51 straight to the corpus Phase 1 backfill entries; the security (#52/#53), safety-routing (#54/#55 ADR 0009, #56 ADR 0010), and observability (#57) work had no entries. Backfills all six in merge order, style-matched to the surrounding entries, content distilled from the merged PRs. Co-authored-by: Feuer <feuer@shsl.world> Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

F-e-u-e-r merged commit 3d85dea into main Jul 6, 2026
4 checks passed

F-e-u-e-r deleted the corpus/phase2-islam-thematic-cue branch July 6, 2026 19:58

F-e-u-e-r pushed a commit that referenced this pull request Jul 15, 2026

Merge origin/main into corpus/phase2-hinduism-canon (resolve CHANGELO…

a38354d

…G + test adjacency with #51) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

F-e-u-e-r mentioned this pull request Jul 16, 2026

docs(changelog): backfill the missing #52–#57 entries #58

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat(corpus): mark the《古蘭經》thematic cue interpretation-only, never [Text]#51

feat(corpus): mark the《古蘭經》thematic cue interpretation-only, never [Text]#51
F-e-u-e-r merged 1 commit into
mainfrom
corpus/phase2-islam-thematic-cue

F-e-u-e-r commented Jul 6, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

F-e-u-e-r commented Jul 6, 2026

What

Mechanism

Acceptance criteria

Two deliberate design choices

Verification

Relationship / follow-up

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant