`intent-fidelity-reviewer` Gains Its Cross-Doc And Kind-Classification Roles

Re-scoped on 2026-09-21 by the product thinker: the consistency pass only. The shape role goes to the kinds lint (itd-34) and the overlap question to the pre-pass (itd-42); the headless oracle leg this record named no longer exists. The press release and scope below are read through this paragraph and the Decisions section.

Press Release

abcd's intent-fidelity-reviewer agent grows from one role to three. Role 1 (per-intent fidelity) grades a shipped intent's acceptance criteria against delivered reality and writes MET/MET_WITH_CONCERNS/NOT_MET/INCONCLUSIVE verdicts back into the intent's ## Audit Notes. This intent ships Role 2 (cross-document consistency — surfaces terminology drift, premise contradictions, scope leakage, sequencing impossibilities, naming conflicts across the brief + intents corpus) and Role 3 (kind classification — examines whether intents' declared kind still fits the corpus and surfaces suggested reclassifications). With all three roles live, /abcd:intent consistency and /abcd:intent shape move from documented command surface to working command surface. The reviewer becomes the corpus's continuous fidelity auditor.

"The corpus had drifted in three places I hadn't noticed," said Nia, facilitator. "itd-29 still used epic throughout its press release copy; itd-34's discipline lifecycle said one thing and adr-9's brief amendment said another; itd-37 had been silently bundle-shaped for two intents now without anyone reclassifying. /abcd:intent consistency and /abcd:intent shape surfaced all three in one pass. None of them were urgent. All of them would have bitten."

Why This Matters

The intent-fidelity-reviewer agent is named in docs/reference/commands.md as the engine behind three distinct review verbs on /abcd:intent:

A 2026-05-16 working-log entry named this gap: "No flow-next spec owns internal/core/lint or the intent-fidelity-reviewer agent" — partially addressed by spc-12 (predecessor store) for Role 1, but Roles 2 and 3 still have no owner. spc-12 (predecessor store)'s spec explicitly bounded itself to Role 1; the other two roles "are NOT in scope: they are a later Pass A/B/C agent spec" (spc-12 (predecessor store) ## Overview).

This intent ships those later epics as a standalone intent whose scope covers both roles — the two roles share substrate that should land together:

• Same agent file (agents/intent-fidelity-reviewer.md). Roles 2 and 3 add prompt sections to the same agent, with the same prompt_version family, the same injection-canary fixture, the same itd-5 discipline gates. Shipping them separately means two agent CHANGELOG bumps two weeks apart and twice the discipline overhead.
• Same oracle infrastructure. Both roles call _build_cli_oracle() and follow the same itd-6 cascade ordering. Both ride on itd-47's autonomous-mode oracle fix.
• Same command-surface conventions. Both follow the [<itd-N>] optional-ID pattern (bare = scan corpus, with ID = scan one intent); both write findings to a structured output the facilitator can act on; both honour the bare-command-as-help convention.

The kind is standalone per itd-34: a formal bundle-member requires multiple intent files sharing a bundle: id and an spec_id, and itd-48 is one intent file. The "bundle" framing here is metaphorical — two reviewer roles bundled into one spec's scope, not the schema bundle-member kind. The capture-time suggested_kind: bundle-member advisory was a category error and is preserved as historical advisory data in the frontmatter; reclassification_history records the null → standalone correction.

The intent is project-agnostic: every abcd project that uses the intent corpus benefits from cross-doc consistency checking and kind reclassification suggestions. The reviewer's behaviour is shaped by the framework, not by any one project's subject matter.

What's In Scope

Role 2 — cross-document consistency

• Extend agents/intent-fidelity-reviewer.md with a Role 2 prompt section that takes the brief + intents corpus as input and surfaces:
  • Terminology drift — uses of forbidden synonyms, glossary-undeclared terms, or terms used inconsistently across documents.
  • Premise contradictions — two intents declaring incompatible facts/assumptions about the same surface.
  • Scope leakage — an intent's ## What's In Scope overlapping another intent's scope in a way that creates double-coverage or contradictory coverage.
  • Sequencing impossibilities — intent A claims to depend on intent B, but B's scope cannot satisfy A's dependency as written.
  • Naming conflicts — same noun used for different concepts; different nouns used for the same concept.
• Implement /abcd:intent consistency [<itd-N>] routing in commands/intent.md (the row already exists in the doc but the routing is stubbed).
• Findings emitted as structured records grouped by judgement category (terminology drift, premise contradictions, scope leakage, sequencing impossibilities, naming conflicts) — spc-29 (predecessor store) ships finding categories, not mechanical lint codes. Mechanical cross-doc categories (schema/state contradictions, reference rot, acknowledgement gaps) and any associated lint-code namespace are deferred to a follow-up intent.
• Reports land at .abcd/logbook/audit/consistency-<ts>/report.{json,md} (JSON internal, MD render — per the transparency invariant).

Role 3 — kind classification review

• Extend the same agent with a Role 3 prompt section that examines whether intents' declared kind still fits the corpus and surfaces suggested reclassifications. Inputs: an intent (or all intents) + the corpus the intent sits in. Outputs: on scoped runs, a scoped_verdict.verdict of KIND_OK / KIND_DRIFT / INCONCLUSIVE; on every run, a top-level suggestions[] list whose entries carry suggestion_type ∈ {kind_change, bundle, supersession}. KIND_DRIFT pairs with an embedded scoped_verdict.suggestion that is also reconciled into suggestions[] (matched by finding_signature).
• Implement /abcd:intent shape [<itd-N>] routing as an on-demand verb. Continuous pre-commit shape scanning is deferred to a follow-up intent — spc-29 (predecessor store) ships only the on-demand surface.
• Pairs with /abcd:intent reclassify (already documented) — when Role 3 surfaces a finding, reclassify is the action verb that commits it.
• Reports land at .abcd/logbook/audit/shape-<ts>/report.{json,md}. Findings carry suggestion_type ∈ {kind_change, bundle, supersession} with the matching arm fields (current_kind + suggested_kind, bundle_members, or superseded_by); scoped runs additionally emit the KIND_OK / KIND_DRIFT / INCONCLUSIVE scoped_verdict. spc-29 (predecessor store) does not introduce a mechanical lint-code namespace for this persona.

Shared

• Both roles share spc-12 (predecessor store)'s testing pattern: golden fixtures + at least one injection-canary per role (per itd-5).
• Both roles call _build_cli_oracle() from itd-47's extended version, so they run in headless Ralph mode.
• Pre-commit hook wiring is deferred to a follow-up intent for both roles. spc-29 (predecessor store) ships the on-demand verbs (/abcd:intent consistency and /abcd:intent shape) only; the previous draft's promise of a intent-fidelity-consistency pre-commit hook and Role 3's "runs continuously in pre-commit" surface are explicitly out of scope.

What's Out Of Scope

• Role 1 modifications. spc-12 (predecessor store) shipped Role 1; this intent does not edit it.
• lifeboat-oracle agent. itd-5's named reviewer; out of scope here.
• Auto-fix for findings. Role 2 and Role 3 surface findings; a human (or /abcd:intent reclassify) acts on them. No "auto-apply-suggested-reclassification" verb.
• Cross-document fidelity beyond the corpus. External docs (PRs, external research) are out of scope; only brief/ + intents/ are the inputs.

Scope Conditions

None stated.

Mechanism

We expect a whole-corpus read to find contradictions no single-record review can, because terminology drift, a premise that two records state differently and a scope that leaked are visible only with both ends in view; shown wrong if a pass over the current corpus finds nothing the ledger did not already hold.

Acceptance Criteria

• Given abcd intent consistency runs bare, when it completes, then a dated report exists under the reviews shelf listing each contradiction found across the brief and every intent (terminology, premise, scope, sequencing, naming), with both ends quoted and located, and the commit it read named.
• Given the report's findings, when the pass files them, then each is an issue naming both ends with the report as its evidence, and a finding the ledger already holds is linked to the existing record rather than filed twice.
• Given abcd intent consistency <itd-N>, when it runs, then the pass is scoped to that intent against the corpus and the report says so.
• Given the pass has run, when the tree is inspected, then the brief and the intents are unchanged; the binary assembled the input and validated the return, and the judgement rode the host.

Decisions

Ruled by the product thinker on 2026-09-21, in the interview that gave this intent its spec:

1. The consistency pass only. The shape role is the kinds lint in itd-34; the overlap question is the pre-pass in itd-42.
2. A report, and a capture per finding, deduplicated against the ledger.

Open Questions

None open; the standalone-versus-two question this record carried is moot with one role left.

Routed Deferrals (spc-33, predecessor store)

spc-33 (predecessor store)'s Phase 3→4 cleanup sweep routed its cluster-A and G1 deferrals into this intent (the routed_from frontmatter backlinks). This intent shipped the consistency pass alone and none of the five, so they are tracked by the ledger record iss-2609260926323349, a future-work seed that names them and hands the Role-3 item to the kinds lint in itd-34. The list below is what was routed here, kept as the record of that routing, not as scope this intent delivered:

• spc-33:A1 (predecessor store) — Role-2 mechanical half: schema/state contradictions, reference rot, acknowledgement gaps → internal/core/lint --cross-doc lint codes XD002 / XD006 / XD007 (the mechanical cross-doc categories this intent defers under "Open Questions → Mechanical Role 2 categories").
• spc-33:A2 (predecessor store) — Role-2 pre-commit hook (blocking-vs-advisory policy) — the hook wiring this intent lists as deferred under "Shared → Pre-commit hook wiring is deferred".
• spc-33:A3 (predecessor store) — Role-3 pre-commit scheduling: the mode="pre_commit" seam exists; the hook wrapper + per-commit-cost policy is the deferred follow-up.
• spc-33:A4 (predecessor store) — chunked corpus review. DORMANT — triggered-by bundle_overflow: true. The Role-2 collector / Role-3 classifier set a bundle_overflow manifest flag on overflow and stop; this item activates ONLY when a persisted report actually shows bundle_overflow: true. Routing it here does NOT make it active implementation scope.
• spc-33:G1 (predecessor store) — consistency-report reconciliation key non-convergence: the R1 finding_id recipe is prose-derived and re-hashes differently across re-runs. Fix is structural-key / deterministic-decode / an addressed report state on consistency_report.schema.json — owned by the predecessor store's spc-29 (this intent owns the reviewer + the consistency-report contract).

Related

• spc-12 (predecessor store) (intent-fidelity-reviewer agent — Role 1) — the foundation this intent extends, via the shared agent file.
• itd-47 (autonomous-mode oracle gates) — precondition. Roles 2 and 3 need headless oracle access; itd-47 ships it.
• itd-31 (cross-document fidelity reviewer; superseded by this intent on 2026-05-27 — file moved to intents/superseded/itd-31-cross-document-fidelity-reviewer.md, superseded_by: itd-48) — the original intent that named Role 2 as a concept; itd-48 absorbs its substance.
• itd-34 (three intent kinds) — the source of the kind taxonomy that Role 3 audits.
• itd-5 (prompt-quality additions discipline) — applies to the agent edits in this intent (prompt_version, CHANGELOG, injection canaries, capability_scope).
• docs/reference/commands.md — the doc whose consistency and shape rows this intent makes real.
• a dated working-log entry (2026-05-16) — the gap entry that motivated this intent.

Grounds

• pursued: the autonomous run builds forty-eight intents against this corpus, and a contradiction between two of them is a stop condition it cannot resolve; we expect the first whole-corpus pass to find contradictions the ledger does not hold; shown wrong if it finds none

Audit Notes

Fidelity review — receipt rcp-80414ddf96b9 (verifier intent-auditor claude-fable-5-1).

Provenance: intent-auditor@claude-fable-5-1 · rubric_hash sha256:effa65b3e9e88ff29433b443ec2be159522a8b0b71cf1434526514aa61edb13e · prompt_hash sha256:420fede6b8a9ca127ff19d7173a9d26718dc6e8ee0c6d9a33022efa0451feb9c Input attestations: diff:internal/core/intent/consistency.go, internal/core/capture/consistency.go, internal/surface/cli/intent_consistency.go, commands/intent.md (Consistency section) at BASE 4ce401c5 (tree 376356ce669156f45b423eb7858a952ffeb0f9d6); live go run ./cmd/abcd intent consistency --json on that tree on 2026-09-26@-;

Acceptance rollup: MET 4 · MET_WITH_CONCERNS 0 · NOT_MET 0 · INCONCLUSIVE 0

Per-criterion verdicts:

• ac-1 — MET: the ingest renders .abcd/work/reviews/< date>-consistency/00-summary.md with review_of_commit pinned to the commit the emit read and one row per finding with both ends quoted and located; TestConsistencyIngestWritesTheDatedReport asserts the path, the pin and the ends, and TestIntentConsistencyRunsEndToEnd drives it through the CLI evidence: internal/core/intent/consistency.go:679 — "reportRel := ReviewsShelfRelDir + "/" + dirName + "/00-summary.md"" evidence: internal/core/intent/consistency.go:1060 — "fmt.Fprintf(&b, "---\nreview_of_commit: %s\n", rv.Commit)" evidence: internal/core/intent/consistency.go:277 — "func headCommit(repoRoot string) (string, error)" evidence: internal/core/intent/consistency_test.go:276 — "TestConsistencyIngestWritesTheDatedReport" evidence: internal/surface/cli/intent_consistency_test.go:79 — "TestIntentConsistencyRunsEndToEnd"
• ac-2 — MET: capture.IngestConsistency files one issue per finding with the report as its evidence and answers Linked: true when an open record already names either end, and the intent core records the link on the row instead of filing twice; TestConsistencyIngestLinksAFindingTheLedgerHolds pins it evidence: internal/core/capture/consistency.go:22 — "func IngestConsistency(repoRoot string, payload []byte, date string) (intent.ConsistencyIngestResult, error)" evidence: internal/core/capture/consistency.go:37 — "return intent.ConsistencyFiling{IssueID: id, Linked: true}, nil" evidence: internal/core/intent/consistency.go:689 — "if filing.Linked {" evidence: internal/core/intent/consistency_test.go:330 — "TestConsistencyIngestLinksAFindingTheLedgerHolds"
• ac-3 — MET: assembleConsistency scopes to the named intent and the prompt body says so, the rubric requires every finding to have an end in that intent, and the scoped emit and ingest are tested in core and through the CLI evidence: internal/core/intent/consistency.go:315 — "if scope != ConsistencyScopeCorpus {" evidence: internal/core/intent/consistency.go:471 — "- scope: %s (%s against the rest of the corpus)" evidence: internal/core/intent/consistency.go:111 — "scoped run: when the scope is one intent, every finding has at least one end in that intent" evidence: internal/core/intent/consistency_test.go:247 — "TestConsistencyEmitScopesToOneIntent" evidence: internal/core/intent/consistency_test.go:354 — "TestConsistencyIngestScopedRun" evidence: internal/surface/cli/intent_consistency_test.go:122 — "TestIntentConsistencyScopedAndRefusals"
• ac-4 — MET: the emit assembles the corpus and writes only the local-tier request and corpus files (a live emit on the BASE tree changed nothing under git), the ingest validates the returned findings and refuses with nothing written, the tree-unchanged assertion is in the ingest test, and the judgement is a host pass over the request as the intent-audit shape is evidence: internal/core/intent/consistency.go:221 — "fsutil.WriteFileAtomic(filepath.Join(dir, rcp+".request.md"), []byte(doc), 0o644)" evidence: internal/core/intent/consistency_test.go:187 — "TestConsistencyEmitAssemblesTheCorpusAndWritesOnlyTheLocalTier" evidence: internal/core/intent/consistency_test.go:317 — "AC 4: the brief and the intents are unchanged." evidence: internal/core/intent/consistency_test.go:385 — "TestConsistencyIngestRefusesWithNothingWritten" evidence: commands/intent.md:868 — "intent consistency --json # the brief and every intent"

Gap audit:

• honoured:
  • the consistency pass (Role 2) is a working command surface, on the CLI and the plugin page evidence: internal/surface/cli/intent_consistency.go:26 — "Use: "consistency [< itd-N>]"" evidence: commands/intent.md:865 — "## Consistency: where do two records contradict each other?"
  • a report and a capture per finding, deduplicated against the ledger (decision 2) evidence: internal/core/capture/consistency.go:13 — "with the report as its evidence, and a finding an open record already holds"
  • the report pins the commit it read and marks a dirty corpus evidence: internal/core/intent/consistency.go:504 — "- review_of_commit: %s" evidence: internal/core/intent/consistency_test.go:590 — "TestConsistencyADirtyCorpusIsMarkedDirty"
• diverged:
  • the press release promised Role 3 (kind classification) and /abcd:intent shape as a working surface; decision 1 moved the shape role to itd-34's kinds lint, and no shape verb exists on the CLI or the page evidence: commands/intent.md:4 — "audit [< itd-N>] | audit --owed [--max < n>] | audit --issue-drift [--strict] | consistency [< itd-N>]" evidence: .abcd/development/intents/shipped/itd-48-intent-fidelity-reviewer-roles-2-3.md:52 — "The consistency pass only."
  • the reviews shelf holds no report produced by this pass: the verb shipped without a first whole-corpus run, so the Mechanism's conjecture (the pass finds contradictions the ledger does not hold) is unexercised; the two consistency-named reports on the shelf predate the verb and were hand-written evidence: .abcd/work/reviews/2026-07-07-roadmap-consistency/00-summary.md:1 — "# Roadmap-Consistency Review - Consolidated Summary" evidence: internal/core/intent/consistency.go:70 — "const ReviewsShelfRelDir = ".abcd/work/reviews""
• missing: (none)