feat(research): add Zotero classification audit evidence

[seonghobae]

Current source-scope and derived audit repair — 2026-09-06

Exact head 6dff8c2ee42cfeb7bf8688c1f7e95989b61be266 normally preserves prior #11 1dc032598b41a35d52c09d8690c871e07365d7e3 and parent #10 fdf8b8d70c05bcb76c55cb6336c9bf31b5e42ce4. Base remains feat/zotero-golden-set-evaluation; OPEN Draft. Runtime is 935e035, final test delta 23178a9.

RED ebdd852 reproduced findings 3934799129 and 3934994550: forged audit reaches governance, and duplicate parent/child coordinates count as provenance-complete. Source digest still represents source, not counters. The selected alternative is shared deterministic audit recomputation before governance; unique nonblank parent/linked-child keys are required for complete provenance. The earlier reply rejecting source-digest mixing does not remove the need for this derived-audit validation. Unknown original-source authenticity and duplicate proposal authorization are not claimed.

PR10 source partition validation, inventory, pending ancestry and v2 approval binding remain intact. Original prediction-mismatch and independent-approval tests now use coherent attacker-controlled counts and retain their rejection expectations. Changing duplicate candidates together with their count remains a subsequent duplicate-owner binding Gap, not authority granted here.

Exact-head locked workspace tests75/15unfiltered suites including2doctests; runtime strictClippy, warnings-denied rustdoc, release build, format, CI contract and diff checks passed. Independent read-only source review reran17integritytests and found no new regression; this is not GitHub approval. Unchanged pinned coverage gate passed140/140functions1101/1101normalizedregions182/182normalizedbranches. Raw1502/1505lines2332/2343regions177/182branches are below100%.

Fresh native Zotero screenshot plus accessibility state both displayed3719items with list rows and attachment icons. Previous Mac-lock limitation is no longer current. Private bibliographic content/screenshots are not committed; this is only visual observation, not completed reclassification. Actual decisions/approvals/writes remain unchanged. Hosted exact-head checks, independent approval, protected integration, release and later restoration/review/write adoption remain pending. Earlier checkpoints below are historical.

Latest bounded-read integration checkpoint

Exact head 1dc032598b41a35d52c09d8690c871e07365d7e3 normally merges parent 4bb633305b04a1dd4c4ce526806c9469bcb79fd3 while retaining previous child 11d158b105cbd03edc34452358ebb3ff445e388e. Base remains feat/zotero-golden-set-evaluation. The #9 whole-snapshot elapsed-time repair and its RED/GREEN evidence are inherited without reverse-merging later features or discarding predecessor deltas.

This exact head passes Rust 1.98.0 locked workspace tests=60 suites=15, including doctests and excluding filtered subprocess duplicates, strict all-target Clippy, warnings-denied rustdoc, formatting, the existing CI contract and diff checks. Log: /private/tmp/conceptweave-deadline-pr11-20260906.log. Intermediate coverage is not inferred from owner/final-endpoint coverage. No new dependency, actual paper read/decision, Zotero mutation or authority issuer was used.

Draft and protected prerequisites remain. Local tests are not hosted GREEN, independent approval, merged/released source or evidence for another head. Earlier checkpoints below are retained history, not the current head.

Current source-integrity note — 2026-09-05

• Exact head: autoresearch/zotero-reclassification-sep04@082710e8c9a0e37ddae528e305cd75ab3574d0bb.
• Exact base: feat/zotero-golden-set-evaluation@e7d4e59f1b55b5954c5f8436527bc96e7ef2fb13.
• This head inherits PR feat(research): add steward golden-set evaluation #10 root e7d4e59f1b55b5954c5f8436527bc96e7ef2fb13 through ordinary merge ancestry. The source-snapshot digest binds complete captured raw provider JSON and the actual typed classifier inputs; source evidence and derived proposals retain separate identities.
• GoldenSetApproval.proposal_digest is required and binds the complete proposal records used for evaluation. The current proposal digest is checked before the caller-owned governance verifier. Do not backfill old receipts: regenerate evidence and obtain a new approval bound to the reviewed evidence.
• Keep Draft. This note does not claim current exact-head hosted GREEN, independent approval, protected merge, live Zotero mutation, or governed publication. Root, predecessor, and terminal-stack local test evidence is not transferred as per-PR hosted evidence.

Earlier heads, runtime snapshots, campaign counts, and verification statements below are historical notes, not current acceptance evidence.

Historical PR notes — original text retained

Outcome

Adds aggregate classification-audit evidence on top of PR #10: snapshot count, bibliographic count, proposal count, provenance-complete count, abstentions, duplicate candidates, zero-success-path failures, and per-disposition totals. The report stays local and aggregate; it does not promote classifier output or Zotero metadata to governed semantic truth.

Exact stack — 2026-09-05

• base: PR feat(research): add steward golden-set evaluation #10 exact b821b0fd485201810650b6a8aed438f9f34b6ad8;
• exact child head: cb365fbf175ae6588aa6176cf9e99ae4d85fcdca;
• the current non-force successor preserves the classification-audit delta while adopting current feat(research): add steward golden-set evaluation #10/Foundation ancestry;
• Draft/open/mergeable. Hosted exact-head Product/security/review evidence must materialize on cb365fb...; predecessor and local evidence do not transfer as GREEN.

Current review disposition

Two prior current-line findings were checked against the actual contracts rather than accepted mechanically.

The P1 request to include audit_summary in snapshot_digest is not a valid repair for this boundary. ClassificationReport::snapshot_digest and GoldenSetApproval::snapshot_digest are immutable Zotero source-snapshot content identities; TRD §11 states that the digest covers every raw Zotero item in canonical key order. audit_summary is derived aggregate evidence from that same in-memory snapshot and is not used as semantic authority by golden-set evaluation. A future published report/artifact may define its own artifact digest at that publication boundary without redefining source-snapshot identity.

The P2 provenance-completeness finding remains valid. Test-only predecessor b836db910a07cb3c3fc3fb68140d70bcd8d6c6c2 added a parent+linked-note regression requiring a proposal with a blank linked-child Zotero key to be excluded from provenance_complete_count; it did not execute before concurrent source repair. The retained production predicate requires the parent key and every retained child_item_key to be nonblank. The P2 thread remains unresolved until the restacked exact head receives terminal verification; source repair alone is not described as RED→GREEN.

Audit semantics

ClassificationAudit is aggregate-only. A successful report records snapshot/bibliographic/proposal/provenance/abstention/duplicate/disposition totals and zero reported failures; reader errors fail before a successful report is returned. Zotero 9 item version zero remains a valid observed revision coordinate rather than missing provenance. Provenance completeness additionally requires stable identity for every linked child coordinate retained by the proposal.

Inherited PR #10 boundary

PR #10 remains Foundation-dependent. Its current head retains complete observed snapshot identity, raw-snapshot SHA-256 content binding, duplicate-key rejection independent of version, and caller-owned verification of the complete ReviewedGoldenSet. #11 must not become independently merge-ready ahead of #10 or bypass any exact-head finding in its parent.

Zotero boundary

No Zotero mutation exists in this stack. Zotero remains the bibliographic system of record; write-back is a separate Zotero 10+ reviewed capability with authorization, server identity, fresh preconditions, before/after receipts and rollback. Aggregate local counts are evaluation evidence only, never published semantic truth.

Next causal step

Keep Draft. Require exact cb365fb... to execute Product/coverage/rustdoc on one unchanged head. If it fails, repair only the causal failure. Resolve the remaining P2 review thread only after current-head verification. PR #12 and later descendants must retain this current parent ancestry before their evidence is treated as current-stack evidence.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-04T14:43:56.265815Z	c28fcca	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[seonghobae]

Current-head finding: ClassificationAudit::provenance_complete_count treats item_version == 0 as missing provenance. That is not valid for the supported Zotero 9 Local API path. Zotero's primary Local API documentation says earlier versions reported synced versions, with 0 for objects that had never been synced; local edits also did not necessarily change those versions. Version zero is therefore a legitimate provenance coordinate, not evidence absence. The current regression explicitly expects a zero-version item to be incomplete, so the wrong invariant is executable. Add a test-first correction requiring a nonblank key with version 0 to remain provenance-complete, then make the minimum production/test repair; immutable content identity remains the separate PR #10 snapshot-digest lane.

[chatgpt-codex-connector]

💡 Codex Review

ConceptWeave/crates/conceptweave-zotero/src/lib.rs

Line 335 in eda4d7a

if !verify_approval(&golden.approval) {

Bind the approval to the reviewed labels

With a valid receipt and matching snapshot, changing any golden.labels entry still passes verification because verify_approval receives only GoldenSetApproval, and none of that authenticated data commits to the labels. The function then counts altered values as steward-approved truth, allowing golden-set metrics to be corrupted while reported as verified; bind a canonical label digest into the approval or have the verifier authenticate the complete ReviewedGoldenSet.

AGENTS.md reference: AGENTS.md:L18-L18

---

ConceptWeave/crates/conceptweave-zotero/src/lib.rs

Lines 338 to 344 in eda4d7a

	let report_snapshot = report
	.classified_items
	.iter()
	.map(|item| SnapshotItemRevision {
	item_key: item.item_key.clone(),
	item_version: item.item_version,
	})

Bind the approval to the entire classifier input

For snapshots containing child notes, attachments, annotations, or other excluded records, this constructs report_snapshot only from top-level bibliographic classified_items, even though those omitted records are part of the classifier input and can affect linked-child evidence. An approval that correctly lists the complete Zotero item-key/version snapshot therefore fails with SnapshotMismatch; omitting those records makes it pass but leaves part of the source snapshot unauthenticated. Retain and compare the full input item-revision identity rather than the classified subset.

AGENTS.md reference: AGENTS.md:L18-L18

---

ConceptWeave/crates/conceptweave-zotero/src/lib.rs

Lines 352 to 354 in eda4d7a

	if report_snapshot.len() != report.classified_items.len()
	|| approved_snapshot.len() != golden.approval.snapshot_items.len()
	{

Reject duplicate keys even when versions differ

When a report contains the same Zotero key twice with different item versions, both SnapshotItemRevision pairs remain distinct, so this length check passes and a matching approval is accepted. The later BTreeMap keyed only by item_key silently discards one proposal, producing metrics from an ambiguous report instead of failing closed as duplicate report identities should; validate item-key uniqueness independently of version.

AGENTS.md reference: AGENTS.md:L18-L18

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[seonghobae]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: f509ebfa36

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[seonghobae]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 551e6cb1f7

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Include linked child identities in the provenance count

When a bibliographic item has a linked child with a blank key, child_index retains that child in the proposal, but this predicate examines only the parent's key and counts the proposal as provenance-complete. Since the linked child can no longer be identified by a stable source coordinate, the audit overstates provenance completeness for malformed/untrusted snapshots; validate the proposal's linked child revisions as well as its own key before incrementing this count.

AGENTS.md reference: AGENTS.md:L18-L18

Useful? React with 👍 / 👎.

[seonghobae]

Current-stack correction (2026-09-05): exact base #10 bf35f97a1d5d2538e3370cf70a70b4294c1faed6; exact head 472dd348e8f0f07b154af2f63e868104e7519de5. Non-force restack preserves classification-audit semantics; older body SHAs are historical and do not transfer GREEN.