Skip to content

Curation persistence: versioned panels, ballots and reviewed page compositions #196

Closed as not planned

Description

@holden

Delivery child of #193; refines #97's schema proposal. #156 renders the output, #197 produces proposals, #198 refreshes them. #190 is required before bot claim writes.

Implementation status, 2026-09-27: the first persistence slice is merged in PR #206 at c6bd882f550f80c83a87f82eb4b0fedd81eb493c, following CodeRabbit resolution and an independent re-audit (2,269 tests, 19 doctests, five independent regression probes and populated upgrade rehearsal passed). Configuration/profile identities and versions, manual composition records, operator review/publication receipts and read-time eligibility are implemented. Profiles are proposed identities, not admitted dossiers; the default remains manual-only until a real model and roster are available. No model, panel run, ballot, decision authority, page binding or refresh scheduler was shipped by that slice. This issue remains open for those persistence integrations. The schema below is the broader target; the merged docs/curation/persistence-slice-1.md states exactly what exists and which later child owns each extension.

Two different outputs

A composition selects and orders a sourced lead plus up to three highlights across a page's meanings. It does not assert identity, rewrite a definition, or create an illustrates edge merely by displaying a poem. An exemplar nomination asserts that a specific subject illustrates a specific sense/concept; it uses the existing assertions, assertion revisions, evidence and review-context machinery of #105/#181. A composition can feature an accepted claim, but its own approval never accepts that claim.

Remote dictionary main checked at 828cdf2. Existing revision/evidence/actor types must be reused with real FK types, not replaced with a parallel claim graph.

Stable composition identity and curation configuration

editorial_compositions.id is the durable editorial target for one curation configuration. #201 introduces curation_configuration_id FK; stable identity is scoped by target/scope + configuration ID + language, not by configuration version. Seed only one system-owned global-default configuration now; alternate logical configurations exist only in integrity tests until later authorized stages. Add editorial_composition_memberships(composition_id FK, object_id FK objects, role, inserted_at), unique (composition_id, object_id, role), to describe its current page scope; snapshot membership IDs/revisions in each version and run. It is not a URL slug or the first lexeme returned by a query.

For the ten-page pilot, explicitly allocate these identities and memberships from the existing page resolver's canonical registry-object scope. A normalized sorted membership signature, scoped by scope_kind, language and curation_configuration_id, is unique for active lexical-page compositions and supports alias lookup; it is a lookup key, not the permanent identity. Two configurations may therefore address the same lexical scope with independent composition histories. Configuration-version edits keep the same composition ID, changing frozen inputs and creating later versions rather than new identities. Scope changes preserve the composition ID, update the lookup mapping transactionally and invalidate freshness/approval as appropriate. Ambiguous or split scopes require reconciliation, never a guessed alias. Exact single-lexeme pages and aggregated word pages are different scopes when their membership differs. New targets are provisioned by a bounded service, never by deriving a hash from URL spelling.

#194 owns the explicit audited mapping from a durable page to its default composition, with at most one active default editorial binding per page. This does not prohibit future alternate configuration-scoped compositions for the same target. Validate scope/language compatibility; no URL or shared label establishes the binding. This child does not choose namespaces, redesign routes, add domain taxonomy tables or merge unrelated meanings. The authoritative merged ADR 0004 currently says “at most one active binding per page”; its required default/alternate clarification is recorded in #201/#194 and remains a separate documentation change.

The MVP server-side resolver returns only {configuration_id, configuration_version_id, resolution_policy_version, reason: global_default, evidence_refs} for a stable scope. It must not select an alternate configuration from public query parameters. A missing/disabled default returns unavailable rather than an arbitrary enabled row. Future domain resolution is an extension contract only: versioned stable-ID classifications/editorial assignments, deterministic precedence and global fallback under #201; personal ownership/access/sharing is not implemented now.

Tables and constraints

All new rows have primary IDs and timestamps; version content and completed decisions are immutable. Operational status/pointers may change through audited transitions. Use UTC timestamps. Add unique composite keys wherever referenced below so the database can enforce ownership.

Table Required fields and relationships Uniqueness / checks
curation_configurations stable ID, unique slug, name, ownership_kind = system only, role (global_default/internal-test), state, current version FK, created-by actor/timestamps one enabled global default; current version belongs to configuration; seed only global-default; no user ownership tables now
curation_configuration_versions configuration_id FK, version, manifest hash, panel-selection/size/quorum/abstention/round policy, lead/highlight policy, template version, local_model_config_id FK, creator/activation audit unique (configuration_id, version) and (configuration_id, manifest_hash); immutable manifest; explicit validated activation does not publish compositions
curation_configuration_members configuration_version_id FK, profile_id FK, exact profile_version_id FK, slot/priority, voting_weight (MVP fixed integer 1) unique (configuration_version_id, profile_id) and slot; referenced profile version belongs to profile; reusable roster distinct from actual participants
curator_profiles slug, bot_actor_id FK actors, label, state, current_version_id FK curator_profile_versions unique slug and actor; current version belongs to profile; bot source contract respected; actor does not impersonate historical entity; offered subject must be deceased with sourced admission evidence
curator_profile_versions profile_id FK, version, manifest_hash, dossier JSON, source/edition/rights references, template version, created_by_actor_id FK, required change reason unique (profile_id, version) and (profile_id, manifest_hash); reviewed source excerpts only; immutable
local_model_configs weight repo/tag + immutable digest, quantization, runtime/version, chat template, license, context/output/reasoning limits, sampling, config hash unique config hash; immutable; deployment endpoint/credentials stay runtime configuration, not persona fields
curation_runs composition_id FK, configuration_version_id FK, resolution policy/version/reason/evidence snapshot, input fingerprint/manifest hash, scope and sense-revision snapshot, candidate fingerprint, model_config_id FK, frozen policy JSON/hash/version, routing/template versions, status, attempt/lease owner/expiry/fence, budget reservation, started/completed times, outcome execution key unique (composition_id, input_fingerprint, refresh_generation); one queued/running run per configuration-scoped composition; composite ownership checks tie configuration version to composition configuration; retries resume same logical run with fenced attempts
curation_run_participants run_id FK, configuration_member_id FK, profile_version_id FK, profile_id FK, slot, frozen voting weight=1, status, missing/abstention reason unique (run_id, profile_id) and (run_id, slot); profile version must belong to profile and member must belong to run's frozen configuration version; frozen before first generation; future routed subsets remain inside that versioned roster
curation_candidates run_id FK, canonical candidate/display identity keys, content type, frozen relation_basis (direct_target/located_usage/reviewed_interpretation), category-assignment actor/reason/evidence references, item_object_id FK objects, intended sense_revision_id FK (or explicit concept object), exact typed source/content/sense revision FKs, passage locator, evidence/rights fingerprint, eligibility expiry, optional accepted assertion_revision_id FK unique (run_id, candidate_key); candidate key includes item, meaning and passage; endpoint/revision ownership validated; one declared intended meaning; no discovery-result-only identity
curation_ballots run_id FK, participant_id FK, round (1/2), candidate_id FK, support/oppose/abstain, rationale, evidence refs, structured objections, recorded-at unique (run_id, participant_id, round, candidate_id); composite FKs ensure participant and candidate belong to run; immutable accepted ballot; missing response recorded as participant status, never invented support
curation_nominations run_id FK, candidate_id FK, nominating participant, kind (highlight/illustrates), rationale/generated-note label, evidence refs, validation outcome, optional resulting assertion_revision_id FK unique (run_id, candidate_id, kind); run ownership FKs; existing subject/predicate/meaning duplicate handling remains authoritative; other supporters remain ballots
editorial_compositions stable ID, required curation_configuration_id FK, scope kind/language, current membership signature, current_published_version_id FK, last_completed_at, refresh_after, next_retry_at, refresh generation, desired input fingerprint, lock version unique active (scope_kind, scope_signature, language, curation_configuration_id); configuration version excluded from stable key; published pointer must belong to this composition/configuration and pass publication guard; clock/queue state distinct from published content
editorial_composition_versions composition_id FK, parent_version_id FK (same composition), required change reason, configuration_version_id FK, resolution policy/reason/evidence snapshot (also for manual selections), version, originating_run_id FK (nullable for manual), created_by_actor_id FK, arrangement hash, scope/input/policy snapshot, eligibility fingerprint, derived review status unique (composition_id, version) and (composition_id, configuration_version_id, arrangement_hash, eligibility_fingerprint); run and configuration version belong to composition/configuration; eligibility fingerprint includes configuration version; content immutable, review status derived from append-only decisions
editorial_composition_items composition_version_id FK, role lead/highlight, position, durable item_object_id FK, exact typed revision FKs + passage locator, intended meaning FK, optional assertion_revision_id FK, nomination FK, selection_origin (panel_recommendation/human_override/human_added/manual), attribution/rights refs, concise generated reason + author/profile label unique (version, role, position) and selected item/meaning/passage key; one lead, highlight positions 1–3; exact revision belongs to object; source definition for lead; no stored arbitrary polymorphic revision ID
editorial_composition_reviews composition_version_id FK, reviewer_actor_id FK, decision accepted/rejected/withdrawn/needs_review, reason, reviewed eligibility fingerprint and timestamp append-only authorized operator decisions: the optional path for manual compositions, overrides and corrections, and an operator veto; not the routine gate (see item 2 of Review, retention and publication); idempotency key per review action; no bot reviewer

Transparency additions owned here (#203)

#203's history is a read projection of these records, not a separate generic event store. Add the following narrowly scoped records/fields where an equivalent canonical record does not already exist at implementation time:

  • curation_candidate_proposals(run_id FK, candidate_id FK, participant_id FK, proposed_at, concise rationale, typed evidence refs), unique (run_id, candidate_id, participant_id). Same-run composite FKs preserve every co-proposer instead of keeping just the first nominator. Frozen candidates remain bounded; no extra model calls.
  • curation_round_responses(run_id FK, participant_id FK, round, status, occurred_at, safe reason_code), unique (run_id, participant_id, round), same-run FKs. A successful first response and a missing/invalid second response are different states; never fabricate an abstain ballot for missing output. Finalize a round response only after its permitted retries/deadline; accepted response/ballot outcomes are then immutable and late output cannot replace them. Retry-attempt failures remain in the existing attempt ledger.
  • Run result includes selection_algorithm_version, derived final support counts (recomputable, not independent authoritative votes) and a deterministic selection/exclusion trace referencing candidate IDs and safe reason codes. Configuration versions carry equal weights/policy and required activation/change actor/reason; snapshots must match their configured membership. The human-reviewed relation category is a declared editorial input, not measured confidence.
  • editorial_composition_publications(composition_id FK, previous_version_id FK nullable, published_version_id FK nullable, authority_kind operator/panel_decision, authorizing_review_id FK (operator), authorizing_decision_id FK (panel decision, added with runs), actor_id FK, action publish/withdraw, reason, committed_at, idempotency_key). Old/new versions and the authorizing review or decision belong to the same composition/configuration; publish requires a valid authority for that exact version, withdraw removes the pointer. Unique action key and an append-only receipt in the pointer-switch transaction establish when publication/supersession/withdrawal actually happened. A review timestamp or row update time is not a substitute. If implementation already retains an equivalent authoritative receipt, extend/reuse it rather than duplicate it.

Human edit/override creates a new version, parent lineage and required change reason, leaving old ballots and recommendations intact. A new manually selected item has no invented panel vote. History must show assessed/reviewed/published times separately and diff exact lead/highlight/reason/evidence/configuration changes. Public actor identity is an appropriate display identity/pseudonym, never private email/IP metadata.

Profile versions supply permitted biography/influences/editions, rubric/examples/counterexamples, abstention, documented blind spots and reviewed historical context. #203 projects profile changes and selection history with the same configuration/access filtering as page history; #204 provides the broader method explanation. Both preserve historical writings separately from generated commentary.

Privacy and source retention apply to history/diffs/profile feeds/caches/exports as well as current content. Failed or adverse private nominations and restricted passages remain internal; even a tombstone cannot reveal prohibited names/content. Store only lawful audit metadata after redaction. Never retain/expose raw chain-of-thought, prompt dumps, secrets or manufactured debates as “transparency.”

Typed references: reuse real sense_revisions, content_revisions, source_record_revisions and assertion revisions; validate matching item/object ownership with composite FKs or constraint triggers. A reference to a work may use its durable object plus the exact source-record revision and locator, not a invented content revision. Run manifests can contain schema-validated lists of these typed references, but graph relationships must not be opaque JSON IDs. A short quotation/passage can use an existing content revision plus locator; add a segment only if genuinely missing.

The global-default version initially references the five existing profile versions, 4/5 quorum, two rounds, one panel repair, the pinned model/template and Bierce-first selection policy. The current version pointer belongs to the stable configuration and changes only through validated operator activation with audit. Freeze exact version, participants and resolution reason per run; an activated update cannot change running ballots. Configuration/profile/model/policy edits invalidate affected inputs/drafts without automatically publishing or copying approval. Mandatory source-integrity, rights, withdrawal and claim-review constraints live outside overridable editorial preferences. No separate weights or fine-tuning are required to change configuration.

Presentation uniqueness, freshness and approvals are configuration-scoped; existing semantic claim/evidence identity remains shared and rejection cannot be evaded by switching configuration. Future private modes require authorization and visibility-aware cache separation before any exposure. MVP has no personal ownership/sharing controls or domain tables. Keep one global inference budget across every configuration/test route.

Run attempts may be normalized into a small curation_run_attempts ledger (unique run/fence) and request records (unique request id, participant/round/repair, token/time/status) to support runtime accounting. It is operational history, not another voting/review graph.

Review, retention and publication

  1. A human may manually compose eligible material before a model exists (the optional operator path). Panel output becomes publishable only as a finalized decision that reaches the configured consensus and passes the evidence/policy checks; output that falls short creates drafts only. For illustrates, call the existing contribution path and retain assertion evidence/review contexts. Never write an accepted assertion directly. Existing rejection/withdrawal is a block on automated renomination even if the generic contribution API permits a fresh submission; a human must explicitly reopen it.
  2. Publication authority. Routine publication needs no per-selection human sign-off: its authority is the configured panel's finalized decision on the exact composition version (arrangement, generated reasons, sources and intended meanings), meeting the configured consensus (4/5) and passing every evidence/policy check. The decision must be a genuine, immutable run result of the composition's own configuration version: actual roster participants, recorded ballots and per-round statuses (a missing response is never an abstain or a support), a recomputable selection-v1 trace and support counts. It is cited by the receipt (authority_kind = panel_decision, authorizing_decision_id). An operator may instead accept a version (manual compositions, overrides, corrections), which authorizes an operator publication; an operator rejection vetoes publication. A linked nomination still requires its own current accepted claim review; neither a panel decision nor an operator acceptance accepts a semantic claim. Reuse reviewer authorization/UI conventions; the small composition-review table handles operator presentation decisions, not semantic claims.
  3. In one transaction lock the composition, validate the expected current pointer and scope/evidence/configuration-version/policy fingerprints, recheck all rights/claim eligibility, then switch the published pointer. Concurrent publication returns a conflict instead of overwriting a newer version. Only the publication service switches it, and only with a valid authority: a background job may do so solely on a finalized panel decision, never without one. Record withdrawal; rollback selects a still-eligible previously published version only by an explicit operator action.
  4. At read time revalidate withdrawal, rights expiry, current source/meaning revisions and claim review validity. Suppress an ineligible item immediately and enqueue bounded repair; do not silently substitute a different item. A changed lead can leave the band leadless. Do not rewrite an immutable arrangement to disguise suppression. Public dissent is filtered by the same evidence/privacy rules; draft sensitive nominations stay internal.
  5. Retention applies to candidates, prompts, ballots, notes and historical versions, not just discovery rows. Store durable references plus permitted metadata; do not copy transient payloads into permanent audit JSON. On expiry purge restricted bytes and retain a permitted tombstone/hash/reason. FKs to legally deletable source revisions use a documented tombstone/null-on-delete path that marks the item ineligible; audit retention must never prevent mandatory source deletion. No fabricated quote or duplicate downloaded media.
erDiagram
  curation_configurations ||--|{ curation_configuration_versions : versions
  curation_configuration_versions ||--|{ curation_configuration_members : configures
  curator_profile_versions ||--o{ curation_configuration_members : references
  curation_configurations ||--o{ editorial_compositions : separates
  curation_configuration_versions ||--o{ curation_runs : freezes
  curation_configuration_versions ||--o{ editorial_composition_versions : governs
  curation_configuration_members ||--o{ curation_run_participants : resolves
  curator_profiles ||--|{ curator_profile_versions : versions
  curator_profile_versions ||--o{ curation_run_participants : participates
  local_model_configs ||--o{ curation_runs : configures
  editorial_compositions ||--o{ curation_runs : refreshes
  editorial_compositions ||--o{ editorial_composition_memberships : scope
  curation_runs ||--|{ curation_run_participants : freezes
  curation_runs ||--o{ curation_candidates : snapshots
  curation_run_participants ||--o{ curation_ballots : casts
  curation_candidates ||--o{ curation_ballots : receives
  curation_candidates ||--o{ curation_candidate_proposals : proposed
  curation_run_participants ||--o{ curation_candidate_proposals : proposes
  curation_run_participants ||--o{ curation_round_responses : responds
  curation_candidates ||--o{ curation_nominations : proposes
  curation_nominations o|--o| assertion_revisions : optional_claim
  curation_runs o|--o{ editorial_composition_versions : proposes
  editorial_compositions ||--o{ editorial_composition_versions : versions
  editorial_composition_versions ||--o{ editorial_composition_items : orders
  editorial_composition_versions ||--o{ editorial_composition_reviews : reviewed
  editorial_compositions ||--o{ editorial_composition_publications : publishes
  assertion_revisions o|--o{ editorial_composition_items : optional_exemplar
Loading

Acceptance

  • Migrations/constraints prove cross-run ballots, duplicate persona slots, mismatched object revisions, four highlights and cross-composition pointers cannot be stored.
  • Alias URLs resolve to one pilot default composition; different lexical scopes/configurations stay distinct. Scope change/split and Routing before launch: On overviews, standards-informed namespaces, and stable subject URLs #194 mapping are explicit, tested contracts.
  • Manual operator-approved composition renders without inference or claims; panel output publishes only on a finalized consensus decision that passes the checks, and a draft that falls short never publishes; neither an operator acceptance nor a panel decision can accept a pending claim or revive rejected evidence. Missing, forged, stale and cross-configuration authorities are refused.
  • Two approvals race safely; expired evidence, changed senses, withdrawn claims and changed policy fail publication/read eligibility and retain lawful audit history.
  • Model/profile upgrades and retries cannot duplicate claims or mutate historical ballots. mix precommit passes when implemented.

Configuration foundation acceptance (#201; NOW only)

  • One system-default configuration is seeded; its five members are versioned deceased-person-inspired profiles. No domain/personal resolver, UI or ownership platform is introduced.
  • Two test-only configurations for the same scope/language can coexist with independent run/content histories, freshness, decisions and published pointers. Cross-configuration participant/ballot/version/pointer writes and approval overwrite are refused by ownership constraints.
  • New config version preserves composition identity and old frozen runs, invalidates affected draft eligibility and does not inherit approval or publish. Configured roster growth requires no name-specific orchestration.
  • Default-only renderer rejects alternate selection through public input; one global budget covers all configurations. A later access-control design is required before private/alternate modes become exposed.

Transparency acceptance (#203): reconstruct manual/no-panel and panel histories from canonical records; all five weights equal 1; preserve co-proposers and per-round missing statuses; recompute counts/order from frozen inputs; show human edits and actual publication receipts. Duplicate publication delivery is idempotent, cross-configuration parent/review/pointer links fail, and source/rights redaction suppresses every public historical surface. No competing event truth or fabricated missing rationale.

Activity

  1. holden commented on Sep 27, 2026

    @holden
    MemberAuthor

    Progress: persistence slice 1 is up as draft PR #206 (codex/curation-persistence-196, head b8211ed). It needs review and is not merged. It is also not a claim of readiness: tests passing is not the same as reviewed.

    This slice covers the manual path only: provision composition → manual version → human presentation review → explicit publication with a receipt. The invariant map was written before the migration and is in docs/curation/persistence-slice-1.md. It maps K1–K10 and R1–R7 to their constraints and tests.

    Built here

    • Tables:
      • editorial_compositions: identity (scope_kind, scope_signature, language_tag, curation_configuration_id), where the configuration version is not part of the identity;
      • memberships and chained scope-change receipts;
      • immutable versions and items holding exact references;
      • append-only reviews and publication receipts.
    • Bierce first is enforced from the registry when a version is made, reviewed, published and read.
    • Published.current/1 withholds and never substitutes.
    • Source deletion nulls the reference and is never blocked. Items keep no source text.

    Differences from the draft table above. Each is reasoned in the PR.

    Verified on an isolated test database (MIX_TEST_PARTITION=_curation196):

    • migrate → rollback leaves nothing → re-migrate;
    • 46 curation tests, 0 failures;
    • mix precommit: 2135 tests, 0 failures.

    Nothing ran against the shared dev DB.

    Not in this slice:

    Production selection stays disabled.

    Review asks. Are the seven reconciliations the right ones? Is the identity tuple enough for #205's locale and scope checks? Is the eligibility fingerprint the right thing for a reviewer to accept?

  2. holden commented on Sep 27, 2026

    @holden
    MemberAuthor

    Independent audit: changes required before merge

    Audited PR #206 at b8211edee08e32afcad22320171c8eb218ba8f27; the head is unchanged. This is a bounded repair request, not a request to start panel execution or the refresh scheduler.

    Independent verification: 2,135 tests and 16 doctests pass; the 47 curation/seed tests pass; compilation with warnings as errors passes. Five additional regression probes fail. Formatting fails in two committed curation test files. CodeRabbit completed its review with moderate merge risk and two unresolved findings.

    Required repairs

    1. Deleted claim evidence must not restore eligibility. A published highlight pinned to a claim is withheld after that claim is superseded. Deleting the historical claim revision nulls the reference and makes the highlight appear again. Preserve the difference between an item that never depended on a claim and one whose required claim was deleted, while allowing mandatory deletion.
    2. Rejected defining relations must not authorize leads. LeadRule.defining/1 bypasses Claims.visible(:public). A rejected defines assertion still authorizes its content as a published lead. Use the canonical visibility policy for priority and fallback definitions.
    3. Configuration upgrades must permit a fresh version of unchanged content. Activating a new configuration with a changed highlight limit withholds the old selection, but reissuing the same valid arrangement is rejected as a duplicate. Include the frozen configuration version in the version-deduplication rule and matching unique constraint, without changing permanent composition identity.
    4. Exact work evidence must identify that work. An unrelated artifact can be published with the committed Cupid and Psyche catalog pin. Validate catalog/source identity ownership rather than independently testing that a row and an object exist. Cover the source-record path too.
    5. Note attribution belongs to the authenticated author. Caller-supplied author_label can impersonate another writer and survives publication. Confirmed CodeRabbit finding; derive attribution from the acting principal.
    6. Restore the required checks. Fix formatting in compositions_test.exs and publications_test.exs, document PostgreSQL 15+ (column-list ON DELETE SET NULL), and run mix precommit on an isolated test database.

    Publication-policy correction

    The user clarified that routine curation is intended to publish through agent-panel consensus plus evidence/policy checks, without mandatory human approval for every selection. The current issue bodies and implementation reflect an earlier manual-only brief; this is requirements drift, not merely misleading UI text.

    Reconcile #196/#201 and the implementation documentation with that intent. A manual composition path may remain as an explicitly optional operator mechanism; it must not be presented as the universal future publication gate. Document the versioned authorization contract and the migration/integration boundary for genuine panel decisions. Until that path exists and has verifiable decision records, automated publication remains unavailable. Do not remove authorization, manufacture panel records, or implement inference/scheduling simply to complete this repair. If changing publication authority in this PR, require tests for forged, missing, cross-configuration and stale authorizations before calling it ready.

    Merge exit criteria

    • All five independent reproductions become meaningful passing regression tests, including deletion/supersession and end-to-end publication/read outcomes.
    • Existing committed-connection race and ownership checks still pass; test relevant migration changes in an isolated database.
    • mix precommit passes at the exact reviewed head, with evidence recorded.
    • CodeRabbit's findings are fixed and resolved after verification; its next completed review covers the final head.
    • An independent re-audit verifies the repairs and the clarified contract. No automatic A+ grade from green checks alone.

    Preserve #194/#205 routing ownership, #190 semantic-claim safeguards and the existing word-page work. Keep #196/#201 open for their remaining panel/history/runtime scope. No merge, deployment, shared-database migration or data cleanup is part of this repair request.

  3. holden commented on Sep 27, 2026

    @holden
    MemberAuthor

    Audit repair pushed to PR #206 (head f2fc29c), and the publication contract in this issue is corrected. The PR is not merged, and this is not a readiness claim. An independent re-audit is requested.

    Publication-policy decision

    The body's section "Review, retention and publication" and its acceptance line now read differently. The earlier brief made a human acceptance of every composition version the gate for publication, and said background jobs never switch the pointer. That is not the intended product. Routine curation publishes on the configured panel's finalized consensus decision (4/5), provided the version passes the evidence and policy checks. There is no per-selection human sign-off.

    Audit findings

    # Finding Disposition
    1 Deleting a superseded claim revision revived its item Fixed. required_references, written by the database at insert, keeps a deleted reference distinct from one never used. Tested: publish → supersede → withhold → delete → still withheld
    2 A rejected defines claim authorized a lead Fixed. LeadRule uses Claims.visible(:public) for priority and fallback leads, and a priority entry must itself be eligible to lead. Tested for rejection and withdrawal through reading, publication and creation
    3 A configuration upgrade blocked reissuing an unchanged arrangement Fixed. Deduplication, its unique index and the fingerprint include the configuration version, and the reissue needs its own review
    4 An unrelated object could carry a catalog pin Fixed. A catalog pin must match the object's verified identifier, and a source-record work's record must have materialized the object. Catalog-only pins remain valid
    5 Caller-supplied note attribution (CodeRabbit) Fixed. Attribution comes from the acting principal (note_author_actor_id), and the database refuses anything else
    6 The human-only publication contract Reconciled, as above
    7 Formatting; PostgreSQL 15+ (CodeRabbit) Fixed. The requirement is documented and checked in the migration

    The repairs are a separate, reversible migration (20260927094256_repair_curation_integrity), so databases that ran the first one are repaired by ecto.migrate.

    Verification: at f2fc29c, on dedicated test databases only. mix precommit on a fresh database ran both migrations and 16 doctests, 2153 tests, 0 failures. The upgrade from b8211ed's schema migrated, rolled back to the previous state and re-migrated. The focused curation suite ran 65 tests, 0 failures (18 new in AuditFindingsTest), and the committed-connection races 4 tests, 0 failures. Formatting is clean. Three of the auditor's five probes pass verbatim. The other two now refuse the bad input at creation (:note_attribution_not_an_input, :work_identity_mismatch). One full run hit the pre-existing DataCaseTest claim-lock flake that PR #205 fixes; it is not touched here. Exact commands are in the PR description.

    This issue stays open for runs, ballots, the panel-decision authority, freshness and history.

  4. holden commented on Sep 27, 2026

    @holden
    MemberAuthor

    Final head for re-audit: f43d2f8 on PR #206.

    • Contents: the audit repairs from f2fc29c, a merge of main now that Routing Stage 1: durable pages, path ledger and resolver (#194) #205 has landed, and one test fix. With Routing Stage 1: durable pages, path ledger and resolver (#194) #205's pages tables present, the approval and publication tests check that no page row is written, instead of checking that no page table exists.
    • Precommit: mix precommit at f43d2f8 on a fresh MIX_TEST_PARTITION database gives 19 doctests, 2269 tests, 0 failures.
    • Upgrade path: migrate from b8211ed's schema, roll back the repair, re-migrate, then 65 curation tests and 4 committed-connection race tests, all passing.
    • CodeRabbit: no actionable comments on b8211ed..f43d2f8, and both earlier threads were verified and resolved by CodeRabbit.

    Not merged. This is not an A+ claim: an independent re-audit is requested.

  5. added a commit that references this issue on Sep 27, 2026
  6. holden commented on Oct 21, 2026

    @holden
    MemberAuthor

    Folded into #193, the curation programme's one open tracker, on the owner's word (21 October 2026). The specification here stands; reopen when the work starts. Status is in the consolidation comment on #193.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions