Skip to content

Curation: the next stretch to the agents — #267 and #264 in one audited order, deployment gates, and "ready to prepare the persona plan" (under #254, #253, #193) #268

Closed

Description

@holden

Revised 26 October 2026 after the independent audit (C+, revise before Stage 0). All six of its required changes are applied below; the list is the revision comment on this issue. The owner kept #253 as written: persona implementation waits for #253 §10's full MVP acceptance, including P3 and the ten pilot pages. This issue's end state is therefore "ready to prepare the persona plan", not "ready to build the agents".

Dedicated implementation issue for the next stretch of the curation MVP, under #254 (the plan), #253 (the specification, which governs on any conflict) and #193 (the programme). Written 26 October 2026 against main at 9ccc427. It adds no feature. It puts #267 and #264 in one order, with their parallelism, their deployment order and their gates, and it names everything else that still stands between the programme and persona implementation (§4), so the whole route to the agents is visible in one place.

1. Why this issue

The remaining work is spread over #267 (the runtime made live), #264 (P6, the proposal path), #254's unstarted P3, #253 §10's acceptance and five audit comments (four on #254, one here). None says, in one place, what order the rest runs in, what can run at the same time, what must be deployed before what, and what is still the owner's to decide. This issue does. #267 and #264 stay the build issues; Stage 0 (§5) is what corrects them.

2. Where the programme stands (26 October 2026, verified by the audit)

Part Issue / PR State
P1 published reader #255 / #256 Merged, live. On love is published: composition/version 1/1, 3 items, receipt 5. Its three notes are the accepted placeholders, signed "Account #2".
P2 owner's screen #257 / #258 Merged. The owner's three-page exercise has not happened; #257 closed on 26 Oct with the exercise transferred to #264 §8 step 2a.
P3 PoetryDB pin none Not started, no dedicated issue. Waits on the owner's two decisions (O1).
P4 packet v2 #262 / #265 Merged, live. Packets for love, war, hope, truth and coward build, freeze, verify and fit (the audit's read-only run).
P5 freshness #259 / #261 Merged, live on love: refresh row 1, generation 3, last outcome needs_editor (an input check, not an editorial judgment). #198 stays open: prompt suppression of already-open readers is unproved.
P7 runtime + benchmark #221 / #260, #263, #266 Merged. #266: GO, 9 of 9 (run 1793029211, attempts 52–70). Bench holds 3 model configs, 1 service, 70 attempts; v2 holds 0 / 0 / 0. The authority marker names the bench database; the service is stopped.
P6-0 runtime live #267 Not started. Only its GO box is ticked.
P6 proposal path #264 Not started. Its body still holds the pre-audit instructions and a warning comment that it is unreliable until corrected.
v2 curation state 5 proposed curator profiles, 0 profile versions, no generic editor; one enabled global-default configuration with one immutable manual version.

3. This issue's end state: "ready to prepare the persona plan"

This issue closes when all of these are true, each with its evidence linked here.

  1. The runtime is live on v2. One immutable model config (qwen3.5-9b-v2, generic-editor-v2) registered in v2 with its licence and hash; the service bound to v2; ready --smoke passes; three rehearsal attempts (love, war, hope) posted with readings and a confirmed stop. (Curation P6-0: the editor runtime made live — one model, where it lives, the v2 binding, the benchmark GO and a rehearsal against the real pages (under #254, #253, #193) #267 §12, as corrected in Stage 0.)
  2. One editor has gone end to end through the record, including its failures. The generic-editor profile admitted with a bot actor under a narrow admission rule; a real packet becomes a validated attempt, the attempt becomes a proposed version with model attribution and an attempt link; no bot operation moves a published pointer. Evidence also covers: a crash after the result receipt, a stale or uncertain result recovered, suppression and configuration change during a run, a review reservation recovered, a proposal superseded on arrival, and an offline service that defers without advancing freshness or delaying a reader.
  3. The owner has curated by hand. Three pilot pages curated on /ops/curation with the owner's own reasons under the owner's actual public label; love's placeholders replaced through new attributed versions; one page withdrawn, re-opened and re-published; all three end published.
  4. The baseline exists. Five real, non-abstained proposals on five pilot pages, each reviewed on the full quality form (a dismissal counts only with the full form and its reason); at least one edited or dismissed with a reason; a Take only where deserved; D4 (no self-publication) unchanged. The five forms posted on Curation agent MVP: what an editor does, what a curated entry contains, how the page looks, how the owner and the bots share one record, and how it stays current (under #193) #253.
  5. The shared editor seams are named, each with its file and test, so a later persona plan builds on them rather than beside them:
    • exact On-scope provisioning and server-side configuration resolution; the immutable editor/model/instruction binding and its invalidation on a new configuration version;
    • Candidates → frozen packet → registry, reference and lead-rule validation, the bounds, the omissions and the duplicate-display check;
    • refresh admission, the persisted needs_editor hand-off, the original published parent, generation, token and lease, suppression and recovery;
    • the runtime call under the one global slot and the 30-minute daily inference budget;
    • every non-proposal outcome (no result, unchanged, refused, uncertain), durable review-unit reservation and release, dismissal history, supersession;
    • the trusted proposal writer and its attribution; the trusted human Take/Edit path that preserves authorship, ancestry and compare-and-switch, with the quality form enforced;
    • D8 marks and D9's attempt audit.
  6. P6's post-activation re-audit is complete and satisfactory (Curation P6: the first bot proposes into the owner's screen — attempt to proposal, Take, Edit, Dismiss, the quality form, D8 marks and the D9 attempt audit (under #254, #253) #264 §8's "then the re-audit"), posted on Curation P6: the first bot proposes into the owner's screen — attempt to proposal, Take, Edit, Dismiss, the quality form, D8 marks and the D9 attempt audit (under #254, #253) #264 and linked here. A requested audit is not a completed one.

What a persona will need that this issue does not build, recorded so the persona plan starts from it (from the audit's answer 1): a versioned dossier-to-prompt seam with its own prompt identity, context budget and provenance; a frozen participant, profile and configuration identity per run; run, round and ballot persistence; roster routing; partial-result recovery; equal weights and the fixed 4-of-5; missing versus abstained responses; ten calls plus one repair; the 20-minute panel deadline and lease renewal beyond P5's 10-minute lease; dissent; finalized panel authority. A persona does not use the generic editor's no-dossier admission exception or the owner's Take/Edit controls; participant ballots are not each an owner-review proposal.

4. After this issue: what still gates persona implementation

#253 is unchanged, so persona implementation starts only when #253 §10 is fully accepted. After this issue closes, that still needs:

Remaining MVP item (#253 §10) Owner Can start
P3: the PoetryDB pin, its withhold-and-restore round trip, a pinned poem rendered as a highlight The owner's O1 decisions, then a dedicated P3 issue and session Now, in parallel, as soon as the owner takes O1. P3 touches discovery, the registry and the opening's work tile; agree the components/opening.ex boundary with #264 in Stage 0.
The other five pilot pages showing an opening (or an honest nothing) Owner, with the editor's proposals After item 4 above
§6a demonstrations (owner-curated page receives a proposal without moving; withdrawn page receives none; superseded_on_arrival) #264 Delivery B evidence With items 2–4 above
Restore from a bundle, including the new proposal, mark, dismissal and model-config records A bounded check under #211's bundle After item 4
#198's prompt suppression of an already-open reader Unassigned; needs a delivery decision Any time; do not close #198 on P6 counts

When those are done, the persona plan (drafted any time after this issue) gets its own independent audit, and persona implementation starts. Fastest route to the agents: take O1 now so P3 runs alongside Stages 0–3, rather than after them.

5. The plan

Stage 0  design: #264's resolved contract + #267 reconciled ──► one independent check
            │
Stage 1  ┌─ #267 Delivery A (rehearse, Health card, runbook) ─┐  two worktrees on the Studio,
         └─ #264 Delivery A (disabled, stubbed runtime)      ─┘  in parallel; P3 alongside if O1 is taken
            │
Stage 2  merge #267 and deploy it  ──►  rebase #264, integrated review + precommit, merge, DO NOT deploy
            │
Stage 3  3a #267 live window: idle check, start + rebind to v2, register, smoke, rehearse ×3, stop
         3b #264 migration on v2 (owner's yes), editor disabled ──► then deploy #264's code
         3c editor configuration activated; love re-established under it; owner's three pages;
            five page identities provisioned; proposal path enabled
         3d proposal window: love, then war, hope, truth, coward; five forms; stop
         3e P6 re-audit
            │
         end state (§3) ──► §4's remaining MVP work ──► persona plan audited ──► persona build

Stage 0. Resolve #264's contract and reconcile #267 (a design session, no code)

The nineteen corrections (the fourth audit's §E, on #254) are necessary but not sufficient: several say "define" or "resolve" rather than giving the resolved rule. Stage 0 produces the resolved contract, written into #264's body with its schema, migration and test matrix, and a comment on #264 listing what changed. Corrections 1–2 state today's fulfilled GO (#266), not the fourth audit's open-PR state. It must resolve, at least:

  • Ancestry. editorial_composition_versions_shape requires (version = 1) = (parent_version_id IS NULL), but the proposal's parent is the published pointer captured at admission, which can be null at version 2 or later (after an unpublished draft, a first unaccepted proposal, or withdraw and re-open). Separate published-parent from version ancestry in the scoped migration; never invent a published parent.
  • Retry identity. Refresh.act(:admit, …) keeps a deferred row's generation but recaptures the current published parent (refresh.ex:364). State what stays frozen on retry and recovery, and when changed inputs require a new generation and request key.
  • Enablement versus activation. The configuration manifest (Configurations.manifest/1, private), ready?/1, resolve_default/1 and current/1 know only manual mode. Name their changes, the owner's explicit immutable-version creation and activation step, and what happens to love's published version and the owner's manual versions, which Standing withholds once made under an older configuration version.
  • Carry-forward (O2). A kept reference outside the frozen packet is not a model-selected candidate: give copied owner items a separately verified source identity and truthful human attribution. The arrangement still fits one lead and three highlights: specify precedence and overflow when kept items and model choices exceed it, revalidation when marks change during a run, dedup on the actual arrangement, and revised tests. A blanket "model dropped a kept item → refuse" conflicts with carry-forward and is replaced.
  • Frozen reference reconstruction. Packet.summary stores excerpt_sha256, not the quotation's words_sha256 (confirmed on bench attempt 59). The proposal writer reconstructs against stored frozen identities and hashes, or the attempt persists the needed reference metadata, and refuses drift. The quote's substring hash stays distinct from the full quotation's reference hash.
  • Storage choices. Lifecycle storage (or tightly bounded status mutability) for immutable versions; the honest uniqueness rule for an unchanged Take; durable pre-run review-unit reservation, recovery and release (Gateway's pending units count inference, not reviews); exactly-once attempt-to-proposal linkage.
  • No bypass. The editor's existing direct Publish of the latest held version and the ordinary manual writer (which stamps notes human) must not bypass Take, the quality form or provenance for a proposal.
  • Abstention. The editorial-outcome storage and disclosure contract for a successful abstention, including its 48-hour assessment and cooldown, without loosening RefreshState's outcome CHECK by accident.
  • Backward-compatible migration. The scoped migration must be safe under the old code (Stage 3b applies it before Curation P6: the first bot proposes into the owner's screen — attempt to proposal, Take, Edit, Dismiss, the quality form, D8 marks and the D9 attempt audit (under #254, #253) #264's code is deployed): additive, nullable or defaulted, constraints only widened for the new path.
  • Reconcile Curation P6-0: the editor runtime made live — one model, where it lives, the v2 binding, the benchmark GO and a rehearsal against the real pages (under #254, #253, #193) #267 with §6.1 and §6.2 below, and replace Curation P6: the first bot proposes into the owner's screen — attempt to proposal, Take, Edit, Dismiss, the quality form, D8 marks and the D9 attempt audit (under #254, #253) #264's comment "P6's own build starts after Curation P6-0: the editor runtime made live — one model, where it lives, the v2 binding, the benchmark GO and a rehearsal against the real pages (under #254, #253, #193) #267's Delivery A is merged" with the parallel rule in Stage 1.
  • Regenerate both file manifests from the resolved contract; the overlap in §6 is the briefs' current lists, not a guarantee.

Stage 0 ends with one independent check of the resolved #264 and the reconciled #267 before any build session starts.

Stage 1. Two build sessions in parallel

#264's Delivery A is built disabled, with the runtime stubbed through Req.Test; it needs neither #267's code nor a live model. Each session in its own worktree on the Studio, its own partition (_cp60 for #267, _cp6 for #264), stopping at a reviewed PR (CodeRabbit and one independent review). Coordinate the two full precommit runs; the Studio is shared. File ownership is §6.

Stage 2. Merge order and deployment order are separate

Deploying here means advancing the Studio checkout's main, which the server on 4007 runs from.

  1. Merge Curation P6-0: the editor runtime made live — one model, where it lives, the v2 binding, the benchmark GO and a rehearsal against the real pages (under #254, #253, #193) #267 and deploy it. It has no migration; its Health card must render on v2 with zero model configs and services (§6.2).
  2. Rebase Curation P6: the first bot proposes into the owner's screen — attempt to proposal, Take, Edit, Dismiss, the quality form, D8 marks and the D9 attempt audit (under #254, #253) #264 on it; run the full precommit on the integrated head and review the integrated result (a clean merge is not a compatibility test); merge Curation P6: the first bot proposes into the owner's screen — attempt to proposal, Take, Edit, Dismiss, the quality form, D8 marks and the D9 attempt audit (under #254, #253) #264 on GitHub.
  3. Do not deploy Curation P6: the first bot proposes into the owner's screen — attempt to proposal, Take, Edit, Dismiss, the quality form, D8 marks and the D9 attempt audit (under #254, #253) #264 yet. Its new Ecto fields would be selected before their columns exist: Published.current does Repo.get!(CompositionVersion, …) (published.ex:68), so an attempt_id field breaks every published page until the migration runs. Deployment waits for Stage 3b.

Stage 3. Live steps, each on the owner's named yes in chat

3a. #267's live window (no proposals written):

  1. Confirm the bench slot is idle (budget, sweep, the services row) and handle the already-stopped state.
  2. Start the service and rebind it to v2 (service start --rebind); record the marker's new contents.
  3. Register the pinned config against the running service (setup --accept-license; registration reads /api/version, /api/tags and /api/show, so the service must be up). The owner reads the licence. Record the row id and config_hash; it should equal bench config 3's 5b8b65ad… if the inputs are identical.
  4. ready --smoke; record the receipt. The smoke writes an attempt, so v2 holds four attempts after the rehearsals, not three.
  5. Rehearse love, war and hope, --dry-run first on each; post the records.
  6. Stop, confirm, readings 30 s after.

3b. #264's migration, then its code. On the owner's yes: apply the one scoped migration to v2 from a worktree at the merged head, with the editor still disabled. Verify the old code still serves (it must, by Stage 0's compatibility rule). Then deploy #264's code.

3c. Activation, each step on its own yes:

  1. Seed and admit the generic-editor profile under its narrow admission rule.
  2. Create and activate the immutable configuration version that selects the editor and the model config (Stage 0's contract); re-establish love's publication under it as Stage 0 specifies.
  3. The owner's three pages (§3 item 3): actual label, own reasons, love's placeholders replaced, one withdrawn, re-opened and re-published, all three published.
  4. Provision the composition identities for all five proposal pages through the authorized seam. Truth and coward have none today, nor do war and hope; a visit cannot create them.
  5. Enable the fenced proposal path.

3d. The proposal window. Start the service; love first, then war, hope, truth and coward; the owner reviews each on the full form; stop and confirm. Post the five forms on #253 and the ids here.

3e. The P6 re-audit (§3 item 6).

6. File ownership and interfaces for the parallel builds

  1. refresh_worker.ex belongs to Curation P6: the first bot proposes into the owner's screen — attempt to proposal, Take, Edit, Dismiss, the quality form, D8 marks and the D9 attempt audit (under #254, #253) #264 alone. On main the worker calls only Refresh.claim and Refresh.check; it never calls the runtime, so it cannot meet a down model service. The service_unreachable refusal and backoff, and Curation P6-0: the editor runtime made live — one model, where it lives, the v2 binding, the benchmark GO and a rehearsal against the real pages (under #254, #253, #193) #267's test 4, move to Curation P6: the first bot proposes into the owner's screen — attempt to proposal, Take, Edit, Dismiss, the quality form, D8 marks and the D9 attempt audit (under #254, #253) #264's fenced editor step, keeping the rule: offline inference never advances editorial freshness or delays a reader. Move every Curation P6-0: the editor runtime made live — one model, where it lives, the v2 binding, the benchmark GO and a rehearsal against the real pages (under #254, #253, #193) #267 reference together: D-F, the §5 seam row, §6 scope, §7 files, §8 test 4, §12 acceptance and §13's prompt. Curation P6-0: the editor runtime made live — one model, where it lives, the v2 binding, the benchmark GO and a rehearsal against the real pages (under #254, #253, #193) #267's runbook says visits invoke an editor only after Curation P6: the first bot proposes into the owner's screen — attempt to proposal, Take, Edit, Dismiss, the quality form, D8 marks and the D9 attempt audit (under #254, #253) #264 is enabled.
  2. health_live.ex is shared by card, over a defined interface. Curation P6-0: the editor runtime made live — one model, where it lives, the v2 binding, the benchmark GO and a rehearsal against the real pages (under #254, #253, #193) #267 owns the "Editor runtime" card; Curation P6: the first bot proposes into the owner's screen — attempt to proposal, Take, Edit, Dismiss, the quality form, D8 marks and the D9 attempt audit (under #254, #253) #264 adds its own card below. The card is non-mutating and absent-data safe: it must not call Readiness.check or Gateway.budget unchanged, because both reach Gateway.service!, which inserts (gateway.ex:34–39); it reads a SELECT-only snapshot and renders "none" where v2 has no config, service or refusal. Pending review units belong to Curation P6: the first bot proposes into the owner's screen — attempt to proposal, Take, Edit, Dismiss, the quality form, D8 marks and the D9 attempt audit (under #254, #253) #264's card; Curation P6-0: the editor runtime made live — one model, where it lives, the v2 binding, the benchmark GO and a rehearsal against the real pages (under #254, #253, #193) #267's card shows inference minutes only and does not imply zero review work from an idle slot. Stop history: inference_services stores no window or stop history, so the card shows current state only; the confirmed stop and readings live in the rehearsal and window records. Both sessions agree this interface in Stage 0, then rebase and review the integrated result.
  3. components/opening.ex is Curation P6: the first bot proposes into the owner's screen — attempt to proposal, Take, Edit, Dismiss, the quality form, D8 marks and the D9 attempt audit (under #254, #253) #264's for the disclosure changes; if P3 runs in parallel, its work tile is agreed with Curation P6: the first bot proposes into the owner's screen — attempt to proposal, Take, Edit, Dismiss, the quality form, D8 marks and the D9 attempt audit (under #254, #253) #264 in Stage 0.

7. Owner decisions

Decision Default
O1 P3's two rules: the poem-to-page relation (#253 D6) and PoetryDB supersession (#254 P3 "Identity") Open. Deferring them does not block this issue, but P3 gates #253 §10 and therefore persona implementation; taking them now lets P3 run in parallel (§4).
O2 Kept items outside the packet's cut (love's published quotation 248575 is omitted over_quota) Trusted carry-forward with precedence and overflow rules, resolved in Stage 0; P4's policy unchanged. If Stage 0 cannot resolve it without new scope, P6 stops and says so.
O3 The five proposal pages love, war, hope, truth, coward. Each has an admissible Bierce lead; all five packets verify and fit (audit: love 10/32, war 9/29, hope 8/17, truth 7/16, coward 6/6). Revalidated at run time; identities provisioned in 3c. They do not test no-lead pages or live exemplars.
O4 The public actor label replacing "Account #2" The owner's own choice at 3c step 3, recorded with the new attributed versions.
O5 Parallel builds Yes, after Stage 0 settles §6 and the deployment rule. Parallel building authorizes no live step and waives no review.

8. What this issue does not do

No code, migration, model call, service start or database write. It changes no part's brief directly; Stage 0 does that. #253 is not amended. Personas, ballots, panels, a second model, the Mac mini, #190 (needed only for bot nominations) and automatic publication are out, as in #253 §10.

9. Audit record

  • 26 Oct 2026: independent audit of the first version, C+, "revise before Stage 0". Its six required changes are applied in this revision. Its recommendations are kept as notes: preserve the measured cut-v1 baseline; freeze evaluation pages and the held-out policy before any persona tuning; update docs/curation/runtime-operations.md (it still says source and Postgres stay internal) and Delivery: curated opening section with five local persona editors and visit-driven refresh #193's stale body status when those are next reconciled.
  • Next: the independent check that closes Stage 0.

10. Acceptance

Activity

  1. holden commented on Oct 26, 2026

    @holden
    MemberAuthor

    Independent audit of #268 — 2026-10-26

    Verdict: revise before Stage 0. The status summary is substantially accurate, and parallel stubbed builds are feasible, but the proposed agent-readiness gate and live deployment sequence are not safe as written. The nineteen corrections are necessary, not yet a sufficient executable contract.

    Assessed part Grade
    Plan overall C+
    §3 end state C
    §4 order / parallelism C+
    §5 scope moves B−
    Sufficiency of the nineteen corrections C+
    §6 defaults B−

    These grade this plan's present implementation readiness, not the quality of unbuilt #264/#267 code.

    Evidence and limits

    I read #268, #193, #253, #254 and all its comments, including all four audits, the reconciliation and the final P7 GO; #267, #264, #198, #257, #259 and #262 and their relevant delivery comments; both 26 October benchmark records and runtime-operations.md. I inspected the curation schemas, refresh/profile/configuration/composition/review/authority/runtime paths, curation LiveViews, Health, and curation migrations, and compared the relevant constraints/triggers with the actual database catalog.

    Audited main: 9ccc42748284a2610d45509aba6c8b1a157ecf46. Studio HEAD, its existing origin/main and GitHub main agreed. I did not pull. GitHub independently confirms PRs #256/#258/#260/#261/#263/#265/#266 merged; its open-PR collection was empty. There is no Proposals or Rehearsal implementation on this main. That establishes the delivered-code position, not the absence of private unpublished work.

    I connected to the Studio as holden and read its AGENTS.md. Both database inspections used PostgreSQL 18.2, port 5434, cluster 7693849764459364596, SELECT only, with default_transaction_read_only=on. A separate read-only Repo process, without the application/Oban/endpoint, built, froze, verified and checked estimated fit for the five proposed pages, plus greed/bestiality. I independently recomputed both benchmark verdicts from committed raw samples with main's Bench.summarize/verdict; this called no model.

    Own suite, once: MIX_ENV=test MIX_TEST_PARTITION=_audit268 DD_NO_OBAN=1 mix precommit, detached external-drive worktree at the audited SHA, deps/data and every replay gzip symlinked from main: 19 doctests, 2,986 tests, 0 failures; seed 86311; ExUnit 405.4 s; exit 0. The worktree was clean and was removed; main remains clean. This tests main, not the unbuilt proposal path.

    No live/bench writes, migrations, jobs, model calls, pulls, downloads, service start/stop/rebind, or requests to/restarts of 4007 were performed. I did not exercise signed-in owner controls or remeasure layout. Historical owner approvals and confirmed stops are documented evidence; I did not recreate them. Process/listener inspection found no model-service pid file, listener on 11435, or matching dictionary model process.

    §2: claims versus verified state

    Claim Independent result
    P1/P2 merged; love published; owner exercise outstanding Confirmed merge/code state. v2 has 1 composition, 1 version, 3 items, 1 review, 5 composition-publication receipts. Composition/version 1/1, final receipt 5; all three notes remain the quoted placeholders, signed Account #2. This is not three screen-curated pages. #257 is closed with its explicit transfer to #264, rather than evidence that the exercise happened.
    P3 not delivered No pin path in the audited curation code; source-record works remain unsupported by PublishedReader. No dedicated open P3 pin implementation issue appeared in the PoetryDB issue search. The owner decisions remain open.
    P4/P5 merged and live Confirmed code, applied refresh table, live packet builds and love's refresh state. Refresh row 1, configuration version 1, idle, generation 3, token 4, admitted parent 1, no retry/failure/dirty/lease pending. Last outcome needs_editor / fingerprint_version, completed 2026-10-26 13:59:58.254863 UTC, next check 28 October at the same UTC time. This remains an input check, not an editorial judgment.
    P7 GO, benchmark separate from v2 Recalculated morning 8/9 NO-GO and follow-up 9/9 GO. Bench has 3 model configs, 1 service, 70 attempts, 210 ledger entries; v2 has 0/0/0/0 respectively. Run 1793029211, attempts 52–70, has 13 accepted selections, 6 abstentions, no refusals, all recorded swap readings zero. Config 3, qwen3.5-9b-v2 / generic-editor-v2, hash 5b8b65adc7709921eb5cd06ba9d4dc1b96404e82ab04a5eb59bb709021da62cd. Bench service available, epoch 1, fence 70, no holder. Across bench's history there are 70 reserves/releases/charges and no net reservation outstanding.
    P6-0/P6 unbuilt Confirmed on main and v2. #267 has only its GO acceptance box ticked; #264 still contains the incompatible original instructions and its warning comment.

    Additional SELECT inventory: v2 has 5 proposed curator profiles, 0 profile versions, no admitted generic editor, one enabled global-default configuration, one immutable manual configuration version, zero roster members and one activation receipt. Bench has zero compositions, versions, items, reviews, composition-publication receipts, curator profiles or refresh rows. Page-publication receipts, a different subsystem, number 114 on v2 / 0 on bench.

    Oban by worker/state on v2: RefreshWorker 3 completed; CleanupWorker 672 completed; Discovery.RunWorker 67 completed + 2 cancelled; Quotations.VerifyWorker 337 completed. Bench has no jobs. The refresh generations match P5's two checks and P4's transition. Older audits' discovery totals are historical snapshots, not today's totals.

    The authority marker still names devils_dictionary_runtime_bench on the expected cluster. A database slot marked available is not proof that the OS service is running.

    Answers to §8

    1. Is §3 the right end state; what seams are missing or not reusable?

    It is a useful generic-editor hand-off, but not the governing bar for starting persona implementation. #253 §11 explicitly puts personas/panels after the MVP; §10 includes ten pilot pages, the PoetryDB pin/withhold/restore round trip, freshness/owner-rule demonstrations and recovery evidence. #268 expressly excludes two of those deliverables. Saying “personas lose nothing” is a technical dependency argument, not permission to change the governing scope. Preparing a persona plan can precede MVP completion; calling that “ready to build” cannot, unless the owner explicitly changes #253.

    Item 5 also describes less than the shared editor contract. Add named, tested seams for:

    • exact On-scope provisioning and server-side configuration resolution; immutable editor/model/instruction binding and configuration invalidation;
    • Candidates → frozen packet → registry/reference/lead-rule validation, bounds, omissions and duplicate-display checks;
    • refresh admission, the persisted needs_editor hand-off, original published parent, generation/token/lease, suppression and recovery; the runtime slot fence does not supply these;
    • no-result/unchanged/refused/uncertain outcomes, durable reservation recovery, dismissal/veto history, and supersession;
    • the trusted human Take/Edit path preserving authorship, ancestry and CAS authority, with quality-form enforcement.

    A persona needs a versioned dossier-to-prompt seam and a frozen participant/profile/configuration identity. Current Packet.messages accepts pinned generic instructions; Runtime checks the packet instruction against the immutable model config. A dossier cannot simply be appended without designing its prompt identity, context budget, provenance and validation.

    The persona plan must own run/round/ballot persistence, routing the admitted roster, partial-result recovery, equal weights/fixed 4-of-5, missing versus abstained responses, the ten normal calls/one repair maximum, the 20-minute panel deadline, dissent and finalized panel authority. A panel may outlast P5's current 10-minute running lease; renewal/recovery must be designed. These are future-phase work, not reasons to implement a panel in #264.

    What a persona does not use: the generic editor's special no-dossier admission exception, human quality-form controls, or human Take/Edit buttons. The owner uses those. D8 constraints and D9 audit still apply to persona output. Nor should each ballot automatically become a separate owner-review proposal/reservation; distinguish participant output from the final reviewable composition.

    2. Are the §5 scope moves sound?

    5.1: accept, with all references reconciled. RefreshWorker only calls Refresh.claim and Refresh.check. Refresh reads/computes inputs and completes or defers. Neither calls Runtime.run or the model. A down model service therefore cannot be its present failure.

    The service-unreachable refusal/backoff and #267 test 4 belong to the new fenced editor step in #264. Preserve that test's rule: offline inference never advances successful editorial freshness or delays readers. Move #267's D-F assertion, §5 seam row, §6 scope, §7 file, §8 test, acceptance and implementation prompt together, not just one bullet.

    Nothing in rehearsal itself needs this worker behavior: #267 calls Runtime directly and can record its own refusal. Its “last worker refusal” Health row must tolerate none/not yet implemented until #264 lands. The operating-window runbook must identify that visits actually invoke an editor only after #264 is enabled.

    5.2: accept layout ownership, reject the implied sufficiency. Separate cards do not define their data/API dependencies. In particular, Readiness.check calls Gateway.service!, and Gateway.budget defaults through service!; service! executes INSERT ... ON CONFLICT DO NOTHING. Using those APIs unchanged is not a read-only Health implementation. It must use an explicitly non-mutating snapshot/query and handle absent live rows/configs.

    Gateway.budget reports inference milliseconds, not pending proposal review units. That count belongs to #264's durable accounting. #267 must display unavailable until that seam exists, or put the review count on #264's card. Do not infer zero pending review work from an idle inference slot. Also name the evidence source for “last window / stop confirmed”: ServiceProcess.stop returns confirmation and removes the pid file; inference_services does not store window/stop history. An absent listener is not a historical confirmed-stop record.

    5.3: refuse as an unconditional agent-phase scope change. P3 need not block #267/#264 engineering or the first generic proposal. It still gates MVP completion, which #253 currently puts before persona implementation.

    3. Is parallel Delivery A safe, and what if #264 merges first?

    The complete literal intersection of #267 §7 and #264 §5 is:

    File Parallel worktree safety
    lib/devils_dictionary/curation/refresh_worker.ex Both briefs currently own it: unsafe shared contract until Stage 0 moves the service-down behavior entirely to #264. Then #267 should not edit it. Sequential scheduling alone would not make a model-free worker detect a down model.
    lib/devils_dictionary_web/live/health_live.ex Separate worktrees prevent filesystem interference; separate cards reduce textual conflicts. Both can still touch aliases, mount assigns/helpers and tests. Agree the read-only snapshot/optional review-capacity interface first, then rebase and review the integrated result. A clean merge is not a compatibility test.

    No other file is literally listed in both sections. The corrected briefs' manifests must be regenerated: configuration enablement, lifecycle/capacity storage and any extracted Bench measurement helper may require additional files. “Only expected conflict is Health” is a prediction, not evidence.

    Code merge order is not a hard runtime dependency: a correctly disabled, stub-tested #264 could merge before #267 without itself needing Rehearsal. Its live activation would lack #267's command/runbook/evidence, and the current #264 warning comment explicitly requires #267 Delivery A merged first; reconcile that instruction if parallel work is accepted. Any eager call to a future Health API would additionally break the build.

    The larger defect exists in either merge order: both PRs merged, then #267 live work, then #264 migration leaves a deployment gap. Pulling/loading that combined head exposes Ecto schemas with columns/tables absent on v2. For example, Published.current uses Repo.get! on CompositionVersion; adding attempt_id makes that query select a nonexistent column before migration. Composition mode and proposal telemetry create similar exposure. The no-config-change reloader is not a migration barrier.

    Specify a compatible migration-before-new-schema-code deployment, or explicitly isolated deployment of #267 while #264 remains undeployed. This is a static consequence of the proposed schema/read path, not a failure I induced on 4007. Rebase and final combined-head verification remain required regardless of merge order.

    4. Do the nineteen corrections make #264 executable as written?

    Not yet. Their diagnoses remain mostly correct, but several say “define/resolve/specify” rather than supply the resolved contract. Stage 0 must produce that contract and revised tests, not paste nineteen summaries and declare readiness.

    The existing live CHECKs/triggers confirm the original problems: manual-only version/item origins, human version author, immutable versions/items, sourced/deceased profile admission, human-note author equality and actor-label equality, the arrangement/fingerprint dedup index, and refresh outcomes limited to unchanged/needs_editor.

    Further concrete conflicts/unfinished resolutions:

    • Corrections 3/4/7/18 — unpublished ancestry: editorial_composition_versions_shape still requires (version = 1) = (parent_version_id IS NULL). The required parent is the published pointer captured at admission. After an unpublished draft, a first unaccepted proposal, or withdrawal/re-open, that pointer can be null for version 2 or later. Merely adding model origin/bot permission still fails this CHECK. Resolve published-parent versus version ancestry in the scoped migration; do not invent a published parent.
    • Corrections 3/4/16 — retry identity: main Refresh.act(:admit, ...) retains the generation from a deferred row but recaptures the current published parent. Reusing that unchanged in P6 can change ancestry under the same request identity. State what stays frozen on retry/recovery and when changed inputs require a new generation.
    • Corrections 5/18 — enablement is not activation: Configurations.manifest emits manual mode, ready? recognizes only that mode, and resolve_default/current use it. Name changes to these functions as well as the schema/manifest. Add the owner's explicit immutable-version creation/selection/activation step. Standing currently withholds versions made under an older configuration version; plan when love and the owner's manual versions are re-authored/reviewed under the new one.
    • Corrections 12 versus 17 and Curation P6: the first bot proposes into the owner's screen — attempt to proposal, Take, Edit, Dismiss, the quality form, D8 marks and the D9 attempt audit (under #254, #253) #264 test 4 — carry-forward: a kept reference outside the frozen packet cannot be resolved as a model-selected candidate. Give trusted copied owner items a separately verified source identity and truthful attribution. The final arrangement must still fit one lead + three highlights: three kept highlights plus three different model highlights cannot all survive. Specify precedence/overflow behavior, revalidation of changed marks, actual-arrangement dedup and revised tests. A blanket “model dropped a kept item → refuse” conflicts with the chosen carry-forward default.
    • Correction 17 — frozen reference reconstruction: Packet.summary omits the quotation's words_sha256 even though its packet candidate carries it; the summary stores excerpt_sha256 instead. I confirmed that shape on persisted bench attempt 59. The proposal writer must reconstruct against the stored frozen identities/hashes, or persist the necessary reference metadata, and refuse drift. Hashing whatever happens to be at the locator now is not proof of what the model read. Keep quote substring hash/range distinct from the full quotation's reference hash.
    • Corrections 9/13/14 — storage decisions remain outstanding: choose lifecycle storage/status mutability, the honest unchanged-Take uniqueness rule, and durable pre-run review reservation/recovery/release. Preserve exactly-once attempt-to-proposal linkage and immutable selections. Existing Gateway pending units count live inference attempts, not unresolved reviews.
    • Corrections 13/15/19 — the existing Publish shortcut matters: Editor.held/publishable can publish the latest held version directly; ordinary write_version stamps notes human. A proposal strip alone does not repair either path. Ensure ordinary Publish/Save/Edit routes cannot bypass the required trusted Take/quality-form/provenance rules. Require five completed baseline forms even when a proposal is dismissed, with the rejection reason retained; a reason-only dismissal is not one of the five fixed forms.
    • Correction 8: decide the actual editorial-outcome storage/disclosure contract. “Do not add abstained unless enum/CHECK/disclosure change together” prevents one bad implementation but does not specify a successful abstention's 48-hour assessment/cooldown behavior.

    Corrections 1–2 should use today's fulfilled GO rather than preserve the fourth audit's historical open-PR state. Corrections 6/7/10/11 are appropriate narrow changes; they are not permission to disable general immutability, deceased-person admission or human review/publication authority.

    5. Are O1–O5 safe defaults?

    Default Decision
    O1 Safe to defer the two P3 decisions while building the generic path. Unsafe to use that deferral to authorize persona implementation under unchanged #253.
    O2 Correct direction, not yet an implementable default until the bounded merge, copied provenance, omitted meaning/reference, mark-change and test conflicts above are resolved. The explicit stop-on-blocker remains appropriate.
    O3 Accept as the generic baseline, subject to current packet revalidation at actual run time. My read-only results follow below. This is not broad coverage or a held-out panel result.
    O4 Correct owner-owned step, not a chosen public label. Record the actual label and replace the notes through new attributed versions; re-open must end with re-publication, and all three owner pages must finish published.
    O5 Parallel stubbed engineering is reasonable after interface/ownership/deployment corrections. It does not authorize live steps, waive independent review, or make the old briefs executable.

    Own Candidates/Packet results, generic-editor-v2, all freeze/verify :ok and estimated fit :ok at 8,192 context / 1,024 output:

    Page Admissible leads Candidates kept / available Meanings kept / total Omissions
    love 1 10/32 12/48 22 over_quota
    war 1 9/29 12/37 20 over_quota
    hope 1 8/17 12/58 9 over_quota
    truth 1 7/16 12/22 9 over_quota
    coward 1 6/6 12/12 none

    Love/war/hope hashes reproduce the previous audit's values; truth is 554b8b5e84d9b2866d52242678d5aa2c81ba4f1982080a54de19cdbcef3f36dc, coward ee5270fec4c99b6888ddabb16cf7aea457f4ac86b204a77e8f3d68c93b738537. Love's current published item 3, quotation-248575-3, is indeed omitted over_quota. Greed/bestiality build at 7/7 and 3/3, with zero admissible leads. No accepted exemplar was offered on these pages; works on the first three are catalog artworks.

    6. What path/gates are missing?

    Beyond the conflicts above:

    Required changes to #268 — before Stage 0 starts

    1. Correct the authority of §3/§5.3/O1: either retain full Curation agent MVP: what an editor does, what a curated entry contains, how the page looks, how the owner and the bots share one record, and how it stays current (under #193) #253 MVP acceptance before persona implementation, or obtain an explicit owner amendment to Curation agent MVP: what an editor does, what a curated entry contains, how the page looks, how the owner and the bots share one record, and how it stays current (under #193) #253. A narrower “ready to prepare the persona plan” milestone is also honest. Do not silently redefine the MVP.
    2. Expand Stage 0's deliverable: a resolved Curation P6: the first bot proposes into the owner's screen — attempt to proposal, Take, Edit, Dismiss, the quality form, D8 marks and the D9 attempt audit (under #254, #253) #264 contract/schema/test matrix covering the additional ancestry, retry, enablement, carry-forward, Take, reference reconstruction and accounting issues above. Reconcile Curation P6: the first bot proposes into the owner's screen — attempt to proposal, Take, Edit, Dismiss, the quality form, D8 marks and the D9 attempt audit (under #254, #253) #264's existing “start after Curation P6-0: the editor runtime made live — one model, where it lives, the v2 binding, the benchmark GO and a rehearsal against the real pages (under #254, #253, #193) #267 Delivery A merges” comment and all Curation P6-0: the editor runtime made live — one model, where it lives, the v2 binding, the benchmark GO and a rehearsal against the real pages (under #254, #253, #193) #267 worker references with the accepted parallel contract.
    3. Add the deployment gate: named owner approval for the scoped Curation P6: the first bot proposes into the owner's screen — attempt to proposal, Take, Edit, Dismiss, the quality form, D8 marks and the D9 attempt audit (under #254, #253) #264 migration and an old/new code compatibility plan, before new schema-reading code reaches 4007. Separate merge order from deployment order.
    4. Define Health's non-mutating, absent-data-safe interface and review-capacity ownership, including truthful stop-history evidence. Separate template cards alone are insufficient.
    5. Rewrite the live checklist: correct service/registration order; immutable editor configuration activation and its effect on existing versions; all five exact identities; all three manual pages ending published with the actual label/reasons; offline/no-result handling and every required approval.
    6. Restore evidence gates: final rebased/integrated review and checks, P6 post-activation re-audit, five actual forms, MVP/restore/residual-contract ownership, and an unambiguous audited persona hand-off. Closure checkboxes must match the chosen end-state wording.

    These are plan/contract corrections requested by this audit; I have made none of those edits.

    Recommendations, separate from required changes

    • Keep the two implementation worktrees if their owners agree the above interfaces; coordinate expensive suites on the shared Studio. No model is needed for Delivery A.
    • Preserve the measured cut-v1 baseline. Its quotation clustering and unused capacity are legitimate later product questions, not fixes to smuggle into P6.
    • Freeze evaluation pages/packets and held-out policy before tuning personas. These five lead-bearing pages do not test no-lead behavior or live accepted exemplars.
    • Update stale present-tense tracker/runbook statements during their authorized reconciliation: runtime-operations still says source/Postgres stay internal; Delivery: curated opening section with five local persona editors and visit-driven refresh #193's body retains obsolete runtime/freshness status despite its newer pointer comment.

    Ranked risks

    1. High — stale or unauthorized writes / false provenance: refresh authority lost between checking and proposing; copied owner/model notes relabelled; Take or ordinary Publish bypassing the quality gate.
    2. High — deployment outage: new Ecto schemas/telemetry loaded before their live migration.
    3. High — readiness declared against a reduced specification: five forms used to bypass ten-page/P3/recovery acceptance and the P6 re-audit.
    4. Medium — stranded/duplicated work: parent recapture, changed packet under one generation, nil-parent CHECK failure, immutable lifecycle updates and leaked review reservations.
    5. Medium — owner intent lost at the cut/cap: kept items outside the packet, capacity overflow or marks changed during a run.
    6. Medium — GO overinterpreted: the fixed-seed fixture run validates its declared runtime checks, not real-page editorial quality, sparse real-evidence behavior, persona prompting or panel benefit. The crane reason's “general” definition claim remains a concrete quality warning.

    Smallest safe sequence

    1. Correct Curation: the next stretch to the agents — #267 and #264 in one audited order, deployment gates, and "ready to prepare the persona plan" (under #254, #253, #193) #268's scope/gates, then Stage 0 resolves and reconciles the two build briefs with the explicit contracts above.
    2. Build Curation P6-0: the editor runtime made live — one model, where it lives, the v2 binding, the benchmark GO and a rehearsal against the real pages (under #254, #253, #193) #267 and disabled/stubbed Curation P6: the first bot proposes into the owner's screen — attempt to proposal, Take, Edit, Dismiss, the quality form, D8 marks and the D9 attempt audit (under #254, #253) #264 in isolated worktrees under agreed ownership; review both. Merge Curation P6-0: the editor runtime made live — one model, where it lives, the v2 binding, the benchmark GO and a rehearsal against the real pages (under #254, #253, #193) #267 first as proposed, rebase Curation P6: the first bot proposes into the owner's screen — attempt to proposal, Take, Edit, Dismiss, the quality form, D8 marks and the D9 attempt audit (under #254, #253) #264, and verify/review the integrated result.
    3. Before combined code deployment, perform the separately approved compatible Curation P6: the first bot proposes into the owner's screen — attempt to proposal, Take, Edit, Dismiss, the quality form, D8 marks and the D9 attempt audit (under #254, #253) #264 schema migration with the editor still disabled, or deploy only Curation P6-0: the editor runtime made live — one model, where it lives, the v2 binding, the benchmark GO and a rehearsal against the real pages (under #254, #253, #193) #267 until that migration can precede Curation P6: the first bot proposes into the owner's screen — attempt to proposal, Take, Edit, Dismiss, the quality form, D8 marks and the D9 attempt audit (under #254, #253) #264's schema reads. Keep 4007's launch script untouched; this audit grants no operational approval.
    4. On the owner's named approvals, verify the old runtime slot idle, open the v2-bound service window, register the pinned config/licence, smoke, rehearse love/war/hope, stop and retain confirmation/readings. Nothing here writes proposals.
    5. Activate the resolved immutable editor binding on the owner's approval, preserve/re-establish valid manual publications under that configuration, complete the owner's three-page exercise, provision the remaining pilot identities, and explicitly enable the fenced proposal path.
    6. Open the separately approved proposal window: first love, then the remaining chosen pages; five real proposals, truthful Take/Edit/Dismiss, five forms on Curation agent MVP: what an editor does, what a curated entry contains, how the page looks, how the owner and the bots share one record, and how it stays current (under #193) #253, confirmed stop, IDs and recovery/owner-rule evidence. Complete the P6 re-audit.
    7. Finish the still-required MVP work, including the owner's P3 decisions and pin round trip, remaining pilot pages and restoration evidence, unless Curation agent MVP: what an editor does, what a curated entry contains, how the page looks, how the owner and the bots share one record, and how it stays current (under #193) #253 has explicitly been amended. A persona plan may be drafted earlier, but persona implementation waits on the governing gate.
    8. Complete the persona plan's independent audit, with the bounded prompt/run/ballot/publication/evaluation contracts, and link the evidence for every revised §3 item.

    Only GitHub write by this audit: this comment. No issue body, PR, commit, live record, service or server was changed.

  2. changed the title [-]Curation: the last stretch to the agents — #267 and #264 in one audited order, parallel builds, and "ready to build the agents" as a checkable state (under #254, #253, #193)[/-] [+]Curation: the next stretch to the agents — #267 and #264 in one audited order, deployment gates, and "ready to prepare the persona plan" (under #254, #253, #193)[/+] on Oct 26, 2026
  3. holden commented on Oct 26, 2026

    @holden
    MemberAuthor

    Revised after the independent audit, 26 October 2026

    The owner kept #253 as written: persona implementation waits for #253 §10's full MVP acceptance, including P3 and the ten pilot pages. The end state is renamed "ready to prepare the persona plan", and the title with it. The audit's six required changes, as applied:

    1. Authority (§3, §4, O1). Curation agent MVP: what an editor does, what a curated entry contains, how the page looks, how the owner and the bots share one record, and how it stays current (under #193) #253 is not amended. §3 is the narrower, honest milestone; the new §4 lists what still gates persona implementation (P3, the other five pages, the §6a demonstrations, restore from a bundle, Visit-driven curation refresh: coalesced jobs, fenced leases and atomic publication #198's reader suppression), each with an owner and when it can start. The old §5.3 ("P3 does not gate the agent phase") is withdrawn; P3 can instead run in parallel as soon as O1 is taken.
    2. Stage 0 expanded (§5). Stage 0 is now a design session that writes Curation P6: the first bot proposes into the owner's screen — attempt to proposal, Take, Edit, Dismiss, the quality form, D8 marks and the D9 attempt audit (under #254, #253) #264's resolved contract, schema, migration and test matrix: ancestry against the version-shape CHECK, retry identity, enablement versus activation (and love's standing under a new configuration version), carry-forward precedence and overflow, frozen reference reconstruction (words_sha256), lifecycle and review-unit storage, the Publish and manual-writer bypass, abstention, and a backward-compatible migration. It reconciles Curation P6: the first bot proposes into the owner's screen — attempt to proposal, Take, Edit, Dismiss, the quality form, D8 marks and the D9 attempt audit (under #254, #253) #264's "after Curation P6-0: the editor runtime made live — one model, where it lives, the v2 binding, the benchmark GO and a rehearsal against the real pages (under #254, #253, #193) #267 Delivery A merges" comment and every Curation P6-0: the editor runtime made live — one model, where it lives, the v2 binding, the benchmark GO and a rehearsal against the real pages (under #254, #253, #193) #267 worker reference, and ends with one independent check.
    3. Deployment gate (Stage 2, 3b). Merge order and deployment order are separate. Curation P6-0: the editor runtime made live — one model, where it lives, the v2 binding, the benchmark GO and a rehearsal against the real pages (under #254, #253, #193) #267 deploys on merge (no migration). Curation P6: the first bot proposes into the owner's screen — attempt to proposal, Take, Edit, Dismiss, the quality form, D8 marks and the D9 attempt audit (under #254, #253) #264 merges but is not deployed until its migration is applied on v2 on the owner's yes, because new Ecto fields would otherwise be selected before their columns exist (Published.current, published.ex:68).
    4. Health interface (§6.2). The card is SELECT-only and absent-data safe, never calling Readiness.check or Gateway.budget unchanged (both reach Gateway.service!, which inserts). Pending review units belong to Curation P6: the first bot proposes into the owner's screen — attempt to proposal, Take, Edit, Dismiss, the quality form, D8 marks and the D9 attempt audit (under #254, #253) #264's card. Stop history comes from the rehearsal and window records, not the card.
    5. Live checklist rewritten (3a–3e). Idle check, start and rebind, then register (it needs the running service), smoke (one more attempt), rehearse, stop. Configuration activation and love's re-establishment under it. The owner's three pages end published with the actual label. All five page identities provisioned (truth and coward included). Offline and no-result handling in §3 item 2. Each step on its own named yes.
    6. Evidence gates restored (§3, §10). Integrated rebased review and precommit; recovery and failure evidence; five non-abstained proposals on the full form; the P6 post-activation re-audit as a completed, satisfactory gate; checkboxes matching the new end state.

    Kept from the first version: the worker's service_unreachable refusal moves to #264 (the audit accepted it, with every #267 reference moved together); the five pages (love, war, hope, truth, coward); parallel stubbed builds, now after Stage 0.

  4. holden commented on Oct 26, 2026

    @holden
    MemberAuthor

    Independent re-audit of #268 — 2026-10-26

    Verdict: GO for Stage 0's design/reconciliation work. NOT yet GO for Stage 1 builds or live activation. Overall grade: B+, up from C+. The revised issue now represents an honest generic-editor milestone and preserves #253's full MVP gate before persona implementation. The old audit's six required topics have been addressed in the plan; the design obligations they identify have not yet been discharged in #264/#267.

    Assessed part Grade Qualification
    Plan overall B+ Sound revised direction; two P3 integration/deployment details below still need to enter Stage 0.
    §3 end state A− Correctly narrowed to “ready to prepare the persona plan”; fuller failure/recovery evidence and completed P6 re-audit now required.
    §5 order / parallelism (formerly §4) B+ Core migration-before-code sequence repaired; optional parallel P3 must respect the same deployment boundary.
    §6 scope moves/interfaces (formerly §5) B+ Worker transfer and Health contract are sound; P3's overlap is larger than its work tile.
    Nineteen corrections' sufficiency C+ pending Stage 0 Correctly acknowledged as unresolved design work; unchanged build briefs are still not executable literally.
    §7 defaults (formerly §6) B+ Gate authority and five-page provisioning fixed; O2 remains explicitly design-gated.

    This is a re-audit of the revised planning issue, not its future Stage 0 independent check and not acceptance of unbuilt code.

    Evidence and limits

    I reread the current #268 body and revision comment, compared them with my first audit, and reread the current #264/#267/#253/#254 bodies and comments. I rechecked the relevant main source, especially the proposed P3 interaction with Packet.

    Exact main remains 9ccc42748284a2610d45509aba6c8b1a157ecf46, independently matching Studio HEAD, its existing origin/main and GitHub main. Studio working tree is clean; GitHub's open-PR collection is empty. #264 and #267 have not yet been rewritten. That is expected before this revised Stage 0; it is not evidence that Stage 0 is complete.

    Fresh database checks were SELECT only, PostgreSQL 18.2, port 5434, expected cluster 7693849764459364596, with default_transaction_read_only=on on both connections:

    • v2: 0 model configs/services/attempts; 1 composition, 1 version, 3 items, 1 review, 5 composition-publication receipts; 5 proposed curator profiles and 0 profile versions. Refresh row 1 remains idle, generation 3/token 4, needs_editor/fingerprint_version; last completed 26 October 13:59:58.254863 UTC, next assessment 28 October at the same UTC time.
    • Bench: 3 configs, 1 service, 70 attempts, 210 ledger entries; no compositions/profiles/refresh rows. Config 3 remains generic-editor-v2 / qwen3.5-9b-v2 with hash 5b8b65adc7709921eb5cd06ba9d4dc1b96404e82ab04a5eb59bb709021da62cd; service available, epoch 1/fence 70, no holder. Attempts 52–70 remain 13 accepted selections and 6 abstentions.
    • Oban on v2: RefreshWorker 3 completed; CleanupWorker 673 completed; Discovery.RunWorker 67 completed + 2 cancelled; Quotations.VerifyWorker 337 completed. The cleanup count rose by one since my first snapshot; it is not a new curation proposal.
    • The authority marker still names bench; no model-service pid file or listener on 11435 was found. The version-parent and refresh-outcome CHECKs remain unchanged.

    No suite rerun: source/main did not change, and this is a planning revision. My earlier independently executed result remains 19 doctests, 2,986 tests, 0 failures, seed 86311, exit 0, at this same SHA. It is historical evidence from the first audit, not a second run. Likewise the first audit's independently reproduced packet builds and benchmark verdicts were not rerun here.

    No worktree, code, database, issue body, service or server was changed. No pull, migration, job, model call, service operation or request to 4007 was performed. The sole GitHub write for this re-audit is this comment.

    Disposition of the first audit's six required changes

    Previous requirement Revised plan Disposition
    Preserve #253's authority New header/§3 milestone and §4 restore full MVP acceptance, including P3 and ten pilot pages, before persona implementation. Resolved in the plan.
    Resolve rather than paste the nineteen corrections Stage 0 now requires the actual contract/schema/migration/test matrix, the additional conflicts, reconciled briefs and an independent check before builds. Correctly gated; design still outstanding.
    Migration before new schema reads #267 deployed first; #264 merged but withheld; backward-compatible migration from a worktree, old-code check, then #264 deployment. Core defect resolved; exact deployment-head/P3 boundary needs the clarification below.
    Read-only Health interface Explicit SELECT-only, absent-data-safe snapshot; pending reviews on #264's card; no invented stop history. Resolved as a build contract.
    Correct live activation/owner steps Start/rebind before registration; smoke counted; immutable editor binding, love's standing, actual label, three final published manual pages and all five identities. Resolved as an ordered checklist; Stage 0 must define enablement separately from activation.
    Restore review/evidence gates Integrated-head checks, failure/recovery evidence, five full forms, completed satisfactory P6 re-audit and remaining MVP ownership. Resolved in the plan.

    Answers to the original six audit questions, against the revision

    1. Is the end state appropriate and are the shared seams complete enough?

    Yes for preparing the persona plan. The revised list now includes provisioning/configuration, packet and reference validation, refresh authority, failure/recovery, reservations, the trusted human path and D8/D9. It explicitly distinguishes future persona prompt/run/ballot work from generic infrastructure, and removes human controls/no-dossier admission from the persona's own responsibilities.

    The full MVP and satisfactory P6 re-audit remain prerequisites to their stated next steps. This issue's closure is not persona implementation permission. In §4, interpret “the other five pilot pages” as an inventory against all ten acceptance rows: five proposals and five forms do not necessarily mean five published openings. A dismissed proposal is evaluation evidence, not a public opening; an unprocessed page is not an “honest nothing.”

    2. Are the scope moves sound?

    The worker transfer and revised Health boundary are sound. The worker remains model-free on main. Moving the refusal and test to #264's new editor step is correct, and the revision now requires moving every contradictory #267 reference. Health's snapshot must stay independent of mutating service creation and of the unbuilt review allocator.

    The former P3 gate waiver is withdrawn, correctly. However, the newly promoted parallel P3 path needs more integration ownership than components/opening.ex alone; see required addition R1.

    3. Are parallel builds and merge/deployment order safe?

    Conditionally yes. #264 is explicitly disabled/stubbed; both sessions have independent partitions; rebased integrated review/precommit is mandatory. The original literal shared files remain refresh_worker.ex and health_live.ex, with the former transferred wholly to #264. The revised manifests must be checked again after Stage 0.

    The two-PR deployment sequence is now coherent: #267 code without a migration, then a compatible #264 migration before #264 code. Reversing remote merge order would still not inherently break a correctly disabled/stubbed implementation, but it is outside the chosen reviewed sequence and complicates the deployment boundary.

    Optional P3 can add commits/migrations to the same main history, so “deploy #267” must identify an exact compatible commit, not simply advance to whatever GitHub main currently contains. Required addition R2 closes that remaining ambiguity.

    4. Are the nineteen corrections now sufficient?

    The issue no longer claims they are. That is the right correction. The ancestry CHECK, recaptured parent on same-generation retry, immutable lifecycle, Take uniqueness/provenance, out-of-cut kept items/cap, frozen reference reconstruction, review reservations, Publish bypass and abstention contract are now named Stage 0 deliverables.

    They are not solved by naming them. The current #264 body still asks for operations incompatible with main; its old warning comment still gives the serial dependency. Stage 0 must replace those instructions, regenerate both file lists/tests and get its independent check. This re-audit does not tick that checkbox.

    Do not force an unnecessary fixed implementation from the wording “separate published-parent from version ancestry”: the checked requirement is truthful admitted-published-parent identity, valid version lineage and a deliberately compatible CHECK/API contract. The design should state exactly which fields carry each.

    5. Are O1–O5 safe?

    • O1: safe to leave open for this narrower milestone. Taking it enables P3 design/build, subject to its own approved brief, integration and deployment gates; it does not make P3 immediately deployable.
    • O2: safe as a design-gated direction, not a completed precedence/overflow/provenance rule. Stop if it cannot be resolved within the requested slice.
    • O3: keep the verified five pages and runtime revalidation. All five identities are now explicitly provisioned. The baseline still does not cover no-lead pages or live accepted exemplars.
    • O4: correctly remains the owner's actual choice and new attributed versions, with all three manual pages ending published.
    • O5: safe after the Stage 0 check, including R1/R2 if P3 joins the parallel work. It grants no model/service/database approval.

    6. Are gates or steps still missing?

    The major prior omissions are now present. Two bounded integration requirements remain for Stage 0:

    Required additions to the design/deployment contract before Stage 1

    R1 — P3 must integrate the complete source-record-work path, not only the public tile.

    #254 P3 explicitly adds newly pinned source-record works to Candidates. On audited main, Packet's work path still assumes catalog works:

    Consequently, a new Candidates work containing a source-record revision without a catalog reference cannot simply flow through the present Packet implementation: construction can raise on missing catalog, and the probe cannot reconstruct the intended source-record reference. Refresh.inputs itself builds Packet, so this affects freshness as well as eventual proposals.

    This is a static integration finding, not a live reproduction: I did not create a pin or run new code. It does not block Stage 0 or the two generic build sessions if P3 remains isolated.

    Add to Stage 0/P3's dedicated brief: owners and tests for Candidates, Packet build/freeze/verify/probe/summary, proposal reference reconstruction, PublishedReader and the work tile, plus the pin-revision/retirement effects on freshness and D8. Either deliver these compatibly together before pins become offerable, or keep that new offering gated until they are ready. This is integration of the existing P3 requirement, not permission for a new matcher or cut-policy change.

    R2 — pin deployment commits and gate the whole deployed history.

    Record the exact #267 deployment SHA/schema prerequisites. While #264 is merged but held back, any later P3 commit on main may be its descendant. Deploying that P3 head by advancing the shared checkout would also deploy #264 before Stage 3b.

    Therefore specify: no deployment of a head containing #264 until #264's compatible migration is complete; P3 has its own schema/code compatibility and owner-approval gate; final integration checks cover the actual head being deployed. Parallel build permission is not independent deployment permission on one linear main checkout.

    This does not require changing the chosen #267 → #264 order. It makes its existing “DO NOT deploy” rule enforceable when the optional third build is used.

    Recommendations, not new blockers

    • Assign the still-unassigned Visit-driven curation refresh: coalesced jobs, fenced leases and atomic publication #198 reader-suppression delivery and bundle/restore check early. The revised closure rule appropriately requires each remaining item to acquire an owner/issue or recorded decision.
    • Preserve the revised separation between admitting/activating the immutable editor configuration and enabling the proposal path. Stage 0 must define the off state and ensure a visit during the owner's manual exercise cannot start an unapproved model call.
    • Use an explicit ten-page acceptance inventory rather than assuming the five baseline proposals all become publications.
    • Keep the original runtime GO's limitations and measured cut-v1 baseline; P3 integration or personas must not silently turn that old measurement into proof of a new task.

    Ranked remaining risks and smallest safe sequence

    1. High at integration — source-record pins offered to catalog-only Packet/proposal code, stranding refresh/proposal work.
    2. High at deployment — advancing main past the held Curation P6: the first bot proposes into the owner's screen — attempt to proposal, Take, Edit, Dismiss, the quality form, D8 marks and the D9 attempt audit (under #254, #253) #264 commit through a P3 deployment, exposing unmigrated schema reads.
    3. High until designed/tested — the original refresh-authority, ancestry, attribution and review-reservation conflicts. These are now honestly gated rather than concealed by the plan.
    4. Medium — confusing completed generic evaluation with full MVP or persona readiness. The revised scope substantially reduces this risk.

    Smallest safe next action: start Stage 0 now. Resolve and rewrite #264/#267, include R1/R2 if P3 proceeds alongside them, and obtain the independent check of those actual resolved briefs. Then follow the revised disabled parallel builds → integrated review → pinned #267 deployment/window → approved compatible #264 migration → #264 code → separately approved configuration/manual/proposal steps → five forms → satisfactory P6 re-audit. Full MVP acceptance still precedes persona implementation.

    No additional suite, model run or live activation is needed to start that design work.

  5. holden commented on Oct 26, 2026

    @holden
    MemberAuthor

    Stage 0: output ready for its independent check (26 October 2026)

    Stage 0's design session has finished. #264's body is now the resolved contract (§4 rules R1–R19, §5 schema and migration table, §7 regenerated files, §8 a 36-test matrix each tied to its rule, §10 Delivery B as Stages 3b–3e), and #267's body is reconciled with §6. A Stage 0 comment on each lists what changed and the audit point each change answers (#264's, which also quotes every replaced section verbatim; #267's). Nothing was built, started or written beyond those two bodies and three comments. No build session should start until the independent check is posted and its required changes are applied.

    Basis. main at 9ccc427 (Studio checkout and origin/main agree; no #264 or #267 worktree or branch exists). The v2 catalog read with default_transaction_read_only=on: every curation CHECK, unique index and trigger body named in #264 §3 and §5 was read from pg_constraint, pg_indexes and pg_proc, and the rows from SELECTs; bench attempt 59 confirms Packet.summary/1 stores excerpt_sha256 and no words_sha256. Not done: no test suite (a design session), no packet build (love's cut counts are the audits'), no process or listener check.

    What Stage 0 settled, in one line each (the rules are in #264 §4):

    • Ancestry. parent_version_id stays version history (one line; the shape CHECK untouched); the published parent captured at admission lives on the new curation_proposals row, null when nothing was published; Take's parent is the proposal.
    • Retry identity. Key proposal:<composition>:<configuration version>:<generation>; the parent, model config and packet hash are frozen per generation; a deferred retry keeps its generation only if the pointer has not moved and its attempt is absent or answered (changing refresh.ex:364); drift or an answerless attempt opens a new generation.
    • Enablement versus activation. Binding in an immutable configuration version (ready?/1 named), activation as today (love's version stops standing and is re-published in the same sitting), enablement a separate switch that a new activation invalidates.
    • Carry-forward (O2). Kept items travel as verified copies with the owner's signature; kept first, the model fills the rest, overflow recorded; marks read at the write; dedup on the merged arrangement.
    • Frozen references. Resolution through the frozen packet with hash equality; words_sha256 added to the summary; drift refused.
    • Storage. A proposal lifecycle table (versions stay immutable); the honest unchanged-Take rule (the dedup index partial on manual versions); review units reserved on the refresh row before inference and released by every exit, exactly once by construction; one proposal per attempt.
    • No bypass. A model version is never accepted or published (code and DB); Publish never holds one; the ordinary form never loads a model note.
    • Abstention. New outcomes proposed, abstained, no_change with the enum, CHECK and disclosure changed together; 48-hour window.
    • Compatibility. Every migration change additive or widening; new restrictions only on rows the old code never writes; rollback rules stated.
    • §6. service_unreachable and its test moved to Curation P6: the first bot proposes into the owner's screen — attempt to proposal, Take, Edit, Dismiss, the quality form, D8 marks and the D9 attempt audit (under #254, #253) #264; the Health card SELECT-only and absent-data safe, inference minutes only, no stop history; the live order fixed; the shared health_live.ex interface agreed (three lines per issue). After regeneration, health_live.ex is the only file both briefs touch.

    The re-audit's required additions. R2 (deployment pinned to commits) is in #264 §10 3b and §11 and #267 §9: #267 is deployed by fast-forwarding the Studio checkout to its recorded merge commit; no head containing #264 is deployed before 3b; a later P3 commit has its own compatibility check and yes, and waits for 3b if it descends from #264's merge; integrated checks run on the exact head deployed. R1 (P3's source-record works) is recorded for P3's future brief, with #264's side fixed: #264's writer accepts only the catalog work shape Packet builds on main and refuses any other (:unsupported_work_reference), so a pin cannot reach a proposal until P3 extends Candidates, Packet (build, freeze, verify, probe, summary), the proposal reconstruction, PublishedReader and the tile together, with the pin's revision and retirement effects on freshness and D8.

    For the check, and for a future P3 issue: the components/opening.ex boundary. #264 edits only the disclosure (about/1, assessed/1, the selected-by and attempt lines). The highlight tiles (highlight/1 for works, work_title/1) and the source-record work path above are P3's; #264 does not touch them. The re-audit's ten-page acceptance inventory (§4) is #268's, not #264's: five proposals and five forms are not five published openings.

    Timing note on 3c-5. "Enable the fenced proposal path" stays step 5, but #264 takes it at the start of 3d's window, after the service is up, so that no visit meets an enabled editor while the service is down.

    Housekeeping left to the owner. #254's P6-0 and P6 checklist lines still say the corrections are carried by #267 (Stage 0 may not edit #254). #264's earlier comment was left unedited; the new comment replaces its rule explicitly. #264's body is about 62,600 characters of GitHub's 65,536, so the check's changes there need to stay concise.

    Owner decisions Stage 0 could not settle from the specification

    Each has the default #264 uses (marked OD-n in its body) and what follows from it.

    Decision Default Consequence
    OD-1 Storage beyond #253 §6's list. §6 says the record needs origin/attempt_id, the freshness table and the D8 marks, "nothing else". #253's own rules (§6a dismissal and supersession, §7's 30 review units, §10's per-item form, §5a step 5's attribution on Take) and #268's separate enable step need more. Add the minimum in #264 §5: curation_proposals (lifecycle and form), copied_from_item_id on items, review_unit_reserved_at on refresh rows, the enable columns on the configuration, kind on profiles, three outcome values. A blocker if refused: P6 stops at its migration. Without these the record cannot keep versions immutable and still record Take, Dismiss and obsolescence, count review units, or copy owner and model words truthfully.
    OD-2 The editor's bot actor needs a sources row (a bot actor must have a source, one bot per source), and the home page prints the count of all sources ("27 sources so far"). Source generic-editor, kind bot, inactive; #264 excludes kind: :bot from the home count (one filter in home_live.ex). Readers see no change. The ops Health tables gain one zero row. The alternative is "28 sources so far".
    OD-3 Love's opening disappears at activation. Activating the editor's configuration version makes love's published version stop standing (approval never transfers), until the owner re-publishes it under the new version. Activate and re-publish in one sitting (3c-2). A few minutes of honest-empty on /on/love. Avoiding it would need a carry-over feature #253 does not authorize.
    OD-4 The owner's public label. No seam exists to change an account actor's label (it is set to "Account #N" at first use); #253 §10 accepts "Account #2". Keep "Account #2", unless the owner asks now for a narrow seam: a reviewer-only task changing only the signed-in account's label (old notes keep the label they were signed with). Asking after #264's Delivery A reopens it.
    OD-5 "Request a proposal." A page just checked admits nothing for 48 hours, so 3d cannot go page by page without an operator control. Build it (#264 R18): reviewer-only, shown only while enabled; marks the page dirty, lifts its backoff, admits as a visit would. If refused, 3d waits on windows and input changes; love, re-published at 3c-2, could wait two days.
    OD-6 The proposal rules O2 leaves open. (a) Precedence: kept items first, the model fills the remaining places, the rest recorded as overflow. (b) A kept item that is ineligible or no longer on the published version is not carried, and the proposal says so. (c) A newer proposal makes the older pending one obsolete; a pending proposal from the same packet, marks and parent skips the call. (d) Take only while the proposal is the page's newest version (Dismiss with the form stays). (e) A validation-refused answer is a failed inference (backoff, new generation), not an assessment. As listed. Changing (a)–(d) changes #264's tests 17–20 and 24; (e) follows #253 §7 ("a failed inference never advances last_completed_at").
    OD-7 The public words for an editor's assessment in the disclosure. "Judged by the editor: it proposed a change, which waits for the owner's review"; "…the candidates were too thin to choose from, so it proposed nothing"; "…nothing to change". Wording only. It says publicly that a proposal is waiting.
    OD-8 Between proposal windows: disable the editor, or leave it enabled. Disable at the end of each window, enable at the start of the next. Disabled, pages between windows complete as P5 checks ("checked, not judged"). Enabled, due pages defer service_unreachable with backoff and fill the Stale tab. Neither proposes between windows.
    OD-9 A page that abstains does not count. Which page replaces it? Justice (one admissible lead; 7 of 9 candidates in the audit), then patience, then nepotism, each provisioned on the owner's yes. Changes O3's list for the baseline only. Truth and coward have the thinnest packets (7 of 16, 6 of 6).

    Next: the independent check of #264 and #267 as now written.

  6. holden commented on Oct 27, 2026

    @holden
    MemberAuthor

    Independent check of Stage 0: #264 and #267 against main and the v2 catalog (27 October 2026)

    This is the one independent check #268 §5 requires before any build (#269's D1). I did not write Stage 0. I checked the rewritten #264 and the reconciled #267 against the code and the catalog, not against their own prose.

    Verdict: GO for Stage 1 builds, once required changes 1–6 below are applied to the two bodies as written. These are text changes only. Applied verbatim they need no further check; anything beyond them does. Not GO for any live step: 3a–3e each keep their own owner's yes. Overall grade: B+ (A− with the changes). The contract is accurate about main almost line by line, and its migration is backward compatible. But two of its rules would fail in use. "Request a proposal" never reaches the editor on a page that has already been checked, which would stall 3d. And once any proposal exists, the owner's Publish of their own unchanged version is refused.

    Part Grade Why
    #264 §3 seams, line references and catalog claims A Every function, line, CHECK, trigger and index named exists as described. One slip: R14.5's LeadPolicy.statement/2 (it is statement/1).
    §5 migration, R19 compatibility A− Every row the old code writes still passes. R19.2 says "git pull", which contradicts R2 (required 3).
    R3–R6 hand-off, retry identity, review units B+ Generations, keys and reservations neither strand nor duplicate work. R18's request is a no-op on a checked page (required 1).
    R7 frozen references A Hash equality refuses drift. words_sha256 is where Stage 0 says it is.
    R8 carry-forward (O2) B+ Fits one lead and three highlights on the pilot pages. A merged arrangement that fails the lead rule is unmapped (required 4).
    R10–R13 lifecycle and no bypass B The DB guards are the right choke points. Publish's held path blocks the owner (required 2).
    R17 / #267 §3.4 Health interface B+ Agreed, SELECT-only and absent-data safe in both bodies. #267 claims a reviewer gate /ops/health does not have (required 5).
    #267 reconciliation A− It no longer owns the worker's refusal anywhere.
    Live steps vs #268 3a–3e, re-audit R1/R2 A− In order. R1 and R2 are present in both bodies.
    OD-1..OD-9 defaults B+ All safe. OD-5 works only with required change 1, and OD-1 should also cover the §6a parent reading (required 6).

    Evidence and limits

    • SHA: the Studio checkout /Volumes/LLM Models/dictionary/src/dictionary is on main at 9ccc42748284a2610d45509aba6c8b1a157ecf46, equal to origin/main, with a clean working tree.
    • Cluster: 7693849764459364596, PostgreSQL 18.2, port 5434. Every connection to devils_dictionary_v2 and devils_dictionary_runtime_bench ran with PGOPTIONS='-c default_transaction_read_only=on', confirmed on.
    • Catalog reads: pg_constraint, pg_indexes and pg_trigger for the 14 tables §3 and §5 rely on, and pg_get_functiondef for the 14 functions behind their triggers. Also \d of the configuration, source and refresh tables, and the columns of curator_profile_versions, inference_attempts, inference_ledger_entries and inference_services.
    • Rows: v2 is unchanged since 26 October and matches Curation P6: the first bot proposes into the owner's screen — attempt to proposal, Take, Edit, Dismiss, the quality form, D8 marks and the D9 attempt audit (under #254, #253) #264 §3. It has 1 composition, 1 version and 3 items (all "Account Bump image_processing from 1.12.2 to 1.13.0 #2", actor 5; item 3 is sense revision 248575, quotation:3), 1 review, 5 receipts, 1 activation, 27 sources (none bot), 5 profiles and 0 profile versions, and 0 model configs, services and attempts. Refresh row 1 is idle, generation 3, token 4, needs_editor/fingerprint_version, due 28 Oct 13:59:58 UTC. Oban has RefreshWorker 3 completed. Bench attempt 59's summary keeps a sense quotation's excerpt_sha256 and no words_sha256.
    • Code read at that SHA: refresh.ex (admission, claim, check, outcome/3, complete, fail, fences, release, defer, triggers, suppress, reopen, published, counts) and refresh_worker.ex. Also configurations.ex, compositions.ex (create_version/3, not_duplicate/4, note/2, one_display_identity/1), curation_live/editor.ex (held/2, publishable/3, write_version/2), curation_live.ex record/1, packet.ex (candidate, freeze/1, summary/1), gateway.ex (service!/1, admit/3, budget/2), readiness.ex, runtime.ex, endpoint.ex, publications.ex, reviews.ex, standing.ex, lead_policy.ex, health_live.ex and its test, home_live.ex, sources.ex/source.ex, router.ex, proxy_guard.ex, installation/bundle.ex, dd.runtime.ex, pilot.ex, config/config.exs (Oban) and config/dev.exs (DD_DATABASE).
    • Not run: I did not run the test suite. main has not moved since the first audit's independent run at this SHA (2,986 tests, 0 failures, seed 86311), and that result is historical, not repeated here; this check is of a design. I built no packets: love's 10 of 32 and the other cut counts are the audits'. I executed no Elixir, mix task or worktree, made no request to 4007, and started no service. The only GitHub write is this comment.

    What holds, verified against main and v2

    Required changes (to the bodies, before any build session starts)

    Characters: #264's body is 62,621 of 65,536; changes 1–4 and 6 add about 1,280 net, leaving it near 63,900 (about 1,600 spare). Each block below is the exact text.

    1. #264 R18 and test 34: a request must reach the editor.
    outcome/3 answers {:unchanged, "unchanged"} whenever the stored fingerprint equals the new one, whatever the pending reason (refresh.ex:613–620), and R3.1 completes an unchanged outcome "exactly as today". Every page in 3d will normally have a stored fingerprint by then: love from the first visit after its 3c re-publication, and war, hope, truth and coward from their first nothing_accepted check after 3c-4. So request --slug love would complete unchanged and call nothing.

    R18, after "then admits as a visit would (admit/2, every admission rule applying)." insert:

     `outcome/3` assesses a claim whose `pending_reason` is `requested` as `{:needs_editor, "requested"}`, as it assesses a periodic run `due`; otherwise a page whose fingerprint is unchanged since its last check completes `unchanged` (`refresh.ex:613–620`) and the editor is never called.
    

    Test 34, replace "admits one run; a second request" with:

    admits one run that reaches `Proposals.step/3` although the inputs are unchanged since the last check; a second request
    

    2. #264 R13.2 and test 28: a proposal must not block the owner's own Publish.
    publishable/3 (editor.ex:530–535) publishes the held version only while it is Compositions.latest_version/1, the newest of any origin. Once any proposal exists, pending or dismissed (versions are immutable), the owner's held manual version is refused {:stale_parent, proposal}, and Save of the same arrangement is {:duplicate_version, id} (not_duplicate/4). The owner could not re-publish an unchanged version, for instance after a withdrawal and re-open (#269 W5).

    R13.2, after "while Save's :expected_parent stays the newest version of any origin." insert:

     Publish's held path (`publishable/3`, `editor.ex:531`) compares the held version with the newest **manual** version: a newer model version, pending or resolved, never makes it stale.
    

    Test 28, replace "Publish never publishes the proposal," with:

    Publish never publishes the proposal but still publishes the held manual version (also after a Dismiss),
    

    3. #264 R19.2: deployment is the pinned fast-forward.
    Replace "so deploying is a git pull and one request, with no restart." with:

    so deploying is §10 3b's fast-forward to the recorded SHA and one request, with no restart (never a `git pull`: §11).
    

    4. #264 R6, R7.6 and R8.3: a merged arrangement the checks refuse.
    R7.6 checks the merged arrangement against LeadRule, the bound and one_display_identity/1, but R6 maps only an ineligible chosen item. Two cases slip through. A carried lead may no longer be the admissible lead. Or, on a fallback-lead page, the model's lead may equal a carried highlight, leaving no lead. Both repeat on every retry (same packet, marks and fixed-seed answer), so mapping them as "dirty" would re-run inference on every visit. None of the five pilot pages reaches them today, but the build must not improvise.

    R6, a new row after the drift row:

    | accepted, merged arrangement refused by `LeadRule`, the bound or `one_display_identity/1` (R7.6) | complete `no_change` / `merge_refused` | advances, +48 h | released | the attempt only |
    

    R7.6, after "A chosen item that is ineligible now closes the generation (R6)." add:

     A merged arrangement refused for any other reason completes `merge_refused` (R6).
    

    R8.3, replace "the model's highlights fill the remaining places in the model's order" with:

    the model's highlights fill the remaining places, up to the configuration version's `max_highlights`, in the model's order
    

    5. #267 §3.4 and test 4: /ops/health has no reviewer gate.
    On main, /ops/health is in the :operator live session with ProxyGuard only (router.ex:115–117), and health_live_test.exs mounts it signed out. Making it reviewer-only would be a router change to an existing page that #267's §7 does not list.

    §3.4, replace "Reviewer-only, 404 through the tunnel like the rest of /ops." with:

    404 through the tunnel like the rest of `/ops` (`ProxyGuard`); `/ops/health` has no reviewer gate on `main` and this issue adds none.
    

    §8 test 4, replace "it is 404 through the tunnel and requires a reviewer." with:

    it is 404 through the tunnel.
    

    6. #264 §5: name the §6a parent reading as part of OD-1.
    #253 §6a says the bot writes a proposal "with parent_version_id N", N being the published version at admission. R9.1 keeps parent_version_id as lineage and puts N on the proposal row. That is the right design, because editorial_composition_versions_shape requires a parent after version 1 even when nothing is published. But #253 governs and is not amended, so the owner should take this reading explicitly.

    §5's first paragraph, after "the rest below is what §4's rules need, and is OD-1 on #268." add:

     So is R9's reading of #253 §6a's "`parent_version_id` N": N (the published version at admission) is `curation_proposals.published_parent_version_id`, and `parent_version_id` stays lineage, because the shape CHECK needs a parent after version 1 even when nothing is published.
    

    The owner decisions, OD-1 to OD-9

    Recommendations (not required)

    1. Lock order. Publications.publish/3 and withdraw/3 lock the composition and then update refresh rows (Refresh.published/2, suppress/1), while P5's complete/3 locks the refresh row first. The step's write takes both, so state "composition first, then the refresh row" in R9. Otherwise a publication landing during the write can deadlock: one side aborts, and nothing is lost.
    2. §5, curation_configurations trigger: fire only when editor_enabled_version_id changes, so activate/4 (which leaves it in place, R1.5) is untouched.
    3. Take and the partial dedup index. A proposal equal to an earlier taken arrangement, under the same configuration version and fingerprint (one the owner later replaced), would make Take hit editorial_composition_versions_dedup_index. Either add a matches_manual case to R8.4 or name the refusal.
    4. R14.5 should name a new clause of LeadPolicy.statement/1; there is no statement/2.
    5. Curation P6-0: the editor runtime made live — one model, where it lives, the v2 binding, the benchmark GO and a rehearsal against the real pages (under #254, #253, #193) #267 §3.2: the first ready --smoke on v2 also inserts the import actor "Curation runtime operator (CLI)" (dd.runtime.ex:414). v2 has none today, so list it.
    6. R19.3: say that "revert the code" means moving the Studio checkout back to Curation P6-0: the editor runtime made live — one model, where it lives, the v2 binding, the benchmark GO and a rehearsal against the real pages (under #254, #253, #193) #267's recorded SHA, not a revert commit deployed by fast-forward.

    Applying changes 1–6 to the bodies and recording each owner decision on this issue is the driver's work under #269. This check does not tick #268's Stage 0 box itself.

  7. holden commented on Oct 27, 2026

    @holden
    MemberAuthor

    Stage 0 check: required changes applied to #264 and #267 (27 October 2026)

    The check's required changes 1–6 are applied verbatim, with no other text changed:

    The six recommendations are not applied. Each stays on the check comment for the build sessions and their reviewers.

    The Stage 0 box in §10 is left unticked for the owner. The owner's decisions OD-1 to OD-9 come next. Each will be recorded here individually, with the check's notes on OD-1 (it also covers change 6), OD-2 (/sources/generic-editor would become a public page, and the bot source would feed mix dd.export.replay) and OD-5 (it works only with change 1).

  8. holden commented on Oct 27, 2026

    @holden
    MemberAuthor

    Owner decisions OD-1 to OD-9, recorded 27 October 2026

    The owner answered each decision separately, in chat with the #269 driver session, after the Stage 0 check and its applied changes. Each question carried the check's notes. None is a blanket acceptance of the defaults. #253 is not amended by any of them.

    Decision The owner's answer
    OD-1 Storage beyond #253 §6's list Accepted, with R9's reading of §6a. #264 §5's additions stand: curation_proposals, copied_from_item_id, review_unit_reserved_at, the enable columns, profile kind and three outcome values. #253 §6a's "parent_version_id N" is stored as curation_proposals.published_parent_version_id, and parent_version_id stays lineage (the check's change 6).
    OD-2 The editor's bot source Default, with its effects recorded. Source generic-editor: kind bot, inactive. The home page's count excludes kind: :bot. Accepted as known effects, with no code change: /sources/generic-editor becomes a public page, and the source is included in mix dd.export.replay.
    OD-3 Love's opening at activation Activate and re-publish in one sitting (3c-2). A few minutes with no opening on /on/love is accepted.
    OD-4 The owner's public label (#269 D4) Keep "Account #2". No label seam is built.
    OD-5 "Request a proposal" (R18) Build it, as R18 reads with the check's change 1.
    OD-6 The proposal rules (a)–(e) Accepted as listed, as R8 and R10–R11 read now, including change 4's merge_refused row and the max_highlights cap.
    OD-7 The disclosure's words for an assessment Accepted: "judged by the editor: it proposed a change, which waits for the owner's review"; "…the candidates were too thin to choose from, so it proposed nothing"; "…nothing to change".
    OD-8 Between proposal windows Disable at the end of each window and enable at the start of the next. Between windows a check completes as P5's does, and the next window uses Request or the 48-hour window.
    OD-9 A replacement for an abstaining page Justice, then patience, then nepotism. Each is provisioned on its own owner's yes when needed.

    With D1 applied and D2 recorded, #269 W1 (#267 Delivery A, partition _cp60) and W2 (#264 Delivery A, partition _cp6) may be proposed as build sessions. Neither authorizes a live step.

  9. holden commented on Oct 27, 2026

    @holden
    MemberAuthor

    O1 taken (27 October 2026). The owner decided P3's two rules, recorded on #269 as D3 (comment). First, the pin is the poem-to-page relation. Second, the owner declares a successor: a changed PoetryDB text withholds the published item :revision_superseded only once the owner declares it the successor, until a version pinning the new text is published. #253 is unchanged.

  10. holden commented on Nov 1, 2026

    @holden
    MemberAuthor

    Closed 1 November 2026 into #298. §3 items 1 and 5 are proven, 2 is tested for crash and offline and partly observed live, 3 and 4 were superseded by the 29 October amendment, 6 is satisfied by the later audits. Its end state, 'ready to prepare the persona plan', is overtaken by #275's single-curator work and now by #298 §5 C4–C5.

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