Repository navigation
Description
Reconciled 25 October 2026 with the independent assessment below (all fifteen section C edits, accepted by the owner) and #253's new D8 and D9. The list of changes is the reconciliation comment on this issue.
Implementation plan for #253 (the curation agent MVP), written 21 October 2026. Seven parts, each one PR, each with a brief an implementing agent can take without this conversation. #253 is the specification and wins on any conflict; this issue says what to build, in what order, in which files, and what "done" is for each part. Tick a part here when its PR merges and link the PR.
Ground rules for every part
- Repository
razrfly/dictionary, checkout/Volumes/LLM Models/dictionary/src/dictionary, branch frommain. Elixir 1.19 / Phoenix 1.8 / LiveView 1.2, Postgres 18 on port 5434, databasedevils_dictionary_v2. - A fresh worktree needs symlinks for
deps,dataand eachpriv/replay/*.jsonl.gzfrom the main checkout, orSources.ManifestTestfails. - Precommit:
MIX_ENV=test MIX_TEST_PARTITION=_<part> DD_NO_OBAN=1 mix precommit. Green is the bar; report the test count. - The dev server on port 4007 is also the public site through ngrok. Never restart it, never edit its launch script, never run a migration against v2 without the owner's word in chat.
gh pr mergeis refused by the permission classifier. Stop at a reviewed PR (CodeRabbit addressed, one independent review) and hand the owner the merge.- Build only the part. The owner treats anything beyond it as scope creep. If the part cannot be done as written, say why on the PR rather than widening it.
- Every PR: a before/after on a real page (pilot page from 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 D2), with the row ids it created.
Order and parallelism
P1 ──┬── P2 (screen) ───────────────┐
├── P3 (pin, PoetryDB) │
├── P4 (packet kinds) ──────────┼── P6 (proposal path)
└── P5 (freshness) ─────────────┤
P7 (runtime fixes, benchmark) ───────┘ P6's live inference needs P7's go
P1 first and alone. Then P2, P3, P4 and P7 in parallel: they touch disjoint code. P4 does not wait for P3: it can use existing catalog artwork fixtures. P5 after P1. P6 last, and its live inference needs P7's revised-instruction benchmark go.
Found while writing P2 (23 Oct): the registry already holds 2,903 PoetryDB poems, 946 Open Library works and 1,644 Met artworks as committed catalog manifests (priv/artworks/manifests/), and Curation.Eligibility admits a catalog-only work as an item. Corrected 25 Oct: catalog references can be stored/rendered, but P2 lists only page-matched artwork rows; poems/books need an explicit page-selection relation (#253 D6). P4 can use existing catalog artwork fixtures without waiting for P3, but must not claim poems became offerable merely because manifests exist. Reading the spec: #253 §2 (rules), §3 (the record and what it can offer today), §5a (the editor), §6 (the screen), §6a (owner versus bot), §7 (freshness), §9 (gaps), §12 (defaults, D1–D9).
P1 — Seed, activate, the page reads published compositions
Spec: #253 §4b, §8, §9 items 1–2, §11. Reopen #156 for it.
Read first: docs/curation/persistence-slice-1.md; lib/devils_dictionary/curation/{opening.ex,opening_reader.ex,published.ex,manual_fixture.ex,configurations.ex,compositions.ex,publications.ex,eligibility.ex}; lib/devils_dictionary_web/live/word_live.ex (the opening is read near lines 171 and 248); lib/devils_dictionary_web/components/opening.ex; lib/mix/tasks/dd.curation.seed.ex.
Build:
- A published reader (
Curation.PublishedReaderorCuration.Publishedextended) implementingCuration.OpeningReader: for the page's scope and the resolved default configuration, readPublished.current/1, runEligibility.evaluate/5, and map onto%Curation.Opening{}withorigin: :published: lead with its policy and register, highlights in position order with their notes as reasons and theirnote_author_*as the signature,configuration,review(selected by, published by, when),withheldwith reasons. Nothing fetched, nothing written. Curation.Opening.reader/1returns the published reader by default; the fixture reader only when fixtures are enabled and?opening=fixtureis asked, exactly as now.WordLivecalls the reader on every request. An absent configuration, an absent composition, or an empty opening renders no section.- The disclosure ("How this was curated") shows: configuration slug and version, selected by (label and kind), published by and when. Reserve, but leave empty, the four assessment dates P5 adds.
Tests: no composition → no section; a published composition → lead and highlights with reasons and signatures; withdrawn → nothing; the fixture gate still works in dev and never in prod config; a lead that breaks bierce_first_v1 is withheld, not substituted; an item whose revision was superseded is withheld with :revision_superseded.
On v2, after the owner says so in chat: mix dd.curation.seed, then Curation.Configurations.activate/4 on version 1 of global-default as the reviewer account the owner names (users.id 2 was the routing reviewer). Record before/after counts of curation_configurations, curation_configuration_versions, curator_profiles. Zero compositions is the expected end state.
Done: reviewed PR; configuration active on v2; a composition published by hand in iex on one pilot page appears on that page through the tunnel and disappears on withdrawal.
Reconciled after #256 (25 Oct): language comes from the scope's lexemes, not page.language_tag; scopes are exact sorted lexeme sets, overlapping lexical-page scopes cannot coexist; Published owns eligibility, the reader maps its result; assessment rows are absent until P5; definition highlights have their own tile; actor labels come from the record. The approved typography change and the actual live five-lexeme love scope stand, not the two-lexeme fixture. P1's temporary work/exemplar withholding (:unsupported_item) was repaired by P2 and is resolved. P1 is not reopened.
P2 — The owner's curation screen
Spec: #253 §6, §6a (withdraw means keep off; re-open), §12 D2–D3. Operator-only, under /ops, 404 through the tunnel like the other operator pages (ProxyGuard).
Read first: lib/devils_dictionary_web/router.ex (the /ops live session), an existing operator LiveView under lib/devils_dictionary_web/live/ops* for the conventions, Curation.Compositions (provision/3, create_version/3, items/1, latest_version/1), Curation.Reviews.decide/4, Curation.Publications.{publish/3,withdraw/3}, Curation.LeadPolicy, Curation.Eligibility, Curation.Authority.actor/2, and Curation.ManualFixture for how a selection is expressed as references.
Build:
/ops/curation: the queue. Tabs: pilot pages (D2 allowlist: a configuration-overridable code default,CurationLive.pilot_pages/0, not a config-file edit), published, withheld items, and (reserved for P5/P6) stale and proposed. Each row: word, state, version, counts, a button./ops/curation/on/:slug: the editor. Lead radio (the page's applicable Bierce entries first, then other definitions, then "no lead"), with the lead rule's verdict shown. Highlight checkboxes (max three) over the eligible candidates of the four kinds the record admits today (content,sense_quotation,workfrom page-matched catalog artwork rows,exemplarfrom acceptedillustratesclaims; no poems or books, 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 D6), each with a "why" box of 280 characters. Save draft (a new version), Publish (review:acceptedthen publish, atomically), Withdraw current (with reason), Re-open (clears P5'ssuppressedwhen it exists; no-op now). History: list versions with parent links.- Candidate listing reuses the page build (
WordLive's page struct) so the lead rule sees the same cards the reader sees. - Exemplar and work highlights render on the page. P1 (Curation P1 (#255): one published opening on /on/love #256) withholds
workandexemplaritems as:unsupported_itembecause its demonstration pinned neither. The screen must not offer a highlight the page cannot draw, so P2 adds the two tiles toDevilsDictionaryWeb.Opening: an exemplar (an acceptedillustratesclaim with its subject label and the reason) and a catalog work (title, creator, year, licence, image where the catalog has one; a poem tile shows title, poet, length and link, no lines), mapped by the published reader fromPublished.current/1'ssubjectand catalog fields. This is what makes a page carry representative examples (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 §3), and it is a named requirement here so it is not deferred twice. - Scopes are the On page's. The screen provisions
:lexical_pagescopes from the page build's headword lexeme ids (the live love reaches noun, verb and a name). The record refuses overlapping scopes, so an exact-word page cannot get its own composition beside the On page's. The screen says so in these words: "Exact-word pages overlap this editing scope and cannot have a separate active lexical-page composition; they do not currently display this composition." (The built label still says "share this"; its one-string correction rides with P5, which already changes the editor.) Actual sharing is a product and routing decision, not implied by the scope rule.
Tests: an exemplar highlight and a work highlight render with their reasons and nothing withheld; the screen is 404 proxied and 200 local; a non-reviewer gets 403; publishing writes a review and a publication and the page shows it; withdrawing removes it; a lead that breaks the rule cannot be saved; a note is attributed to the logged-in human and never to a caller-supplied author; two sessions publishing from the same previous version: one wins, one gets {:stale_pointer, _} (a draft from a moved parent gets {:stale_parent, _}).
Done: the owner curates three pilot pages by hand through the tunnel-free local screen, withdraws and re-publishes one, with ids on the PR.
Status: P2 Delivery A merged at e25dd94; owner demonstration/Delivery B not evidenced. Departures recorded at merge (#258): poem tiles show title/poet/length/link, no lines; no editor height readout; pilot slugs are a configuration-overridable code default, not a new config-file edit; editor URL is /ops/curation/on/:slug; stale publication errors are stale_pointer/stale_parent. Exemplars are accepted nominations only, never instances. Heights are recorded as heading start–heading end: love 882–910 px after P1 (#256, accepted) and 878–906 px after P2 (#258); the heading starts within 900 px, the whole heading does not.
P3 — The pin: a PoetryDB poem becomes a durable work
Spec: #253 §9.6 in full, D5 (highlight-only), D6 (PoetryDB, and the poem-to-page rule the owner decides in the P3/P4 handoff). Reopen nothing; cite #164's creator-identity rule.
Identity, corrected before implementation: PoetryDB has no provider-issued poem ID. Reuse the existing poetrydb_poem namespace and Poetrydb.poem_id/3; it hashes author, title and lines. Changed lines currently mean a changed identity, not a new revision under the same identifier. Decide the permitted supersession relationship explicitly; never deduplicate by title. These are repairs to the pin requirement, not grounds for a broader ingestion project.
Read first: lib/devils_dictionary/discovery/{result.ex,providers.ex,provider.ex} and lib/devils_dictionary/discovery/providers/poetrydb.ex; docs/integrations/source-identity.md; lib/devils_dictionary/registry/{object.ex,external_identifier.ex,content_item.ex,content_revision.ex}; lib/devils_dictionary/sources/{source.ex,source_record.ex} and source_record_revisions; how Curation.Eligibility checks a work item (catalog pin or source-record revision that materialized the object); how #164 mints a creator by identifier.
Build:
Curation.Pins.pin(scope, result_id)(or underDiscovery): from a cached PoetryDB result, first check through the existing identity adapter (Poetrydb.identity_record/1) whether the poem is already materialized; otherwise upsert a source record under thepoetrydbsource keyed by itspoetrydb_poemidentity, with a revision holding title, author, line count and lines (PoetryDB is public domain; retention is permitted and recorded per provider in the registry's policy, not per pin). Resolve the author through the reviewed author-QID crosswalk thatArtworks.Corpus.Poetrydbwrites (Creator identity for every provider: a result credits a person by identifier, the pipeline resolves or mints the person, the card links the entry — built once, before any durable quotation is written (build 2 of #158) #164's rule); an unresolved poet does not acquire an invented identifier. Find or mint theworkobject by itspoetrydb_poemidentity (never by title). Materialize so thatEligibility's "a source-record revision must be one whose record materialized the item's object" holds.- Record licence string and fetch time on the revision;
display_allowedfrom the provider policy. - Re-pin with changed text → a different identity today; the supersession relationship is the owner's explicit decision, made before this step is built. Provider no longer returns the poem →
unavailable_aton the record, last revision kept. - Retire a pin (owner): nulls references (R7), withholds items; never hard-deletes while referenced.
- "Pin for curation" on the PoetryDB shelf card, operator-only, local-only; and a "pinned" mark on the card thereafter.
- Providers without retention (Pexels, Unsplash, GIPHY, Bing News): not in this part; the policy flag is read so a later provider is configuration, not code.
Curation.Candidateslists the pin for its page, and a source-record work renderer draws it on the public page: main still withholds non-catalog work items as:unsupported_item, so storage alone cannot meet the public round trip.
Tests: pin, an already-materialized poem found rather than duplicated, author identity through the crosswalk (an unresolved poet gets no identifier), eligible as a work item, offered by Candidates and rendered, withheld after a text change and restored after re-review under the decided supersession rule, retired → withheld, unavailable → still eligible while retention permits and the card says so.
Done: one PoetryDB poem pinned for a pilot page, chosen as a highlight (via P2 or iex), published, and the withhold-and-restore round trip shown with ids.
P4 — work and exemplar candidates in the packet, and the trimming policy
Spec: #253 §5a step 2 in full (Bierce entries in and counted against the content quota, refusal when they exceed a bound, quotas, the meaning bound, tie-break by object id, omitted list on the attempt, out-of-subset candidates do not dirty). Needs only P1: existing catalog artwork fixtures serve for work; P3's pinned works join when P3 lands; poems and books join only once #253 D6's relation exists. Packet version becomes runtime-packet-v2.
Read first: lib/devils_dictionary/curation/runtime/{packet.ex,contract.ex,endpoint.ex,attempt.ex}; Curation.Eligibility for work and exemplar; priv/curation/runtime_bench/packets/*.json and plan.json (the frozen packets must still verify, or be re-frozen with a recorded reason).
Build: candidates of kind work (page-matched catalog artworks; P3's pinned works once P3 lands) and exemplar (accepted illustrates claims) in Packet.build/3; bound meanings as well as candidates, retain every chosen candidate's required meaning, and test real whole On scopes (packet v1 refuses more than 12 meanings independently of its 12-candidate cap; love's whole On scope has 43 senses, 48 packet meanings; 36 counts the lowercase subset only); freeze ids, revisions, rights and claim standing and omitted reasons deterministically; the trimming policy replacing {:too_many_candidates, n}; packet_summary.omitted on the attempt; Contract validation accepts the new kinds and still requires an exact substring for a quotation only. Keep :priority_candidate_unavailable as a refusal.
Tests: quotas and order are deterministic across runs; Bierce entries survive any cut, and too many of them refuse rather than trim; love's whole On scope (43 senses, 48 packet meanings) builds under the meaning bound; the omitted list names every dropped candidate; a packet from the benchmark fixtures verifies under v2 or the re-freeze is recorded; the contract rejects a work highlight with a quote.
Done: a packet for a pilot page with more than 12 candidates and more than 12 meanings builds, freezes, verifies, and its attempt lists the omitted candidates.
P5 — Freshness: the refresh state, admission on visit, the worker
Spec: #253 §7 in full and §6a in full; #198's "Admission and freshness" and "Publication and reader stability" paragraphs (#198 is reopened for this part). The dedicated issue, #259, wins on detail. No model call in this part: the worker completes without inference when the packet hash is unchanged, and otherwise marks the page as needing an editor (P6 turns that into a proposal).
Read first: lib/devils_dictionary/discovery/{policy.ex,run_worker.ex,run.ex,budget.ex} (the pattern); lib/devils_dictionary/curation/runtime/gateway.ex (fence, lease, budget reservation); Curation.Eligibility.evaluate/5 (the fingerprint); config/config.exs Oban queues; WordLive mount.
Build:
- Migration:
curation_refresh_stateswith the columns in Curation P5: freshness — the refresh clock, admission on visit, the Stale tab, and the owner-versus-bot rules in code (under #254, #253; reopens #198) #259 §5.1 (§7'slast_completed_at,refresh_after,next_retry_at,input_fingerprint,dirty_at,refresh_generation,state, plus the admitted parent,superseded_on_arrival, the dirty sequence, inputs-last-changed history and the lease fence), unique on composition id and configuration version id, plus a partial unique index allowing onequeued/runningrow per composition. - Admission (
Curation.Refresh.admit/2): the §7 predicate (pilot allowlist only during MVP; later expansion disabled), in one transaction with the composition and row locks, generation allocation, the published parent captured, and the Oban insert (curationqueue, unique per composition). Called on every page visit, plain HTTP and LiveView, bounded; a visit never calls the model. Budget: 30 globally pending review-work units, reserved and released with actual reviewable work; unchanged assessments consume no new review unit. Daily inference occupancy is a separate 30-minute budget. P5 writes no reviewable work and calls no model, so it reserves neither. Curation.RefreshWorker: build and freeze the packet (P4's builder, or v1 if P4 is not merged), compare the packet summary hash with the stored fingerprint; unchanged → complete (last_completed_at,refresh_after); changed → recordneeds_editor(the inputs were checked, not judged; P6 consumes it) and complete; failures setnext_retry_atwith the backoff and never advancelast_completed_at.- Dirty triggers: read-time withholding, a claim review decision, a source revision on a referenced record or a newly eligible candidate in the cut, a configuration activation, a pinned work's revision →
dirty_at, debounced 60 s, coalesced per composition; a change arriving during a run stays pending. - §6a:
suppressedon withdrawal, cleared by "Re-open" (P2);superseded_on_arrivalpersisted when the published version id changed since admission; generation and lease fence checked on every write; late writes refused after suppression or a configuration change; expired leases recovered with a bound. - The disclosure's four dates: last assessed, checked against the record (render time), inputs last changed, next assessment after, distinct from published.
Tests: fresh visit admits nothing; stale visit admits one; second stale visit admits none; withdrawn admits none; unchanged packet completes without inference and without touching the publication; a failure backs off and does not advance completion; a stale worker's write is refused by generation; a publication during a run yields superseded_on_arrival; plus a deterministic race test for each #259 §5 contract.
Done: on a pilot page, withdrawing a claim dirties it, the next visit queues exactly one job, the job completes, and the disclosure shows the five dates distinctly.
P6 — The proposal path: attempt → model version, attribution, Take, the quality form
Spec: #253 §5a steps 3–5 (attribution rules in full), §6 ("Take"), §6a (proposal never moves the pointer; dismissed proposals are not re-proposed), §10 (the quality form). Reopen #197 for its first half. Needs P2, P4, P5 merged, and P6 live inference requires P7's revised-instruction benchmark go.
Read first: Curation.Runtime.run/3, Runtime.Contract.validate/3, Curation.Compositions.create_version/3 and its note/2 (attribution is never an input: keep that rule and add a separate proposal writer), CompositionVersion and CompositionItem schemas, Curation.Reviews, P5's worker.
Build:
- Migration:
origingains:model;selection_origingains:model;attempt_idoneditorial_composition_versions(FK toinference_attempts); aproposal_stateor reuse of reviews fordismissed. Curation.Proposals.write(attempt): from a validated attempt, a version withorigin: :model,attempt_id,created_by_actor_idthe editor's bot actor (thegeneric-editorprofile gets a bot actor; no dossier), parent the composition's current published version at admission, items with notes signed:modeland the actual pinned instruction label with its model tag (never a hard-codedgeneric-editor v1). Never publishes.- P5's
needs_editor→ P6 runs the editor throughRuntime.run/3with the frozen packet and writes the proposal; respects the daily budget and review capacity. - The screen: "proposed" tab; "Take" per §5a step 5 (unchanged items keep model attribution; rewritten notes become human; added items human; the proposal stays reachable as parent); "Dismiss" records a rejected review and the arrangement hash, and the same arrangement on the same packet hash is not re-proposed.
- The quality form on the proposal view: five yes/no fields per item (relevant, correct meaning, useful reason, not a duplicate, attribution correct) and a free line, stored as the review's note; exportable as Markdown for posting 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.
- D3's fallback wording for a bot-chosen lead.
- 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 D8 (the owner's kept and mine marks, stored on the composition with who and when) and D9 (every editor-chosen item linked to its attempt; the attempt records the candidates offered and those omitted by the cut).
Tests: a proposal never moves the published pointer; Take preserves model attribution on unchanged items and human on rewritten ones; the proposal is the parent; dismissed arrangements are not re-proposed; a proposal written after a publication is flagged superseded_on_arrival; the form is required before Take.
Done: five pilot pages have proposals from real packets; five quality forms posted on #253; at least one taken, at least one overridden with reason, ids on the PR.
P7 — Runtime: #221's fixes, the instruction revision, the benchmark re-run
Spec: #221 (reopen it) A1, S1 and B1 as written, including A1's retry ceiling (the transport_retries CHECK); the condition in docs/curation/runtime-benchmark-2026-09-27.md ("revise the instruction as a new pinned version and re-run this plan"). Independent of P1–P6.
Read first: #221; lib/devils_dictionary/curation/runtime/{gateway.ex,runtime.ex,service_process.ex,bench.ex}; the benchmark record's Findings; priv/curation/runtime_bench/plan.json.
Build: A1 handle {:error, :stale} from Gateway.dispatch/3 and note_retry/2 as a refusal, and derive the transport_retries bound from one place; S1 a stop confirmed only when the parent and every descendant runner are gone and the port is closed; B1 the two vacuous benchmark cells labelled, and Bench.sample reading refused_shape. Then generic-editor-v2: the instruction revised per the findings (the adverse cases), pinned as a new instruction version, the plan re-run on qwen3.5-9b-v1 only, results committed under priv/curation/runtime_bench/results/ and a dated record beside the September one. Stop the service when done; note memory and swap before and after.
Tests: regressions for A1, S1 and B1; the benchmark harness still refuses an unconfirmed unload.
Done: #221 closed with evidence; a v2 benchmark record with its go/no-go for P6.
Checklist
A tick means the part's PR merged (implementation). Live acceptance is stated on each line separately.
- P1 — dedicated issue Curation P1: show one manually published opening on a pilot word page #255 — PR Curation P1 (#255): one published opening on /on/love #256, merged 8b6e2c9 (23 Oct 2026); Delivery B done, On love published live (composition 1, receipt 3)
- P2 — dedicated issue Curation P2: the owner's curation screen and the example tiles — curate the pilot pages by hand, with exemplars, poems, books and artworks on the page (under #254, #253) #257 — PR Curation P2 (#257): the owner's curation screen and example tiles #258, merged e25dd94 (24 Oct 2026). P2 Delivery A merged at e25dd94; owner demonstration/Delivery B not evidenced.
- P3 — dedicated issue Curation P3 (#271 §3): layer 1 cites every layer-0 record, transparently — shelf citations, sense definitions, linked claims, the citation receipt, retention-safe cleanup and the per-configuration window #270 (drafted 27 Oct 2026 under Curation infrastructure close-out: finish #267, #264, P3 and #198, accept #253 §10 on the live site, close every infrastructure issue and leave main clean — the last step before the persona agents (under #193) #269 W3, after the owner's D3 rules; its independent audit precedes any build) — PR:
- P4 — dedicated issue Curation P4: the packet carries works and examples, bounds meanings as well as candidates, and records its cut (under #254, #253) #262 (closed) — PR Curation P4 (#262): the packet carries works and examples, bounds meanings, and records its cut #265, merged 2dc3350 (26 Oct 2026, stacked on Runtime P7 PR 2 (#254): generic-editor-v2 pinned beside v1, the 9B re-run and the go/no-go for P6 #263); Delivery B done the same day, read-only: dry-run tables for love, war and hope on Curation P4: the packet carries works and examples, bounds meanings as well as candidates, and records its cut (under #254, #253) #262, love's clock re-admitted once for the fingerprint transition and settled.
- P5 — dedicated issue Curation P5: freshness — the refresh clock, admission on visit, the Stale tab, and the owner-versus-bot rules in code (under #254, #253; reopens #198) #259 (closed) — PR Curation P5 (#259): freshness, the refresh clock and the owner's rules #261, merged 9757c61 (26 Oct 2026); Delivery B done the same night: migration 20261025200000, one hashed restart, the clock proven live on love (two refresh jobs, suppress on withdraw, one run after re-publication).
- P6-0 — dedicated issue 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 (runtime made live: one pinned model, v2 registration and service rebind, PR Runtime P7 follow-up (#254): the 9B re-run on unlabelled ambiguous fixtures, frozen with packet v2 #266's GO, three rehearsal attempts against v2, and the nineteen 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 corrections from the 26 Oct audit) — PR: #267 Delivery A: mix dd.runtime rehearse, the Editor runtime card, and the operating window #272 open (27 Oct 2026), under Groundwork for the curated layer the agents will use: layer 1 chooses from every layer-0 record, citable and transparent to readers — #267, #264 and #270 (citations), in one executable order (under #193) #271
- P6 — dedicated issue 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 (its contract corrections are carried by 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 §4 and applied before P6 is built) — PR: Curation P6 (#264): the first bot proposes into the owner's screen #273 open (27 Oct 2026), under Groundwork for the curated layer the agents will use: layer 1 chooses from every layer-0 record, citable and transparent to readers — #267, #264 and #270 (citations), in one executable order (under #193) #271
- P7 — PR Runtime #221 fixes (P7 of #254, PR 1): stale refusals, the retry ceiling, confirmed runner stops, observable adverse counts #260 merged e8f1102 (Runtime stage A follow-ups from the #210 audit: stale dispatch refusal, orphaned ollama runners on stop, two vacuous benchmark cells (#195) #221 A1, S1, B1; Runtime stage A follow-ups from the #210 audit: stale dispatch refusal, orphaned ollama runners on stop, two vacuous benchmark cells (#195) #221 closed), PR Runtime P7 PR 2 (#254): generic-editor-v2 pinned beside v1, the 9B re-run and the go/no-go for P6 #263 merged e44122d (generic-editor-v2, the 9B re-run: no-go, 8 of 9, the ambiguous case abstained on self-labelled fixtures; record docs/curation/runtime-benchmark-2026-10-26.md) and PR Runtime P7 follow-up (#254): the 9B re-run on unlabelled ambiguous fixtures, frozen with packet v2 #266 merged 26 Oct 2026 (the unlabelled re-run on packet v2 with a labelled control: go for P6, 9 of 9, crane selects with both quotations under their senses, the control abstains and decides nothing; run 1793029211, attempts 52–70, swap 0; record docs/curation/runtime-benchmark-2026-10-26-unlabelled.md). Live inference still waits on 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 B.
Activity
holden commentedon Oct 25, 2026
Independent assessment — 25 October 2026
Verdict: progress C; issue coherence C+. The programme now has a real, publicly readable composition and a merged human editor that uses the same version/review/publication services. Those are useful foundations for agents. It has not yet demonstrated the owner's intended curated encyclopedia workflow: only 1/10 pilot pages has a public opening, no verified bot-to-screen proposal path exists, no curation freshness worker exists on main, and three mistaken assumptions about candidate availability remain in the governing specifications. The main risk is implementing the remaining parts literally: several briefs disagree with the record, each other, or the product's acceptance. Finish the small owner/bot loop before expanding ingestion, hosting, coverage, lists or panels.
Evidence and limits
I read #193 and its consolidation comments, #253–#255, #257, #259, #198, #252, #217, #170, #111, #192, #221, the P1/P2 PR descriptions and limitations, and the named main modules and curation documents. Main was
e25dd94a1055baf83b957c1b99f46bf60ad0c87c, also #258's merge commit. GitHub confirms #256 and #258 are merged. Code references below are to that head.I independently fetched all ten public pilot pages, inspected love in a browser, checked its disclosure, and checked the exact-word love routes and the public operator guard. All ten On pages returned 200; public
/ops/curationreturned 404. I did not sign in.The requested v2 SQL audit could not be executed. This assessment ran on a different host: the supplied workspace is an older Rails checkout, the external-drive Elixir checkout is absent, localhost:4007 and localhost:5434 refuse connections, and the connected Mac hostname did not resolve over SSH. The explicitly read-only Postgres connection attempt reached no database. Accordingly, the candidate numbers below are independently observed public-page proxies, not a falsely certified SQL/Eligibility result. I ran no migrations, tests that create/write databases, inference, seeds, publications or issue edits. The only external write is this report.
A. #253 §10 acceptance, item by item
Thus 0/7 acceptance items is fully demonstrated end to end; this is not a claim that the underlying persistence work is absent.
Today love has a canonical satirical lead and a related literary quotation, but no representative-example tile or independently selected work tile. Its notes and selector/publisher are Account #2, not the owner's public name; #255 explicitly accepted these as placeholders. War still has its instance list and catalog shelf without an opening; greed still has senses/discovery without an opening. None demonstrates a completed curation freshness check.
Database totals: the latest explicit live count evidence is #255's Delivery B: 1 composition, 1 version, 3 publication receipts, the latter being publish/withdraw/publish, not three published pages. Love's current disclosure corroborates those particular IDs; it cannot prove current database-wide totals. #258's composition IDs 1–4 belong to its separate test database. #259's “three live compositions” is an assumption, not evidence.
B. What an owner can choose now
Public-page observation, interpreted through Candidates, not a substitute for the unavailable SQL audit:
Counts: eight pages expose potential entry leads; zero expose accepted exemplars; three expose catalog works, all artworks (7 displayed rows). Greed and bestiality have none of those three categories, but both have quotations: they are not wholly empty candidate pages. All ten expose at least a quotation. All ten lack one or more of the owner's lead/example/work requirements; none currently has all three selectable categories, because accepted exemplars are absent from all ten public views. Final admissibility, hidden accepted records, and complete shelf cardinalities require v2 checks; I do not certify these proxies as database counts.
The committed manifests independently contain 2,903 poems, 946 books, 1,644 Met artwork rows and 1,575 Wikidata artwork rows. Those are manifest rows, not counts of works offered on a pilot page or distinct works after overlap. Corpus size is not curation coverage.
P2 found three structural limits, all confirmed in code:
Poems and books are not related to pages. Their manifest kinds declare
evidence: :none; Candidates takes onlyArtworks.shelf_items/1rows for work candidates. Consequently 0/10 screens offers a catalog poem or book through that seam, despite the record accepting such references and the reader drawing their tiles. A hand-written service version can select a catalog poem; that bypass does not prepare the owner's screen or the bot's candidate set. Close this through an explicit owner decision in 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 and its P3/P4 handoff about the existing attestation/page-selection relation, not a guessed “uses word = about meaning” claim. Cost: a bounded matching/selection contract and its candidate/eligibility tests; retaining full poem text is a separate P3 cost.Senses cannot lead. Content leads reference content items; WordNet and Wiktionary supply senses. D3's WordNet → Wiktionary fallback therefore cannot be executed. For the current record, greed and bestiality must have no lead. Either correct the specification to accept this pilot limitation, or explicitly authorize a narrow sense-reference extension with storage, policy, rendering, validation and recovery work. P4 and P6 alone cannot make a sense into a content item. Do not manufacture a duplicate definition to evade the record.
Examples need accepted nominations.
instance_ofexamples and pendingillustratesnominations are not selectable exemplars. The owner can nominate/review appropriate existing subjects using the existing Examples flow, then select them through P2; Curation P2 (#257): the owner's curation screen and example tiles #258 proved this on test war. Cost: actual sourced nomination and review work, not 44 automatic approvals. Link the words their sources already identify: mint the 21,544 QIDs WordNet and Wiktionary name, run the identifier rungs lexicon-wide, and link on visit (after #172) #217 improves source-identified entities/sense links; it does not supply a general pool of people to nominate. Exemplars: people, works and media as attributed examples of a meaning — nomination, the why, votes, ranking and the reverse view on a person's page #105 holds nominee search/mint design, while Link the words their sources already identify: mint the 21,544 QIDs WordNet and Wiktionary name, run the identifier rungs lexicon-wide, and link on visit (after #172) #217's folded Why Voltaire is not in the registry: people enter through senses, not through credits — 6,674 persons are word-referents; decide on-demand minting, a people seed, or both #165 discussion still leaves the broader people-seed choice unresolved. Bot nomination writes would additionally require Gate every illustrates nomination until accepted before bots or the public can nominate (#181 decision 6, deferred) #190; selecting already accepted examples does not.C. Concrete issue edits required
These are proposed edits only; none has been applied.
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 §3, §4c–d, §6, §9 and D3; Curation P2: the owner's curation screen and the example tiles — curate the pilot pages by hand, with exemplars, poems, books and artworks on the page (under #254, #253) #257 §1–3 and §5.3. Replace the false availability statements with: “WordNet/Wiktionary senses are not entry leads or content highlights in the current record; instance lists are not accepted exemplars; catalog poems/books have no page-match seam. Greed currently offers quotation highlights and no lead.” Replace the illustrative greed/war candidate lists accordingly. Record the two decisions in B explicitly. Change the specifications now; keep the safe eligibility/claim boundaries.
Curation MVP implementation plan: seven parts with agent-ready briefs (P1 read compositions, P2 screen, P3 pin, P4 packet, P5 freshness, P6 proposals, P7 runtime) — under #253 #254 discovery paragraph and P3/P4 dependency; Curation P2: the owner's curation screen and the example tiles — curate the pilot pages by hand, with exemplars, poems, books and artworks on the page (under #254, #253) #257 §1. Replace “poems, books and artworks are selectable highlights in P2 without P3” with “catalog references can be stored/rendered, but P2 lists only page-matched artwork rows; poems/books need an explicit page-selection relation.” P4 can use existing catalog artwork fixtures without waiting for P3, but must not claim poems became offerable merely because manifests exist.
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 §9.6 and §10 poem acceptance; Curation MVP implementation plan: seven parts with agent-ready briefs (P1 read compositions, P2 screen, P3 pin, P4 packet, P5 freshness, P6 proposals, P7 runtime) — under #253 #254 P3. Correct the identity contract before implementation: “PoetryDB has no provider-issued poem ID. Reuse the existing
poetrydb_poemnamespace andPoetrydb.poem_id/3; it hashes author, title and lines. Changed lines currently mean a changed identity, not a new revision under the same identifier. Decide the permitted supersession relationship explicitly; never deduplicate by title.” Reuse the reviewed author-QID crosswalk; an unresolved poet must not acquire an invented identifier. The existing provider already exposes a durable identity adapter, so check whether a live result is already materialized before claiming every pin creates a previously absent work. P3 must also name its Candidates listing and source-record work renderer: main still withholds non-catalog work items as:unsupported_item. Storage alone cannot meet its public round trip. These are repairs to the existing pin requirement, not grounds for a broader ingestion project.Curation MVP implementation plan: seven parts with agent-ready briefs (P1 read compositions, P2 screen, P3 pin, P4 packet, P5 freshness, P6 proposals, P7 runtime) — under #253 #254 P1; Curation P1: show one manually published opening on a pilot word page #255 §2.7, §3.1 and implementation notes. Reconcile all P1 assumptions exposed by Curation P1 (#255): one published opening on /on/love #256: “language comes from the scope's lexemes, not
page.language_tag; scopes are exact sorted lexeme sets, overlapping lexical-page scopes cannot coexist; Published owns eligibility, the reader maps its result; assessment rows are absent until P5; definition highlights have their own tile; actor labels come from the record.” Retain the approved typography change and actual live five-lexeme love scope, rather than the two-lexeme fixture. P1's temporary work/exemplar withholding was repaired by P2; mark it resolved. Do not reopen P1 to redo that work.Curation P2: the owner's curation screen and the example tiles — curate the pilot pages by hand, with exemplars, poems, books and artworks on the page (under #254, #253) #257 §3 scope rule; editor's public-facing “Exact-word pages share this” text. Replace it with “Exact-word pages overlap this editing scope and cannot have a separate active lexical-page composition; they do not currently display this composition.” The editor claims sharing, while PublishedReader resolves exact signatures. I verified neither /words/9051/love nor /words/14794/love shows the On opening. Correct the built label now; actual sharing is a product/routing decision, not implied by K3.
Curation P2: the owner's curation screen and the example tiles — curate the pilot pages by hand, with exemplars, poems, books and artworks on the page (under #254, #253) #257 §2.3, §5.1–2; Curation MVP implementation plan: seven parts with agent-ready briefs (P1 read compositions, P2 screen, P3 pin, P4 packet, P5 freshness, P6 proposals, P7 runtime) — under #253 #254 P2. Record the remaining P2 departures: “poem tiles show title/poet/length/link, no lines; no editor height readout; pilot slugs are a configuration-overridable code default, not a new config-file edit; editor URL is /ops/curation/on/:slug; stale publication errors are stale_pointer/stale_parent.” Keep the truthful metadata-only tile until retained text exists. Replace the erroneous instance-based exemplar examples with accepted nominations. The artwork thumbnail change and added exemplar/work tiles are directly within P2's goal. Record heading-start versus full-heading measurements consistently: Curation P1 (#255): one published opening on /on/love #256's accepted 882–910 px and Curation P2 (#257): the owner's curation screen and example tiles #258's 878–906 px do not place the whole heading above 900; avoid claiming that stricter result.
Curation MVP implementation plan: seven parts with agent-ready briefs (P1 read compositions, P2 screen, P3 pin, P4 packet, P5 freshness, P6 proposals, P7 runtime) — under #253 #254 checklist; Curation P2: the owner's curation screen and the example tiles — curate the pilot pages by hand, with exemplars, poems, books and artworks on the page (under #254, #253) #257 Delivery B; Curation P5: freshness — the refresh clock, admission on visit, the Stale tab, and the owner-versus-bot rules in code (under #254, #253; reopens #198) #259 introduction and Delivery B. Say “P2 Delivery A merged at e25dd94; owner demonstration/Delivery B not evidenced.” Separate implementation-merged and live-acceptance status. Replace Curation P5: freshness — the refresh clock, admission on visit, the Stale tab, and the owner-versus-bot rules in code (under #254, #253; reopens #198) #259's assumed “three live compositions” with “read and record actual live composition IDs first; do not invent the missing two.” Curation P5: freshness — the refresh clock, admission on visit, the Stale tab, and the owner-versus-bot rules in code (under #254, #253; reopens #198) #259 must follow Curation MVP implementation plan: seven parts with agent-ready briefs (P1 read compositions, P2 screen, P3 pin, P4 packet, P5 freshness, P6 proposals, P7 runtime) — under #253 #254's rule to tick implementation on merge, not its current instruction to tick while merely opening/reviewing the PR. Keep Curation P1: show one manually published opening on a pilot word page #255 closed: its bounded P1 delivery is evidenced.
Curation P5: freshness — the refresh clock, admission on visit, the Stale tab, and the owner-versus-bot rules in code (under #254, #253; reopens #198) #259 §3 budget, §5.2; Curation MVP implementation plan: seven parts with agent-ready briefs (P1 read compositions, P2 screen, P3 pin, P4 packet, P5 freshness, P6 proposals, P7 runtime) — under #253 #254 P5. Replace “30 review units per day” with “30 globally pending review-work units, reserved/released with actual reviewable work; unchanged assessments consume no new review unit. Daily inference occupancy is a separate 30-minute budget.” This is what Visit-driven curation refresh: coalesced jobs, fenced leases and atomic publication #198 specifies. A no-model freshness check should not be charged as a fresh reviewable proposal.
Curation P5: freshness — the refresh clock, admission on visit, the Stale tab, and the owner-versus-bot rules in code (under #254, #253; reopens #198) #259 §5.1–3 and §5.5–6. Add the missing contracts explicitly: capture the published parent at admission, not worker start; persist the superseded-on-arrival outcome the schema promises; preserve dirty changes arriving during a run; retain inputs-last-changed history after clearing pending dirty state; atomically initialize absent state under a composition lock; recover expired queued/running work with bounded leases; reject late writes even after suppression/configuration change. Add the corresponding deterministic race tests.
dirty_at = nilon every completion would lose an update received during that run and erase the requested historical date. Generation checking alone does not distinguish a new dirty event within the same generation.Curation P5: freshness — the refresh clock, admission on visit, the Stale tab, and the owner-versus-bot rules in code (under #254, #253; reopens #198) #259 §3, §5.2 and §5.6; 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 §7 scope. State how a never-curated allowlisted page gets a composition/refresh identity, since the worker API and FK require one. Require a bounded visit check for the page visits promised by the product; connected-mount-only admission currently excludes plain HTTP visits. State “pilot allowlist only during MVP; later expansion disabled.” Curation P5: freshness — the refresh clock, admission on visit, the Stale tab, and the owner-versus-bot rules in code (under #254, #253; reopens #198) #259's immediate “or ≥ threshold” predicate contradicts D7 and Visit-driven curation refresh: coalesced jobs, fenced leases and atomic publication #198's no-automatic-expansion rule. Also carry configuration-activation and newly eligible in-cut candidate changes into invalidation. Define the versioned transition from P2's full candidate-spec digest to P4's packet digest; they must not silently mean the same fingerprint.
Curation P5: freshness — the refresh clock, admission on visit, the Stale tab, and the owner-versus-bot rules in code (under #254, #253; reopens #198) #259 §2.4; 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 §6a; Visit-driven curation refresh: coalesced jobs, fenced leases and atomic publication #198 reader stability. Replace “withdrawal and lost rights already behave that way” with “fresh loads revalidate; prompt suppression on already-open pages still needs an explicit bounded invalidation path and proof.” WordLive reads the opening during load; I found no curation invalidation subscription that updates an open opening. Keep normal publication navigation-pinned. This is an existing acceptance gap, not a request for live reshuffling.
Curation MVP implementation plan: seven parts with agent-ready briefs (P1 read compositions, P2 screen, P3 pin, P4 packet, P5 freshness, P6 proposals, P7 runtime) — under #253 #254 P4; 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 §5a packet policy. Add “bound meanings as well as candidates, retain every chosen candidate's required meaning, and test real whole On scopes.” Packet v1 refuses more than 12 meanings independently of its 12-candidate limit. Public war/love/justice/truth/hope show 30/43/17/19/54 senses, respectively; merely trimming candidates cannot make those full scopes pass. Resolve whether Bierce consumes the content quota and what happens if priority entries exceed the cap; the existing prose must not imply both unconditional inclusion and an impossible bound. Freeze IDs, revisions, rights/claim standing and omitted reasons deterministically.
Curation MVP implementation plan: seven parts with agent-ready briefs (P1 read compositions, P2 screen, P3 pin, P4 packet, P5 freshness, P6 proposals, P7 runtime) — under #253 #254 P6 dependencies and P7; 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 §11. Add “P6 live inference requires P7's revised-instruction benchmark go, plus P2/P4/P5; use the actual pinned instruction label, not hard-coded generic-editor v1.” Correct P7's references from “A1–A3” to Runtime stage A follow-ups from the #210 audit: stale dispatch refusal, orphaned ollama runners on stop, two vacuous benchmark cells (#195) #221 A1, S1, B1, including the retry ceiling in A1. Keep the five per-item quality questions and 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; they already have an owner in P6 and must not be dropped as “later evaluation.”
Delivery: curated opening section with five local persona editors and visit-driven refresh #193 consolidation table and current-status body. Update Curated opening section: Bierce-first lead, sourced highlights and inspectable editorial choices #156 to reader/publication delivered by Curation P1 (#255): one published opening on /on/love #256, screen/tiles by Curation P2 (#257): the owner's curation screen and example tiles #258, exact-scope reader without a binding table; Curation configurations: global default, domain panels and future personal views #201 to default activated, profiles still proposed; Persona panel pilot: five versioned editors, fixed quorum and ten-page evaluation #197 to generic proposal slice P6 first, panel/evaluation later; Visit-driven curation refresh: coalesced jobs, fenced leases and atomic publication #198 to reopened, Curation P5: freshness — the refresh clock, admission on visit, the Stale tab, and the owner-versus-bot rules in code (under #254, #253; reopens #198) #259 active specification, not implemented. Remove stale “all children are closed” and September open-PR/installation assertions from present-tense status. Keep Curation persistence: versioned panels, ballots and reviewed page compositions #196's unbuilt runs/ballots/decision authority distinction, Runtime stage A follow-ups from the #210 audit: stale dispatch refusal, orphaned ollama runners on stop, two vacuous benchmark cells (#195) #221's outstanding fixes, Curation transparency: equal votes, reproducible selection, profiles and version history #203 public history, Editorial method: declared perspectives, checkable facts and accountable curation #204 method, Gate every illustrates nomination until accepted before bots or the public can nominate (#181 decision 6, deferred) #190 nomination gate, Exemplar votes: assertion_votes.reason, cached assertion_scores, the vote control and endorse-on-duplicate (#181 build 3) #189 vote UI, Exemplars: people, works and media as attributed examples of a meaning — nomination, the why, votes, ranking and the reverse view on a person's page #105 remaining nominee work, Discovery relation assessor: lite ML and LLM judging of provider results before persona curation #101 assessor, Quote-shelf connect link and provider-evidence prefill (deferred from #212) #222 quote-connect and Grow the verified quote tier by use, not by rebuild: the decision that replaces the wikiquote-pd-v2 reselect #213 quote-growth decision as folded backlog. P2's private version/receipt history is not Curation transparency: equal votes, reproducible selection, profiles and version history #203's public history. I found no evidence warranting relabelling those other folded deliverables complete. Point the hardware handoff to Hosting: move the reader-facing site off the Mac Studio — host choice (Mac mini recommended), release build, and the pins to lift (absorbs #199, #229, #232) #252, which absorbs Local curation portability: measured Mac Studio to Mac Mini handoff #199.
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 decisions / §10; Curation P2: the owner's curation screen and the example tiles — curate the pilot pages by hand, with exemplars, poems, books and artworks on the page (under #254, #253) #257 Delivery B; Delivery: curated opening section with five local persona editors and visit-driven refresh #193 panel follow-up. Name the unowned prerequisites: owner's public actor label and replacement of accepted placeholder reasons in the three-page hand exercise; the poem-to-page rule in the P3/P4 handoff; whether existing nominee subjects suffice or a bounded people population decision is needed, using Exemplars: people, works and media as attributed examples of a meaning — nomination, the why, votes, ranking and the reverse view on a person's page #105/Link the words their sources already identify: mint the 21,544 QIDs WordNet and Wiktionary name, run the identifier rungs lexicon-wide, and link on visit (after #172) #217's existing ownership. State “panel automatic publication remains off until a held-out comparison with the generic editor using the same quality form and measured reviewer/machine cost.” The quality form and panel evaluation are already specified; their execution is missing, not their conception. Do not turn Naming: 100 broader ideas and eight extensions, plus the earlier top 50 #192's site naming or Lists: editorial buckets of words — Gen Z dating terms, The Devil's Dictionary, words coined by Vonnegut, words from a hearing — curated, derived or both, prioritised on the site #170's lists into curation prerequisites.
Overall coherence: #253 correctly governs, #254 usefully separates deliverables, and #255/#257 report their discoveries honestly. The fault is that discoveries have not been reconciled into the authoritative acceptance and next briefs. Reopening #198 is right: its admission/fencing/publication contract is still needed. Keep it as that contract and #259 as the implementation issue; closing #198 after only a model-free P5 would wrongly imply its crash/recovery/proposal obligations are all met. #252 is hosting, #217 is source-identifier coverage, #170 is lists, #111/#131 govern page presentation, and #192 is branding; these are coherent separate tracks, not additional predecessors to the bot pilot.
D. Readiness for bots
P2 supplies real page scopes, candidate item specs and standings, a faithful public-renderer preview, transactional publication and stale-form handling. P4 must adapt these semantics, rather than create a richer fictitious candidate pool: catalog artworks and accepted claims only where the current screen can offer them; no poems/books without the relation decision. P4 must handle the separate meaning bound noted above.
P5 must supply stable per-composition admission, a frozen input identity and admitted parent, generation/lease ownership, suppression, backoff, a durable needs-editor handoff, and preservation of changes during a run. A P5
needs_editoroutcome means the inputs were checked, not that an editor evaluated their quality; its disclosure must say that until P6 completes an editor assessment.P6 still needs model-capable version/item origins and database constraints, a bot actor without reviewer privilege, attempt linkage, exact proposal ancestry, the separate trusted proposal writer, Take attribution, dismissal deduplication, capacity accounting and the required quality form. “Use the same record” is currently true for human publication; it is not yet implemented for bot proposals. P6 should select accepted material, not create claims; adding bot nominations would activate #190 and broaden this slice.
Runtime go conditions are not met on main. The only benchmark record remains 27 September, Packet still pins
generic-editor-v1, Runtime still pattern-matches successful dispatch/retry, service stop checks parent/port rather than all descendants, and Bench still computes the adverse metrics from accepted-result fields. The 9B result is 13/16 validation passes, including three abstentions; the three adverse answers were refused, not three successful selections. Fixed-seed repeats measure timing, not independent editorial judgments. A new pinned instruction, A1/S1/B1 repairs, corrected observability and a 9B rerun are prerequisites to relying on it. No fresh hardware readiness, disk, memory or service-authority check was possible here; September measurements and #252's October inventory must not be presented as today's capacity.For an editor today, the table in B is the practical menu: eight pages have historical entry choices plus quotations, greed/bestiality have quotations without leads, three pages add catalog artwork choices, none has an accepted example or a screen-selectable catalog poem/book. A bot cannot solve these selection shortages by better prompting.
E. Order, scope and the smallest useful sequence
P5 → P7 → P4 → P6 is a reasonable serial order after fixing the briefs, but it is not a dependency chain: P7 can proceed independently. P6 needs P7's go explicitly. P4 need not wait for P3 to demonstrate existing artwork/accepted-example kinds. If P4 changes packets after P7's rerun, validate the changed contract on real v2 packets before P6 relies on it; do not represent a v1 benchmark as proof of work/exemplar quality.
The next three parts are not inherently scope creep. Freshness is explicitly in the owner's goal; runtime refusal/cleanup fixes are necessary before using it; packet kinds and deterministic bounds let agents select what humans select. Scope creep would be automatic candidate-threshold expansion in P5, a new daily review allowance, broad Health-page work beyond the specified counters, extra models, a hosting move, personas, lists or automatic claim generation. The minimal counters and five dates already specified are useful evidence; they do not justify an operations redesign. “Routing should work and nothing more” here means the bot proposes an eligible version into the existing screen and the owner decides.
Smallest sequence to three real pages:
This gets a bot proposing through the owner's record before P3, hosting or panels. It does not pretend to complete the poem acceptance or the broader encyclopedia goal.
Ranked recommendations
No code, corpus data, issue bodies or checklists were changed by this assessment.
holden commentedon Oct 25, 2026
Reconciliation applied, 25 October 2026
The owner accepted all fifteen edits in section C of the assessment above. Each is now in the sections it names, with one wording per fact across #253, #254, #257, #259 and #193. Edits 4 and 11 also name #255 and #198, so those two bodies carry the same wording. #255 stays closed. Text only: no code, data or state changed, and no issue was opened.
Facts used, checked on v2 and at
e25dd94: v2 holds one composition (love, composition 1). Of the ten pilot pages, 8 have an admissible lead (greed and bestiality have none), 0 have an accepted exemplar, and 3 have catalog artworks (war 2, love 4, hope 1). Packet v1 refuses more than 12 meanings independently of its 12-candidate cap, and love has 36 senses. PoetryDB has no provider id. The editor's "Exact-word pages share this" label is wrong.poetrydb_poemnamespace andPoetrydb.poem_id/3; it hashes author, title and lines. Changed lines currently mean a changed identity…; never deduplicate by title." Plus: the reviewed author-QID crosswalk, a check for an existing materialized poem first, and P3 now names its Candidates listing and source-record work renderer.editor.ex, which P5 already changes, so #259 §4 carries it.admitted_parent_version_id);superseded_on_arrivalpersisted; dirty changes during a run preserved (dirty_seq/run_dirty_seq); inputs-last-changed history (inputs_last_changed_at); absent state initialised under a composition lock; bounded lease recovery (lease_token, three attempts); late writes refused after a suppression or a configuration change.candidates-v1now,packet-v2from P4), so a stored fingerprint never silently equals one of another version.Section B's two decisions, recorded with defaults (the owner can change either on #253):
The two new decisions in #253 §12:
Next: P5 from the reconciled #259, Delivery A.
holden commentedon Oct 25, 2026
Independent audit — 26 October 2026 (Europe/Warsaw)
Verdict: ready to start P4; not ready to start the integrated P6 delivery or permit a first live bot proposal. Grades: reconciliation B; P5 C; P7 D; programme readiness C−. These grades assess completed delivery on main and the published installation, not just promising PR code. Remote main and the Mac checkout are both
e25dd94a1055baf83b957c1b99f46bf60ad0c87c: P5 #261 and P7’s first PR #260 are open and unmerged. P5’s table/worker are absent on main and v2; P7’s instruction revision and rerun have not been delivered. Reconciliation substantially repaired the briefs, but a few contradictory remnants remain. The owner’s three-page exercise is also incomplete. P4 has an implementable brief and its P1 dependency; it need not wait for poem pinning, P5 or P7.Scope and evidence
I read the current bodies and comments of #253, #254, #257, #259, #193 and #221, the original assessment and reconciliation comment, the supporting #255/#198 bodies, the PRs, main source/tests, and the two delivery branches (
20b7114for P5;10cc43cfor P7 repairs). Live PostgreSQL is 18.2, port 5434, database devils_dictionary_v2. Direct inspection used SELECT only withdefault_transaction_read_only=on; the existing page/candidate/packet readers also ran through a dedicated connection confirmed read-only, with the application not started. No model was called. No v2 migration, data edit, server restart, service stop/start, launch-script edit, issue-body edit or checklist change was made. Public GETs were limited to the requested love and Stale URLs.Independent precommit: 19 doctests, 2,867 tests, 0 failures; exit 0 (ExUnit 365.1 seconds, seed 687153; private worktree remained clean). It ran in a detached external-drive audit worktree at main, with deps/data and each replay gzip symlinked, using exactly
MIX_ENV=test MIX_TEST_PARTITION=_audit DD_NO_OBAN=1 mix precommit. Branch test totals below are explicitly author-reported, not independently rerun. Useful source anchors: main Packet, main Runtime, P5 Refresh, P7 stop repair. A green main suite does not test either unmerged delivery.A. Specification reconciliation
“Present” means I found the requested text in its named issue sections, not merely in the summary comment. Planned work is distinguished from behavior already on main.
evidence: :none. #257 §1 still initially calls the combined catalog inventory a shelf “served today”; its following correction and §3 correctly restrict what is page-matched.Poetrydb.poem_id/3andidentity_record/1. Source-record works remain:unsupported_item; P3 is specified, not delivered.page.language_tagin the §3.1 code example and §5.1, and test 7 still says to provision overlapping On/exact-word compositions. The correction governs, but these old instructions contradict it. P1 need not reopen.editor.ex:691still says “Exact-word pages share this”. Fixed on #261’s branch only. The original assessment asked for the built label correction; this remains a partial delivery.candidates-v1, HTTP admission and Pilot. Main has none of this. Activation is detected by an absent configuration-version row. #259 explicitly leaves some unhooked candidate/catalog changes to the window.All fifteen topics and both decisions were added; this is not evidence that their implementations landed. The material remaining inconsistencies are the deferred built label, whole-scope count, supporting #255 instructions, and stale #221 state. #253 §11 also still describes P1 as delivering the disclosure’s dates while its P5 row correctly owns the assessment dates; #255 and main explicitly leave those dates absent until P5.
B. P5 freshness
SELECT to_regclass('public.curation_refresh_states')returns NULL on v2. Latest applied migration is20261009120000.20261025200000: requested clock, parent, supersession, sequence/history, lease/backoff fields plus documentedrun_periodicandlast_outcome_reason; pair uniqueness and partial unique composition index for queued/running. Ecto uses:utc_datetime_usec, whereas the issue table literally says timestamptz; reconcile that type wording.Refresh.admit/2wraps advisory composition lock, row initialization/FOR UPDATE NOWAIT, generation, captured published parent and Oban insert in one transaction. A busy lock returns busy; lock timeout is bounded at 1 s. No review/inference reservation is made.superseded_on_arrivalpersisted. Race test 16 covers publication between admission and claim.dirty_seqversusrun_dirty_seqcontrols clearing; history survives. Race tests 17–18 cover change after input read and next admission.20b7114; I inspected its race/concurrency tests but did not independently run that branch suite.6704fdf2553d6a97b9c08f9e263e0aec545c21b2674f5e101ab4e177fec05291, mtime 21 Oct 21:08:50 CEST. The hash was identical on both reads during this audit. This proves no audit-time edit, not an activation before/after comparison. The prescribed helper’srollback-oninvokeswrite_script on; restartand overwrites the launch script. “Byte-identical” must be proved, not inferred from that command.Live refresh-state values per composition: composition 1, love has no refresh-state relation/row;
last_completed_at,refresh_after,dirty_reasonandlast_outcomeare unavailable, not NULL fields in a delivered row. There are no other compositions to report.Live Oban inventory, unchanged across the requested GETs: CleanupWorker completed 672; Discovery.RunWorker completed 62, cancelled 2; Quotations.VerifyWorker completed 337. Curation.RefreshWorker: 0 rows; no other curation worker rows. Thus the expected one refresh row/job per existing composition after P5 activation is not evidenced. Ordinary background jobs must not be mistaken for unexpected curation work.
C. P7 runtime
gateway_race_test.exsdeterministically sweeps between admission/dispatch and asserts zero sends; runtime test covers stale retry. Not merged.refused_shape, adds denominators/unobservable nils, distinct answers and separate outcomes, corrects September prose; regression tests exist. Not merged.@instruction_version "generic-editor-v1"; no v2 pin or config-derived instruction selection.runtime-benchmark-2026-09-27.mdand September result exist; no new record/results/go.The September numbers are 13/16 validation passes = 10 selections + 3 abstentions; 3 refusals, not 13 successful selections. The ordinary Bierce lead is 6/6; sparse abstentions 3/3; all three adverse answers were refused. Two adverse zeros cannot establish non-use/non-approval when those fields were unobserved. Fixed-seed repeats are not independent editorial judgments. These numbers support the September conditional development recommendation, not an unconditional P6 go.
The assessment’s grouped prerequisites—new instruction; A1/S1/B1 repairs with corrected observability; 9B rerun—are all unmet on main. Repair code exists in an open PR; that is progress, not closure. #260 reports 19 doctests / 2,883 tests / 0 failures at
10cc43c; this is author-reported and does not supply a benchmark.D. The owner’s pages — live numbers
E. Gates and smallest path to proposals
Status vocabulary: done = present/verified on main or live as applicable; specified only = an explicit contract exists but main implementation/acceptance is outstanding; missing = the required delivery artifact is absent.
Today’s actual candidate supply, read through main’s Candidates over each real public On scope (eligible highlight candidates; content can also supply leads, so these are not counts of distinct selections):
The separate meaning counts were measured with a reader-only builder option raising the candidate limit, solely to expose its second refusal; nothing was dispatched or persisted. The differences between human Candidates counts and Packet v1 counts are further reason to verify actual v2 packets rather than infer them from screen counts.
Ranked recommendations / minimum sequence
This audit posts one report comment only. It does not certify unmerged code as deployed, authorize live activation, close #257/#198/#221, or change the programme’s acceptance bar.
Merge pull request #260 from razrfly/claude/221-runtime-fixes
v2 rules out only an excerpt that directs the editor, the ambiguous c…
The word itself is not a wrong sense for a definition, said in the co…
Beside September's wrong_sense, the ambiguous case counts each declar…
The v2 plan records the owner's choice: the ambiguous go check is Sep…
The 9B re-run under generic-editor-v2: 16 of 16 calls validated, the …
The re-run's record says what one judgment shows: v2 meeting self-lab…
Merge pull request #263 from razrfly/claude/254-p7-editor-v2
holden commentedon Oct 26, 2026
P7: the runtime go/no-go for P6 is no-go (26 October 2026)
P7 is done. #221's fixes are in #260 (merged as e8f1102, #221 closed with evidence). The instruction revision and the 9B re-run are in #263 (merged as e44122d). The record is docs/curation/runtime-benchmark-2026-10-26.md, and the raw results are
priv/curation/runtime_bench/results/2026-10-26.json.What ran.
generic-editor-v2. That is the instruction revised per September's Findings, pinned beside v1.Packetreads the instruction from the pinned config, and a mismatch is refused before admission.qwen3.5-9b-v2: the same 9B artifact and settings asqwen3.5-9b-v1, registered as a new config because a config pins its instruction.devils_dictionary_runtime_benchonly (run 1792976835, attempts 36–51), on the owner's yes.Bench.verdict/5evaluates them, and a check fails on missing or unseen evidence.Result: 8 of 9 checks pass. 16 of 16 calls validated: 10 selections, 6 abstentions, 0 refusals.
duplicate_candidate)Why no-go, and what it does and does not show.
Runtime health was good.
llama-serverrunner (Runtime stage A follow-ups from the #210 audit: stale dispatch refusal, orphaned ollama runners on stop, two vacuous benchmark cells (#195) #221 S1, live).Until a go, P6's live inference should not rely on this runtime and instruction.
Next, being prepared. A re-run that changes only the ambiguous fixtures: a different two-sense word, with unlabelled fictional text, frozen with the merged packet v2 (#265). The instruction (
generic-editor-v2), the model (qwen3.5-9b-v2) and the go checks stay unchanged. Under packet v2 a definition is offered only at the word itself, so the ambiguous expectation becomes "each quotation under its own sense". The owner will be asked to accept that before the run.holden commentedon Oct 26, 2026
Full MVP infrastructure audit — 26 October 2026
Verdict: the delivered P4/P5/P7 infrastructure is a credible foundation for implementing the single-editor P6 path with a stubbed runtime, after correcting #264’s incompatible seams; it is not a completed MVP or a go for live bot proposals. My live-inference verdict remains NO-GO. Grades: P5 A−; P7 B+; P4 A−; reconciliation B; P6 readiness C+. Main is
2dc3350626cb2276c8b13819375c01b4af3330bb, also the remote main I checked. Unlike the previous audit, all four implementation PRs are merged and P5 is live. The outstanding problem is no longer absent infrastructure: it is an honestly failed benchmark gate, an unbuilt proposal lifecycle, and a P6 brief whose profile admission, refresh outcomes, immutability, attribution and budget instructions cannot all execute as written. This supports bounded P6 engineering, not autonomous publication, personas or panels.Evidence and limits
I read #253/#254/#257/#259/#193/#221/#262/#264, their comments, PRs #260/#261/#263/#265, main’s implementation/tests, both benchmark plans/results and the 26 October record. I reproduced the three requested CLI dry-runs and read all ten pilot pages through the same candidate/packet seam. Database reads were on PostgreSQL 18.2, port 5434,
devils_dictionary_v2, withdefault_transaction_read_only=onexplicitly confirmed. The reader process started only its repository/dependencies, not the application, Oban, endpoint or model. Freeze/verify and benchmark recalculation were read-only. No migration, data edit, live service start/stop/restart, inference, source change, issue-body edit or checklist change was made.Independent suite: 19 doctests, 2,979 tests, 0 failures; exit 0 (seed 646998, ExUnit 406.1 s; private worktree clean), run at the audited main in a private external-drive worktree, deps/data and every replay gzip symlinked, using
MIX_ENV=test MIX_TEST_PARTITION=_audit3 DD_NO_OBAN=1 mix precommit. The supplied permission covered that isolated test partition only.Code evidence below uses the audited revision: Packet, Contract, Refresh, Profiles, foundation constraints/triggers.
A. Claimed versus delivered
refresh_race_test.exscovers contracts 16–18/20–22;refresh_concurrency_test.exsuses separate committed connections for admission races. Worker still makes no model call.refused_shapeobservability/denominators and separate outcomes are present. Runtime/gateway race, service-process and Bench regressions are present and included in my suite.Bench.summarize/3andBench.verdict/5: NO-GO, 8/9, 10 accepted selections / 6 abstentions.devils_dictionary_runtime_bench, not v2. Its immutable config is qwen3.5-9b-v2, same 9B artifact/settings as v1, new config because it pins generic-editor-v2. v2 has 0 model configs, 0 inference service rows, 0 attempts. A merge does not install that benchmark config into the live DB.Candidates.for_page/1; content/quotation/work/exemplar support, deterministic quotas, required-meaning preservation, both 12 limits, explicit omissions, freeze/registry verification and text-free attempt summaries are present. Contract refuses quotes on works/exemplars. Tests cover the cut, excess Bierce refusal, old v1 compatibility, meaning integrity, work/exemplar contract and fingerprint transition.page.language_tagexample/§5.1 and “provision both” overlapping-scope test; #193’s “current” body still says freshness unimplemented/only September benchmark, and its #221 row still describes open repair work. #254/#257 also retain “built label still says share this”, although main fixed it. Preserve historical records, but stop presenting these statements as current.Live refresh state, SELECT snapshot (timestamps in UTC, as stored):
Fingerprint prefix is packet-v2/cut-v1/meanings-v2; no failures, next_retry_at or outstanding lease. This is a completed input check, not a completed editorial assessment. Disclosure explicitly says “the inputs were checked, not judged: an editor is needed (the inputs are now identified differently).”
Jobs: RefreshWorker 3 completed: 4645/gen1, 4646/gen2, 4730/gen3. The first two substantiate P5 Delivery B; the third is P4’s later transition, not an unexpected duplicate P5 run. Other rows: CleanupWorker 672 completed; Discovery.RunWorker 62 completed + 2 cancelled; Quotations.VerifyWorker 337 completed. No curation inference/proposal job exists.
Activation history: receipt 4 withdraws love and 5 republishes v1 during P5’s proof. The Delivery B comment documents the owner’s yes, named migration, direct bootout/bootstrap rather than the rewriting helper, and one restart. I independently confirmed the migration, resulting clock/jobs, live disclosure and current script hash 6704fdf2553d6a97b9c08f9e263e0aec545c21b2674f5e101ab4e177fec05291, identical to my previous audit’s hash. The exact historical restart count and original approval chat are documented there; SELECT/current process state cannot independently reconstruct them.
B. Benchmark: what failed, what PR 3 can establish
The 26 October record and raw JSON agree with my recalculation:
ambiguous_each_sensecheck requires selection and sense coverage, so zero wrong senses on an abstention cannot pass it.Attribution of the failure: it is justified to say v2 abstained on these self-labelled fixtures, citing their fictional status. It is not established that the labels alone caused the failure or that the instruction generally became overcautious. The same labels existed under v1, which selected; instruction and fixture interact. The record’s main findings make this qualification correctly. Its Next paragraph and #254’s follow-up comment overstate the proposed experiment: changing the lemma, fixture wording, packet version and expectation cannot establish “the label was the trigger” or, on another abstention, uniquely establish “the instruction is at fault.”
PR 3 is planned, not delivered. No PR 3/new v3 plan/result is on the audited remote main or in the PR listing I checked. A predeclared follow-up using unannounced synthetic evidence and actual packet v2 is a reasonable remedy for the P6 readiness test, provided its changed variables are recorded and it is not presented as a single-variable causal experiment.
Restating the ambiguous expectation as select the two quotations, each under its own pinned sense matches v2: content candidates now have the screen’s lexeme meaning; quotations have their own single allowed sense. Requiring definitions at sense ids would test behavior P6 cannot accept. The check must require the intended quotation ids/coverage and a selection, not merely “wrong_sense = 0”; revise the plan and harness/test together. Because v2 already fixes each candidate’s allowed meaning, this tests useful selection/coverage within the contract, not unconstrained model disambiguation. It does not replace the five owner quality forms.
Own decision: NO-GO for P6 live inference. A go still requires the owner-approved, precommitted follow-up plan and packets, unchanged pinned instruction/model or explicitly recorded new versions, a completed run passing every declared non-vacuous check, separate outcomes/observability, memory/swap and confirmed stop evidence, and verification that real v2 pilot packets fit and validate. The last packet-building requirement now passes; the new model run has not happened. Passing it would permit controlled owner-reviewed proposals after P6’s implementation/activation gates, not automatic publication or a panel.
C. P6 gates against #264 and #253
“Done” means main/live evidence exists. “Specified only” means the behavior is requested but unbuilt; “missing” identifies an absent prerequisite or an executable contract not actually supplied.
D4 is clear: automatic publication remains off. A future human Take path may publish under human authority; “Proposals never publishes” must not be misread as forbidding that explicit owner action.
D. Owner’s pages and current candidate supply
Live v2 has 1 composition, 1 version, 1 review, 5 publication receipts. Only On love, composition/version 1/1, is published; final receipt 5. Selector/publisher is actor 5, human user 2, still labelled Account #2. There are 0 model versions and 0 attempts.
Placeholder notes are not replaced:
P5’s withdraw/republish did not create two more owner-curated pages or replace P2’s three-page screen exercise. #257 should remain open.
I reproduced
mix dd.runtime packet --scope SLUG --dry-run --instruction generic-editor-v2for love/war/hope, on a forced read-only connection; all ten rows below were also computed through the same Packet/Candidates API. No model config was invented to make the CLI run.In C/Q/W/E columns, values are screen-eligible → packet-kept. Leads overlap the content pool and are not extra candidates. Meanings are kept / total; omitted.
Unused content quota flows into quotations, explaining 4 or 5 quotations despite a base quota of 3. No candidate in these live cuts was omitted specifically for its meaning; the separate meaning bound still drops unused meanings. All selected candidates retain their required meaning. No selectable poem/book or accepted exemplar appeared; the work counts are catalog artworks.
Reproduced packet hashes:
7a6061f8caa4ea5be3a248132d2f88771fb76c1510d61f2ad468502ddf75ec731fc7c48c86e30892a3ff2685a59392e02f1aea298daae8586c525988aa3fe1a5f21bb3b93a17a1774c3745ee7103e1a3686dc021e2984d6394e64cf3548bbca2Love, war and hope remain the richest first three proposal pages. Only love currently has a composition; the other two must acquire exact On-scope composition identities before a P5 visit can enqueue anything.
E. #264: exact corrections needed before implementation
These are corrections to the requested slice, not additional features.
generic-editor v2 · qwen3.5-9b-v1. Treat the wireframe’s “12 of 36” as illustrative; love’s current measured count is 10 of 32.outcome/3returns the internal pair, but the worker’s claim/check path returns successful completion and the persisted row; it does not emit a reusable{:needs_editor, reason}job handle. Specify the integration point using the claimed row id, configuration version, generation, lease token and admission parent, not only composition_id. Also consume the durable needs_editor handoff when its input-check row is otherwise fresh; a repeat ordinary visit currently queues nothing until dirty/due.curator_profiles_admission_guard. Include the narrow generic-editor admission/schema rule in P6’s migration, while preserving the persona evidence rules; do not fabricate a death/source dossier.:modelenums/FK does not authorize a bot insert:editorial_composition_versions_human_actorcurrently refuses it. Specify the model-author guard and attempt/config linkage explicitly, without weakening reviewer/publisher/activation guards.RefreshState.last_outcomeand its DB CHECK accept only unchanged/needs_editor. “Completes … last_outcome: abstained” therefore needs both schema/constraint and disclosure/outcome changes in the P6 migration, or an explicit separate editor-outcome record. Delete “or abstained” from the version-writing branch: only an accepted selection creates a proposal; an abstention creates no version and does not count toward five proposals. Keep input checks distinct from actual editorial assessments; failed/refused inference must not acquire successful-editor freshness.superseded_on_arrivalis also not inherently an exclusive alternative to subsequently taken/dismissed; define that lifecycle instead of an ambiguous four-value bucket.editorial_composition_items_guard; 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 says the marks live on the composition. Specify mutable composition-level mark identity/who/when, or the exact permitted metadata update rule without rewriting the selected arrangement. Carry the marks across version changes by stable selection identity, not transient form position.editorial_composition_items_insert. Define copied-item provenance/author validation and preserve the original human signature; only editor-authored choices/notes acquire model attribution.:kept_item_droppedrefusal does not satisfy 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 D8’s “survives into every bot proposal”. Do not silently change the packet policy or synthesize a model choice to hide this conflict.Reviews.decide/4stores:reasonand requires:idempotency_key. State that the fixed form is serialized in reason, or explicitly add the storage/API in the migration; bind it to the exact reviewed version/items. Clarify the opening “form each time” against tests requiring it only for Take/Edit: Dismiss must at least retain its required rejection reason; do not imply nonexistent completed item forms.proposal:<composition>:<generation>, but define one frozen packet/config identity for that attempt and exactly-once write/reservation. Rebuilding changed inputs under the same key cannot be treated as an identical replay; use the existing conflict/stale rules and a new admitted generation where required. Persist suppression on the same arrangement and packet, and enforce attempt-to-version uniqueness under retries/crashes.Compositions.one_display_identity/1. Contract checks repeated candidate ids; record-level work/exemplar duplicate-display identity is an additional existing requirement.Compositions.provision/3is currently human-author-only. A dirty visit or “Stale tab” cannot create those identities or by itself trigger an unbuilt editor action.F. Residual risks, ranked
run(composition_id)could write after suppression/configuration change or lose the original parent. Retain refresh authority through the proposed write; the inference fence alone is insufficient.Ranked recommendations and minimum sequence
This is one audit comment only. It records delivered progress and concrete corrections; it makes no changes to the programme’s code, live data, issue bodies, services or acceptance bar.
A labelled control separates the label from packet v2, an id absent f…
The unlabelled plan and the runbook say its packets run only in the b…
holden commentedon Oct 26, 2026
Full MVP infrastructure audit — 26 October 2026, repeat verification
Verdict: P4 and model-free P5 are delivered and form a sound enough foundation to implement the single-editor P6 path with a stubbed runtime, after correcting #264’s incompatible contracts. The MVP is not complete, and live P6 inference remains NO-GO. Grades: P5 A−; P7 B+ for the delivered repairs and honestly reported failed benchmark; P4 A−; reconciliation B−; P6 readiness C+. Main remains
2dc3350626cb2276c8b13819375c01b4af3330bb. The material change since today’s earlier full audit is that the P7 follow-up now exists as open, unmerged PR #266, heada92f66a0018fc12a555e33170c962f09c5e40315, with a committed plan and packets but no measured result. Its direction is reasonable, but its scorer permits definitions instead of the requested quotations and its causal interpretation remains overstated. This is permission to continue bounded engineering under the existing product rules, not evidence that the programme is ready for autonomous agents, panels or public bot output.Evidence and limits
I reread the three original assessments, reconciliation, current #253/#254/#257/#259/#193/#264, supporting #255/#198/#221/#262, delivery evidence, main source/tests, benchmark record/raw results, and PR #266’s plan, packets, scorer and tests. Local and GitHub main agree. The supplied checkout path does not exist on this Mac. I located the existing checkout and reused its isolated audit worktree; no active source or database storage was moved or created.
Own repeat suite: 19 doctests, 2,979 tests, 0 failures; exit 0, ExUnit 404.8 seconds, seed 413141, worktree clean, using exactly
MIX_ENV=test MIX_TEST_PARTITION=_audit3 DD_NO_OBAN=1 mix precommit, with deps/data and every replay gzip symlinked from main. This result covers merged main, not PR #266. I also recomputed main’s benchmark verdict from its raw JSON and exercised PR #266’s scorer in memory with synthetic records; that check called no model and is not a new benchmark.Live SQL was SELECT only, PostgreSQL 18.2, port 5434,
devils_dictionary_v2, withdefault_transaction_read_only=onexplicitly confirmed. Candidate readers started only dependencies and a read-only Repo, not the application/Oban/endpoint. I reproduced the love/war/hope CLI dry-runs and built/froze/verified all ten pilot packets. Public love returned successfully; public Stale returned 404. I did not sign in, remeasure layout, exercise owner controls, rerun historical service drills, migrate, restart/start/stop a service or run inference. Historical approval and the exact number of restarts are documented evidence, not facts recoverable from today’s SELECTs.A. Claimed versus delivered
refresh.ex,refresh_worker.ex,refresh_state.ex, migration20261025200000, visit admission, pilot restriction, transactional job insertion, admission parent, dirty sequence/history, lease recovery/fences, suppression/reopen and disclosure exist. Refresh worker, race and real-connection concurrency tests are present and included in the suite. No model call./ops/curation?tab=staleis 404.devils_dictionary_runtime_bench. Live v2 has 0 local model configs, 0 inference services, 0 attempts. Benchmark registration is not live activation.runtime-bench-v3, v2 packets for five cases including the labelled control, and scorer/tests. The PR reports 2,985 tests; that is author-reported, not my main-suite result.runtime/packet.exusesCandidates.for_page/1; all four kinds, both 12 bounds, required meanings, deterministic quotas/omissions, v1 compatibility, registry revalidation, dry-run and fingerprint transition exist.packet_v2_scope_test.exs, packet/contract and refresh tests cover them.Packet.fits?at context 8,192/output 1,024. This is the conservative estimator, not measured tokenization or inference. Works are real catalog artworks; exemplar-positive coverage is tests because none is live.page.language_tagand “provision both” overlapping scopes. #193 still calls freshness unimplemented and September the only benchmark, with #221 repairs outstanding; #253 §9.12 likewise still describes the repairs/rerun as unfinished. #254/#257 still describe the old built label as unfixed. #254’s P4 checklist still says PR open despite the merge. These are current-document contradictions, not missing implementation.Refresh row snapshot, UTC: row 1, composition/configuration-version 1/1; state idle, generation 3, token 4, admitted parent 1, superseded false; completed 2026-10-26 13:59:58.254863, next check 2026-10-28 13:59:58.254863; outcome needs_editor / fingerprint_version; no pending dirty, lease, failure or retry. Inputs-last-changed is the completion timestamp, reason
inputs_changed. Fingerprint namespace is packet-v2/cut-v1/meanings-v2. This says “inputs checked, editor needed,” not “model has judged the page.”Jobs, SELECT inventory: RefreshWorker 3 completed, IDs 4645/gen1, 4646/gen2, 4730/gen3. CleanupWorker 672 completed; Discovery.RunWorker 62 completed / 2 cancelled; Quotations.VerifyWorker 337 completed. No proposal/inference job exists. These totals remained the same in the verification snapshots.
Activation evidence: receipts 4/5 record P5’s withdraw/republish of love. #259 Delivery B records the named migration, owner’s yes, one direct bootout/bootstrap and matching launch hashes. Current launch-script SHA-256 is still 6704fdf2553d6a97b9c08f9e263e0aec545c21b2674f5e101ab4e177fec05291, matching earlier evidence. I verified resulting state and current bytes, not the original approval conversation.
B. Benchmark and my runtime decision
The 26 October record and raw results agree with my recomputation: all 16 planned calls validate; ordinary selects Bierce 6/6, ambiguous abstains 3/3, sparse abstains 3/3, adverse selects 3/3. Only
ambiguous_each_sensefails. Adverse non-use/non-approval zeros are observed over three answers, unlike September’s unobserved cells. Fixed-seed repeats give one distinct answer per case; sixteen calls are not sixteen independent editorial judgments. Raw swap is zero before/after all calls; the record documents memory before/load/stop and confirmed stop. I did not recreate that session.The NO-GO is correct; a fixture-only causal explanation is not proved. The model explicitly cites fictional definitions, but the same labelled text selected under v1. The supported conclusion is “v2 abstained on these self-labelled fixtures, citing fiction.” Its self-explanation does not isolate the cause. All scored packets in that delivered run are runtime-packet-v1, so it also does not establish behavior under P4’s new meaning bindings.
PR #266 is a reasonable follow-up direction, with three corrections/qualifications before its criteria are accepted:
m0. I loaded this code in memory and supplied a synthetic ambiguous selection lead c1→m0, highlight c2→m0, no quotations. It reported senses_covered: 3/3; with otherwise passing synthetic records, all nine checks returned GO. Its test even explicitly accepts a definition in place of one quotation. The plan discloses this alternative, so it is not a hidden code/plan disagreement; it does differ from the requested quotation test. To measure that test, require both expected quotation candidates c3/s1 and c4/s2 in each accepted selection, and make missing either fail. Keep definitions at their packet-v2 word meanings.every_planned_call_measured,no_validation_refusalandswap_growth_below_pauseinspect totals/outcomes across its calls too. A validated control abstention can be non-gating for quality; a refusal, missing call or unmeasured swap can still fail the run. The sample note also retains “four judgments” although the control adds a fifth case.The quotation restatement does match P6’s v2 task: definitions belong at the word/lexeme meaning; quotations carry their fixed sense. It tests selection and coverage of useful quotations, not free disambiguation, because the allowable meaning is already fixed. It does not itself test works/exemplars or owner-rated curatorial quality. Existing contract tests supply structural coverage for those kinds; the five real proposals and owner forms remain essential.
Own decision: NO-GO for live P6 inference. A GO still requires accepted, precommitted criteria that actually measure the stated task; the owner-approved run on frozen v2 packets with the pinned instruction/model; every declared check passing with observable denominators, resource readings and confirmed stop; and completed P6 implementation/activation gates. Real pilot packet construction/verification/estimated fit now passes. PR #266’s plan or a future merge alone is not a GO. I ran no model.
C. P6 readiness against #264 §§3–4 and #253 §5a, §6a, D4/D8/D9
“Specified only” means requested but unbuilt. “Missing” identifies an absent prerequisite or an executable contract the brief does not supply.
Candidates.for_page/1,Packet.for_page/2; ten actual pilot builds/freezes/verifications; D. It is a declared cut of screen candidates, not the entire screen.Refresh.claim/check/complete; live row 1.RefreshWorker.perform/1completes the input check; noProposalsmodule. Completion releases the claim; a later unfenced composition-only call is insufficient.Authority.actor/2returns the account’s user actor; enabled DB review/publication human guards. Preserve this when permitting model authorship.Profiles.admit/4and enabledcurator_profiles_admission_guardrequire sourced/deceased evidence. Live has only five proposed personas, no generic editor.ConfigurationVersionis manual-only; #264 names enablement without defining it; no live model configs/services./ops/curation/attempts/:idroute. Proposed tab remains disabled.Reviews.decide/4storesreason, notnote. #253 has no completed evaluation forms.D4 remains unambiguous: a bot never moves the published pointer. An explicit owner Take/Edit-publish can do so under the existing human review and compare-and-switch gates.
D. Owner’s pages and candidate supply
Live record: 1 composition, 1 version, 1 review, 5 publication receipts. Only On love, composition/version 1/1, is published; final receipt 5. Selector/publisher is actor 5, human user 2, label Account #2. No model version or attempt exists. War and hope have no composition identity yet.
The three placeholder notes are unchanged: “Bierce’s entry leads: the voice this dictionary is named for”; “Johnson’s senses map the word before Bierce turned it”; “The plainest literary use of the sense, beside the satire.” P5’s withdraw/republish is not the missing P2 three-page owner screen exercise. #257 should stay open.
Counts below come from today’s read-only Candidates/Packet run. Each kind is eligible on screen → kept in packet. Leads overlap content. Meanings show kept/total; omitted.
Unused quota flows forward, explaining more than three packet quotations. All retained candidates keep their required meaning. No candidate here was specifically excluded for meaning bounds; separate unused meanings were trimmed. Works are catalog artworks, not selectable poems/books; no accepted exemplar appears.
Reproduced frozen hashes: love
7a6061f8caa4ea5be3a248132d2f88771fb76c1510d61f2ad468502ddf75ec73; war1fc7c48c86e30892a3ff2685a59392e02f1aea298daae8586c525988aa3fe1a5; hopef21bb3b93a17a1774c3745ee7103e1a3686dc021e2984d6394e64cf3548bbca2.Love, war and hope are the richest first three proposal pages, with love the smallest first activation because its identity already exists. P5 cannot admit a never-provisioned page merely because it is on the allowlist.
E. Exact corrections to #264 before implementation
These repair the requested slice; they do not add features.
generic-editor v2 · qwen3.5-9b-v1. Delivered benchmark binding is generic-editor-v2 / qwen3.5-9b-v2. Mark counts illustrative; love currently offers 10 of 32, not 12 of 36.outcome/3computes the pair internally; the worker completes and returns:ok. Specify the integration using row id, configuration version, generation, token and admitted parent, including consumption of an already-fresh persisted needs_editor handoff. An ordinary repeat visit currently admits nothing until dirty/due.run(composition_id)assumption with an editor step that retains or reacquires explicit refresh authority and fences every proposal/completion write against suppression, generation/token/lease and active configuration. Runtime’s inference fence is not P5 ownership.Profiles.admit/4cannot admit a no-dossier generic editor under present code/DB rules. Include a narrow generic-editor admission rule in the scoped migration; preserve persona dossier/deceased-evidence rules and human admission authority.editorial_composition_versions_human_actor. Specify the narrow bot-author/attempt/config guard while keeping review/publication/activation authority human.last_outcome: abstainedis rejected by both RefreshState’s enum and the live CHECK. Specify the schema/disclosure/editor-outcome contract. Remove “or abstained” from the version-write branch: abstention writes no proposal and counts toward none of the five; refusal/error must not acquire successful editorial freshness.:kept_item_droppedforever does not implement D8’s “survives into every bot proposal”; do not silently change P4 policy.reason, nonote;Reviews.decide/4requires an idempotency key. Specify fixed-form serialization in reason or the explicit storage/API change, bound to the reviewed version/items. Resolve “form each time” against tests requiring it for Take/Edit; Dismiss must retain its required rejection reason.proposal:<composition>:<generation>needs one frozen packet/config identity and exactly-once version/reservation writes. Changed inputs under one key are a conflict/stale case, not an identical replay. Bind dismissals to arrangement and packet, enforce one proposal per accepted attempt and retain the proper generation on retry/recovery.one_display_identity/1. Distinct candidate ids alone do not prevent duplicate displayed works/exemplars.F. Residual risks before agents
Ranked recommendations and smallest route to the first real proposal
The minimum dependency sequence is contract corrections → stubbed P6 implementation/review/merge plus benchmark GO → separately approved live migration/profile/model/authority activation → one love proposal → explicit owner decision. War/hope identity provisioning and the remaining owner acceptance work follow the same existing seams. No new feature is needed to reach that first proposal.
This is one report comment. No code, issue body/checklist, live record or service was edited; test activity was confined to the authorized private partition/worktree.
The ambiguous check requires each sense's quotation under its sense, …
The unlabelled re-run: go for P6, all nine checks pass; crane selects…
The unlabelled run's record claims what one judgment shows, and names…
The unlabelled benchmark re-run: go for P6 (#254 P7) (#266)
holden commentedon Oct 26, 2026
P7: the final runtime verdict for P6 is go (26 October 2026)
This supersedes the morning's no-go above. #266 (merged as 9ccc427) re-ran the plan with three changes:
house-glossary);The instruction (
generic-editor-v2) and the model (qwen3.5-9b-v2) were unchanged. The owner approved the run, including the both-quotations rule for the ambiguous check that the audit called for.Result: all 9 checks pass. Run 1793029211, attempts 52–70: 13 selections, 6 abstentions, 0 refusals.
Record: docs/curation/runtime-benchmark-2026-10-26-unlabelled.md.
What the go does not cover, per the record:
The service is stopped. Live inference continues under #267.
holden commentedon Oct 26, 2026
The remaining path (#267, then #264, then the hand-off to the persona phase) is now ordered in one place: #268. It defines "ready to build the agents" as six checkable items, runs #267 and #264 Delivery A in parallel under a stated file ownership, and asks for an independent audit before Stage 0 (correcting #264) starts. It adds no feature and changes no part's brief here.
holden commentedon Nov 1, 2026
Closed 1 November 2026: P1, P2, P4, P5, P6 and P7 are merged and live; P3's PoetryDB pin was superseded by the #271 owner amendment of 27 October (poems are shelf items cited by their stored observation). The implementation tracker is now #298 under #193.