Skip to content

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

Closed

Description

@holden

Reconciled 25 October 2026 with the independent assessment on #254: all fifteen of its section C edits, accepted by the owner, and two new decisions, D8 and D9 (§12). Facts checked against main at e25dd94 and devils_dictionary_v2. The list of changes is the reconciliation comment on #254.

Revised 21 October 2026 (evening) after an independent review (B+: A− direction, B− implementation readiness). Six contracts were sharpened: the freshness clocks (§7), owner decisions versus bot refresh (§6a), the pin (§9.6), proposal attribution (§5a, §9.7), an editorial quality gate (§10) and candidate trimming (§5a). The pilot decisions now have defaults (§12). The implementation plan with per-part briefs is #254.

Delivery child of #193, written 21 October 2026 at the owner's request: one place that says exactly what the curation agent is, what a curated entry contains, what the page looks like with it, what the owner can do by hand, how bots do the same through the same record, how an entry stays current, and what is missing before any of it runs. Everything here is checked against main at dc888b7 and the live corpus (devils_dictionary_v2). Where this issue and an older child (#156, #196, #197, #198) differ, this issue wins; the older ones keep their detail.

1. The picture in one paragraph

Every word page keeps its definitions, examples and discovery shelves exactly as they are. Above them sits a curated opening: one sourced lead definition, and a handful of sourced highlights, each with a one-line reason. The opening is a composition: a versioned, reviewed, published record that points at things the registry already holds, never at text someone typed in. The owner can make or change a composition by hand on a screen. An editor bot makes the same kind of composition from the same candidates, through the same record, and says why. Later, five persona editors propose and vote; the result publishes itself when they agree. On every visit the page checks when it was last curated and whether its inputs have changed, and queues a re-curation when it is stale. Nothing is manufactured to fill a slot: an entry with nothing worth showing shows nothing.

2. Rules the system already has, which this issue keeps

These are in code and tests today (PR #206, #212, #195). The agent does not get to argue with them.

Rule Where
Bierce leads. Where the page carries a Devil's Dictionary entry, only that entry may lead. It is "an editorial preference, a choice of voice, not a claim that the entry is literally true". Where there is none, another sourced definition from the page leads, or the lead is empty. Curation.LeadPolicy
A selection is a reference, never text. Every item points at a registry object and the exact revision it was read from (a definition's content revision, a sense quotation's words by hash, a catalog row by manifest and checksum, an accepted illustrates claim). Nothing is keyed by a discovery result id. Curation.CompositionItem, Curation.References
Honest empty. A withheld item is withheld, not replaced. Two highlights are two. No lead and no highlights means no section at all. Curation.Opening, Curation.Eligibility (R6)
Eligibility is re-checked at read time. A superseded revision, a withdrawn claim, a disabled source, changed words at a locator, a drifted catalog row: each withholds the item with a named reason. Curation.Eligibility
A composition is not a claim. Selecting a poem does not assert that the poem illustrates the meaning. An illustrates nomination is a separate claim with its own review (#181, #190). #196, #212
The model sees a frozen packet of at most 12 eligible candidates (and, separately, at most 12 meanings) with excerpts cut at word boundaries, under an instruction that says they are material, not orders. It answers a JSON contract: select or abstain, a lead, up to three highlights, each with a candidate id, a meaning id, a reason of at most 280 characters and, for a quotation, an exact substring. Every answer is re-validated against the registry before it becomes anything. Runtime.Packet, Runtime.Contract
One private local model, one slot, a daily budget (30 minutes of inference), a ledger of every attempt. Personas share the model; there is no model per persona. Curation.Runtime, stage A record
Refresh on visit, never in the request path, with a per-source policy of refresh and retention windows. This is how discovery shelves work and it is the pattern curation copies. Discovery.Policy

3. What a curated entry contains

A composition for a page (a scope of one or more lexemes in one language) has:

The lead. Exactly one sourced definition that the page already shows, or none. Register: satire when it is Bierce, definition otherwise. This is the "canonical definition from somewhere" the owner asked for: it is canonical for this page, it is attributed, and it is one click from the full entry on its card.

The highlights. Up to three in the merged record (Opening.max_highlights/0). Each has a kind, a reason, and an intended meaning (which sense of the word it speaks to):

Kind What it is Where candidates come from today Example
content Another sourced dictionary entry on the page Entries the record holds as content items (Bierce and Johnson on the pilot pages). WordNet and Wiktionary file senses, not content items, so they are never content highlights For love: Johnson's LOVE, n. s. beside Bierce's lead (live since 23 Oct)
sense_quotation A quotation attached to a current sense, pinned by the hash of its words Wiktionary citations; the verified public-domain Wikiquote corpus (1,053 lines over 197 concepts, wikiquote-pd-v1) A Johnson-era citation under the sense it illustrates
work A poem, passage, artwork, book or film that is a durable registry object or a committed catalog row Catalog artworks matched to the page by depiction (Artworks.shelf_items/1 over met-highlights-v1 and wikidata-famous-v1). Catalog poems and books (poetrydb-v1, open-library-v1) can be stored and rendered but have no page-match seam (§12 D6); not yet anything a discovery provider returned (§9) For war: one of its two catalog artworks
exemplar A thing that represents the meaning: an accepted illustrates claim with its evidence Accepted illustrates nominations only (#181). The instances a page lists (44 instance_of claims under war) are not exemplars, and a pending nomination is not one until accepted For nepotism: its one nomination, once it is reviewed and accepted

What the record can offer today (verified on v2, 25 October 2026): 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. Of the ten pilot pages, 8 have an admissible lead (greed and bestiality none), 0 have an accepted exemplar, and 3 have catalog artworks (war 2, love 4, hope 1). The two decisions this leaves are D3 (a sense as lead) and D6 (the poem-to-page rule) in §12.

The owner's three asks map onto these: one canonical definition is the lead; examples of things representing greed are exemplar highlights; a poem that relates is a work highlight.

The why. Every item carries a note of at most 280 characters, with its author kind (human or model) and label. The page shows it as "Why this is here".

Provenance. The configuration and version it was made under, who selected it (the owner, or an editor profile and model digest), who reviewed it, when it was published, and the eligibility fingerprint it was accepted with.

D1, decided for the pilot: three highlights. That is what the record enforces (Opening.max_highlights/0). Five is a later change with the page measurement #111 asks for.

4. The page

4a. Today

A word page as served now, for example /words/198692/nepotism or /on/greed:

┌────────────────────────────────────────────────────────────────────────┐
│ nepotism  /ˈnep.ə.tə.zəm/  noun        4 dictionaries · 2 senses       │
│ forms · origin                                                         │
│ Defined here by: Samuel Johnson · Ambrose Bierce · WordNet · Wiktionary│
├────────────────────────────────────────────────────────────────────────┤
│ Definitions   (4 dictionaries · one open at a time)                    │
│   ▸ Samuel Johnson 1755 · "Fondness for nephews…"                      │
│   ▸ Ambrose Bierce 1911 · "Appointing your grandmother to office…"     │
│   ▸ Open English WordNet 2025 · "favoritism shown to relatives…"       │
│   ▸ Wiktionary · "The favoring of relatives…"                          │
├────────────────────────────────────────────────────────────────────────┤
│ Examples      1 under review                                           │
│   Browser audit Family Table 84 · needs review · cited as an example…  │
├────────────────────────────────────────────────────────────────────────┤
│ Urban Dictionary  (browser-fetched, nothing stored)                    │
├────────────────────────────────────────────────────────────────────────┤
│ Out in the world   17 things · every match says why it is here        │
│   Films 5 · Images · Texts · News 12 · Music · GIFs                    │
├────────────────────────────────────────────────────────────────────────┤
│ Subjects · Related words · Sources                                     │
└────────────────────────────────────────────────────────────────────────┘

4b. With the curated opening (MVP)

Nothing below the opening moves. The opening is the first thing after the headword. Rendered by the existing Opening component (#156 Phase 1), which already draws this shape from a fixture.

┌────────────────────────────────────────────────────────────────────────┐
│ nepotism  /ˈnep.ə.tə.zəm/  noun        4 dictionaries · 2 senses       │
├────────────────────────────────────────────────────────────────────────┤
│ ┌── THE DEVIL'S DICTIONARY · 1911 ─────────────────────────── satire ┐ │
│ │ NEPOTISM, n. Appointing your grandmother to office for the good   │ │
│ │ of the party.                                   Read the whole entry│ │
│ │ Editorial preference: where The Devil's Dictionary defines a word, │ │
│ │ its entry leads. Leading is a choice of voice, not a claim…         │ │
│ └────────────────────────────────────────────────────────────────────┘ │
│ ┌ EXAMPLE ───────────┐ ┌ QUOTATION ───────────┐ ┌ ARTWORK ───────────┐ │
│ │ [thing]            │ │ "…" — author, work,  │ │ [image]            │ │
│ │ Title · source     │ │ year · source        │ │ Title · artist ·   │ │
│ │ for sense ● 1      │ │ for sense ● 1        │ │ year · The Met     │ │
│ │ Why this is here:  │ │ Why this is here:    │ │ Why this is here:  │ │
│ │ one line, signed   │ │ one line, signed     │ │ one line, signed   │ │
│ └────────────────────┘ └──────────────────────┘ └────────────────────┘ │
│ ▸ How this was curated · 21 Oct 2026 · by the owner · checked today    │
├────────────────────────────────────────────────────────────────────────┤
│ Definitions … (unchanged from 4a)                                      │

Opening the disclosure:

│ ▾ How this was curated                                                 │
│   Configuration global-default v1 · lead rule bierce_first_v1          │
│   Selected by: Holden (operator) · 21 Oct 2026 14:02                   │
│   Reviewed: operator publication · version 3 of 3 · history ↗          │
│   Checked against the record: today 09:14 · nothing withheld           │
│   Inputs last changed: 19 Oct 2026 (a new quotation under sense 1)     │
│   Next check: after 23 Oct 2026, or on the next change                 │

When a bot made it, the same block says Selected by: <pinned instruction label> (qwen3.5-9b, digest …) · proposed 20 Oct · published by the owner 21 Oct, and later Decided by the panel: 4 of 5 (Bierce, Voltaire, Hitchens, Le Guin for; Vonnegut against: "…").

When something is withheld since publication:

│ ┌ QUOTATION ───────────┐                                               │
│ │ Withheld: the words at this locator changed since review.          │ │
│ └──────────────────────┘                                               │

4c. A page with no Bierce entry: greed

/on/greed today: two dictionaries (WordNet with two senses, Wiktionary), no Johnson, no Bierce. Bierce has no GREED or AVARICE entry; his nearest is MAMMON, which is a different word and therefore not on this page's scope. The subject greed (Q12819497) is unclassified with no family address. The shelves show 12 films, 15 images (one of them Commons' The Worship of Mammon), 10 texts, 13 news items; artwork and verified quotes are empty.

What the MVP can make of it, honestly: no lead (there is no Bierce or Johnson entry, and WordNet and Wiktionary senses cannot lead, §12 D3) and quotation highlights from its senses. P2's test demonstration (#258) published exactly that: no lead and one 1885 quotation.

│ ┌ QUOTATION ───────────────────────────┐                               │
│ │ "…" — the quotation's source, year   │                               │
│ │ under sense ● noun 1                 │                               │
│ │ Why this is here: one line, signed   │                               │
│ └──────────────────────────────────────┘                               │

The Worship of Mammon is not a candidate: no catalog artwork is matched to greed, and the Commons image on its shelf is a discovery result. The films and the Pexels photos are not candidates either: they are discovery results, refreshed and discarded, not registry objects. This is the single biggest gap between "what the page shows" and "what a curator may choose", and §9 names it.

4d. A page rich in examples: war

The record names 44 instances (Peloponnesian War, Balkan Wars, Crimean War…), but they are instance_of claims, not accepted exemplars, so none is selectable as it stands. An exemplar highlight needs an accepted illustrates nomination first, made and reviewed through the existing Examples flow (#258's test demonstration did this for the Peloponnesian War, on a test partition only); its reason then names the sense ("an instance of sense 1, a conflict between states, not sense 2, a campaign against a condition"). Today war offers its Bierce entry as lead, Johnson, sense quotations and two catalog artworks.

4e. Phone

Same order, one column: lead card full width, highlights stacked, disclosure collapsed. The opening's height is measured the way #111 measures (every <details> closed); #212 Build 3 measured +279 px desktop, +431 px phone for one exemplar behind the gate. The MVP budget: the opening may not push the Definitions heading below the first screen on a 1280 × 900 desktop.

5. The editors

5a. MVP: one generic editor, the owner publishes

generic-editor-v1 is the instruction the 27 September benchmark ran. On the pinned qwen3.5-9b-v1 it gave the Bierce lead 6 of 6 times on ordinary pages, put each quotation under its own sense on ambiguous pages, abstained on sparse pages, and had 3 of 16 answers refused by validation on adversarial packets (it did not follow injected instructions). Warm latency about 3.2 s at p50. The 4B model abstained on everything and is out.

The benchmark's condition stands: revise the instruction as a new pinned version and re-run the plan before relying on it. #221's three runtime fixes (a stale dispatch raising instead of refusing, orphaned Ollama runners on stop, two vacuous benchmark cells) come first.

What the editor does, per page:

  1. The scheduler (§7) hands it a page whose composition is missing or whose inputs changed.
  2. Runtime.Packet.build/3 gathers the eligible candidates in a fixed order: Bierce entries, other definitions, sense quotations. MVP extension: work and exemplar candidates join the packet (today it builds only the first three kinds). The cap of 12 stays. Today more than 12 is refused ({:too_many_candidates, n}), never trimmed; the MVP replaces the refusal with an explicit trimming policy, versioned with the packet (runtime-packet-v2):
    • every applicable Bierce entry is included and counts against the content quota (the lead rule needs it; :priority_candidate_unavailable stays a refusal). If the applicable Bierce entries alone exceed that quota or a packet bound, the packet is refused, never trimmed: unconditional inclusion and a hard bound cannot both hold, so refusal is the default (P4 may put another rule to the owner);
    • then quotas by kind: up to 4 content, 3 sense_quotation, 3 work, 2 exemplar, filled in source-priority order (Bierce, Johnson, then any other entry the record holds as a content item; verified corpus before live; accepted claims by review date);
    • meanings are bounded as well as candidates. Packet v1 refuses more than 12 meanings independently of its 12-candidate cap ({:too_many_meanings, n}), and love's whole On scope (five lexemes) has 43 senses, 48 packet meanings with the five lexeme meanings (36 counts the lowercase love subset only), so trimming candidates alone cannot make a whole pilot scope pass. The packet keeps every chosen candidate's required meaning, bounds the rest deterministically, and is tested on real whole On scopes;
    • candidate ids, revisions, rights and claim standing, and each omitted candidate's reason are frozen deterministically;
    • ties broken by registry object id, ascending, so the cut is deterministic;
    • the omitted candidates are listed on the attempt's packet_summary by id and reason (:over_quota), so an audit can see what the editor never saw;
    • a candidate that becomes eligible outside the subset does not by itself make the composition stale; the fingerprint covers the packet that was sent, and a periodic refresh (§7, refresh_generation) is what re-cuts it.
  3. The model answers the contract. Validation re-checks every reference.
  4. A proposed composition version is written with origin: :model (a new enum value beside :manual), attempt_id pointing at the validated attempt (new column on editorial_composition_versions), created_by_actor_id the editor profile's bot actor, and each item's note signed note_author_kind: :model, note_author_label the actual pinned instruction label and model tag (P7's revised instruction, never a hard-coded generic-editor v1), note_author_actor_id the bot actor. Nothing is published. The manual writer's rule that attribution is never an input stays: the proposal writer is a separate function that takes the attempt, not caller-supplied authorship.
  5. The owner sees it in the queue (§6). Attribution on take and edit: "Take" creates a new version with parent_version_id the proposal, origin: :manual, created_by_actor_id the owner. An item copied unchanged keeps its model note and its model attribution; an item whose note the owner rewrites gets a human note attributed to the owner, and the model's original note is kept on the proposal version, reachable through parent_version_id. An item the owner adds is human-authored. So the page can always say which words are the editor's and which are the owner's, and "Take" can never relabel model text as the owner's.

That is the MVP: bot proposes, owner disposes, one record. It is honest about where the programme is (no panel decision authority exists yet) and it gets curated pages on the site within weeks.

5b. Then: five persona editors and a panel

The roster is decided (#97, #201): Bierce, Voltaire, Vonnegut, Hitchens and Le Guin-inspired profiles, all deceased, each with a dossier grounded in permitted original writing and deceased_evidence_refs. The tables exist (curator_profiles, curator_profile_versions), empty.

How a persona works, concretely:

  • Same model, same packet, same contract. A profile adds a dossier to the instruction: what this editor values, what it refuses, how it reads a definition, in its own words with source refs. No fine-tuning, no model per persona.
  • A run on a page: each of the five answers the packet independently (five calls, serialised through the one slot). Round 1 proposals; round 2 each votes on the arrangement with the most support; at most one repair call. S(c) = count(final_ballot == support); publish when S >= 4. A missing answer is never an abstain ballot.
  • Each persona's reason is kept and shown: "Vonnegut against: a deadly-sin reading is the pulpit's, not the dictionary's."
  • The panel's finalised decision is the panel_decision publication authority (Curation persistence: versioned panels, ballots and reviewed page compositions #196's intended routine path). No per-selection human sign-off; the owner's veto, override and withdrawal remain.
  • A profile is admitted by the owner (Profiles.admit/4), gets a bot actor, and its nominations of illustrates claims go through Gate every illustrates nomination until accepted before bots or the public can nominate (#181 decision 6, deferred) #190's gate like anyone's.

What the five would do differently on greed is the point of having them, and it is also why the MVP starts with one: the panel multiplies every run by five calls and two rounds, and the September programme stalled on building it before anything was on the page.

5c. Expectations of any editor, human or bot

  • Never invents. Chooses from candidates; a quote is an exact substring of its candidate.
  • Names the meaning each item speaks to.
  • Gives one reason per item, under 280 characters, in plain language a reader can judge.
  • Abstains when the candidates are thin. An abstention is a result, logged, and the page shows nothing.
  • Respects the lead rule without being asked.
  • Is reproducible: same configuration version, packet hash, instruction version and model digest give the same attempt identity.

6. The owner's curation screen

Operator-only, behind login, never through the tunnel (it joins /ops). Wireframe:

┌ /ops/curation ─────────────────────────────────────────────────────────┐
│ Queue  [pilot 10] [published 1] [withheld items 0] [stale] [proposed]  │
├────────────────────────────────────────────────────────────────────────┤
│ greed        never curated · quotations only, no lead       [Curate]   │
│ nepotism     proposed · <pinned instruction label> (P6)     [Review]   │
│ love         published v1 · last assessed (P5) · 0 withheld [Open]     │
└────────────────────────────────────────────────────────────────────────┘

┌ /ops/curation/on/greed ────────────────────────────────────────────────┐
│ Lead                                                                   │
│  No admissible lead: Bierce and Johnson have no entry, and WordNet and │
│  Wiktionary senses cannot lead (§12 D3).                               │
│ Highlights (choose up to 3)                                            │
│  [ ] quotation  "…" (1885) · under sense noun 1                        │
│  [ ] work       none: no catalog artwork matches this page             │
│  [ ] exemplar   none accepted                                          │
│  Why this is here: [________________________________] 0/280            │
│ Proposal from the editor (P6): one quotation — "…"          [Take]     │
│ [Save draft]  [Publish]  [Withdraw current]   History: v1 ↗            │
└────────────────────────────────────────────────────────────────────────┘

What it does, all through existing functions: Compositions.provision/3, create_version/3, Reviews.decide/4, Publications.publish/3 and withdraw/3. "Take" copies the bot's proposal into the form under the attribution rules in §5a step 5; the owner can change any item before publishing, and the published version records the proposal (as its parent) and the override with its reason.

Storage this needs: origin: :model and attempt_id on versions (§5a), one new table for freshness (§7), and the owner's marks (§12 D8, P6's). Nothing else.

6a. Owner decisions and bot refresh: the rules

A refresh never silently undoes an owner's decision, and a stale result never replaces newer work. Concretely:

Situation Rule
The owner published version N by hand; a bot runs later The bot writes a proposal (a reviewable draft, origin: :model) with parent_version_id N. The published pointer does not move. The queue shows "new proposal for a page you curated"; the owner takes, edits or dismisses it. Dismissal is recorded (Reviews.decide/4 with :rejected) and the same proposal (same arrangement hash on the same packet hash) is not re-proposed.
The inputs of an owner-curated page change The page is stale (§7). A bot may propose; it may not publish. Read-time eligibility still withholds any item that became ineligible, immediately, with its reason.
The owner withdraws the published version Publications.withdraw/3 means keep this off the page. The composition's refresh state is set to suppressed with the withdrawal's reason; no bot proposes for it until the owner re-opens it ("curate again") from the queue. Withdrawal is never a request to regenerate.
The owner publishes while a bot run is in flight The run finishes against its frozen packet (it holds the inference slot under its fence) and writes a proposal whose parent is whatever was current when the run was admitted (captured at admission, not at worker start). On completion the writer compares the composition's current published version id with the one it was admitted against; if they differ, the proposal is marked superseded_on_arrival and goes to the queue flagged, never auto-taken.
Two publications race (two operator sessions) Publications.publish/3 already compares-and-switches on the previous version id; the loser gets {:error, {:stale_pointer, current}} (a draft from a moved parent, {:stale_parent, latest}).
A recovered or late worker writes after losing its lease Refused by the runtime's fence ({:error, :stale} from Gateway), and by the refresh state's refresh_generation check.
The owner marked an item kept or the page mine (§12 D8) A kept item is carried into every bot proposal until unmarked. A page marked mine still receives proposals, shown in the queue, never auto-taken.
Unique job per scope Is a coalescing optimisation only. The correctness boundary is the refresh-state row's lock, generation and the publication's compare-and-switch, as #198 §"Admission and freshness" and §"Publication and reader stability" specify. Those paragraphs are carried into this issue unchanged.

An open page stays pinned to the version it rendered; a new publication applies on the next navigation, never mid-read. For withdrawal and lost rights, fresh loads revalidate; prompt suppression on already-open pages still needs an explicit bounded invalidation path and proof. Until one exists, no part claims that an open page drops a withdrawn item.

7. Staying current

Copy discovery's design, with its own queue and budget, and #198's clocks, which this section restates rather than replaces. The earlier draft decided freshness by publication age; that was wrong, because a successful assessment can leave the published version unchanged or produce an abstention awaiting review, and neither must re-run inference on the next visit.

One row per composition and configuration version (new table curation_refresh_states):

Column Meaning
last_completed_at the last successful assessment, including an honest abstention or a valid-but-unchanged result. Completion is not publication.
refresh_after when a periodic reassessment is due: last_completed_at + the window (48 h default; 24 h or 7 days per configuration version; an abstention or empty packet gets the empty window, 48 h initially).
next_retry_at failure backoff only: offline runtime, refused dispatch, deferred evidence. Starts at one minute, doubles with jitter to six hours. A failed inference never advances last_completed_at.
input_fingerprint the eligibility fingerprint the record already computes, extended with the packet summary hash (candidate ids and revisions, never excerpt text). Provider request timestamps are excluded, so unchanged evidence never triggers work.
dirty_at set when an input changes (a withheld item at read time, a new eligible candidate inside the packet cut, a claim review, a source revision, a configuration activation), coalesced with a 60-second debounce. Cleared only by a completion that saw the change; a change arriving during a run stays pending for the next. inputs_last_changed_at keeps the date after it clears.
refresh_generation monotonic, allocated per admitted run; retries keep the generation. Separates intentional freshness runs from exact-input duplicates.
state idle, queued, running, suppressed (withdrawn by the owner), deferred (budget or backoff).

On visit (every page visit, plain HTTP render or LiveView mount, after the page is built): a bounded read returns the current published composition. Then one bounded check (an indexed read, at most one admission per composition in flight), with no inference in the request path: admit a run when the page is on the pilot allowlist and has a composition, and (dirty_at is set, or refresh_after has passed, or there is no row), and next_retry_at has passed, and the state is not suppressed, queued or running. Admission locks the composition and the row, captures the published parent, allocates the generation and inserts one Curation.RefreshWorker job in the same transaction. A fresh visit makes zero inference requests; repeated stale visits cause one refresh.

A never-curated allowlisted page has no composition, so it has no refresh identity and admits nothing. It gets one when a composition is provisioned: the screen's first Save draft (Compositions.provision/3) in P5, the editor's bot actor in P6. Its refresh row is initialised on the first admission, under the composition's lock.

What a run does: build and freeze the packet; if its summary hash equals input_fingerprint's packet part and the run is not a periodic one, complete without calling the model (last_completed_at advances, nothing else). Otherwise call the editor (§5a), write the proposal, complete. Until P6, a run stops at needs_editor: the inputs were checked, not judged. The published pointer never moves here.

What the page says, in the disclosure, with distinct words for distinct things: published (the version's publication date), last assessed (last_completed_at; until P6 it says the inputs were checked, not judged by an editor), checked against the record (read-time eligibility, which happens on every render and is not an assessment), inputs last changed (inputs_last_changed_at, kept after the pending dirty_at clears), next assessment after (refresh_after).

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. Both are global, not per configuration. The queue drains dirty before periodic, oldest first.

Scope: pilot allowlist only during MVP; later expansion disabled. Never a corpus-wide scan; widening is the owner's later decision (D7).

The fingerprint is versioned. P5 uses P2's full candidate-spec digest (candidates-v1); P4's packet digest becomes packet-v2. A stored fingerprint of another version never equals a new one, so the switch costs one needs_editor run per page and never a silent match.

The P5 contracts (the published parent captured at admission; superseded_on_arrival persisted; dirty changes during a run preserved; inputs-last-changed history kept; absent state initialised under a composition lock; bounded lease recovery; late writes refused after suppression or a configuration change) are specified in #259 §5.

8. What exists today

This table is as of dc888b7 (21 October). Since then P1 (#255, PR #256) is on main and love (composition 1) has been published live since 23 October; v2 holds one composition. P2 Delivery A merged at e25dd94; owner demonstration/Delivery B not evidenced. #198 is reopened as P5's contract (#259). The runtime and freshness rows are unchanged.

Piece State on main On the live database
Configurations, profiles, compositions, versions, items, reviews, publications, Published.current/1 Built and tested (PR #206) All tables empty; mix dd.curation.seed never run on v2
Lead rule, eligibility, references, excerpts Built —
Opening struct and component Built (Phase 1) Renders only behind ?opening=fixture from a hand-written fixture
Runtime: model configs, one slot, packets, contract, attempts, ledger, mix dd.runtime Built (stage A), benchmarked 0 attempts; 9.3 GB of models on the drive; service stopped when idle
Discovery's visit-driven refresh Built, running The pattern for §7
Examples backbone (instances, nominations, person-subject gate) Built (#181) war names 44; nominations invitation-only
Artwork corpus, verified quote corpus Committed manifests 3,212 works; 1,053 lines over 197 concepts

9. What is missing before the MVP works

In the order the phases need them:

  1. Nothing is seeded or active. Seed, activate version 1. Done in P1 (Curation P1: show one manually published opening on a pilot word page #255).
  2. No page reads a composition. Wire Published.current/1 into WordLive as the default reader. Done in P1.
  3. No screen. §6 does not exist; the operator path is functions only. Built: P2 Delivery A (Curation P2 (#257): the owner's curation screen and example tiles #258).
  4. No freshness clock, queue or worker for curation (§7). Discovery has them; curation has none. P5, 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. The packet knows three candidate kinds. Works and exemplars must join Packet.build/3 for a bot to pick an artwork or an accepted example; a poem or book needs the page-selection relation first (D6). Meanings need a bound too (§5a).
  6. Discovery results are not candidates, and the pin needs its own contract. A film from CineGraph, a poem from PoetryDB, a photo from Pexels is fetched, cached and expired; it is not a registry object and cannot be referenced by a composition item. The MVP adds a pin, with these rules, for one provider first, PoetryDB, end to end:
    • Identity: 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. The author resolves through the reviewed author-QID crosswalk (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. The provider already exposes a durable identity adapter (Poetrydb.identity_record/1), so a pin first checks whether the poem is already materialized before claiming it creates a new work.
    • What is stored: a source record under the provider's source (poetrydb) with a revision holding the poem's title, author, line count and the lines themselves, because PoetryDB serves public-domain texts and its terms permit retention. For a provider whose terms do not permit retaining the material (Pexels, Unsplash, GIPHY, Bing News), the pin stores identity and metadata only and the card renders through the provider's own embed or link; such a pin is a work with no retained text or image. The provider registry's existing display_allowed and retention policy decide which, per provider, in configuration, not per pin.
    • Attribution and rights: the card shows the provider, the author and the licence the provider states; a pin records the licence string and the fetch time. A later rights change on the source record (display_allowed false, source disabled) withholds the item at read time like any other.
    • Revisions and disappearance: because the identity hashes the lines, re-fetched text that differs is a different identity today. Whether it may supersede the pinned work (withholding the item :revision_superseded until re-reviewed) is the explicit supersession decision above, not an assumption. If the provider no longer returns the identifier, the record keeps its last revision and is marked unavailable_at; the item stays eligible only while retention permits, and the card says the source no longer serves it.
    • Deletion: a pin can be retired by the owner; retirement nulls the references (R7) and withholds the item. It is never hard-deleted while a version references it.
    • Who may pin: the owner, from the shelf card ("Pin for curation") and from the screen. An editor may propose a pin (a candidate outside the registry) only in a later phase, under Gate every illustrates nomination until accepted before bots or the public can nominate (#181 decision 6, deferred) #190's gate, never silently.
    • Done looks like: one PoetryDB poem pinned for a pilot page, visible on the shelf as "pinned", chosen as a work highlight, published, withheld after a simulated text change, restored after re-review. It also needs the screen's Candidates listing to offer the pin and a source-record work renderer: main still withholds non-catalog work items as :unsupported_item, so storage alone cannot meet the public round trip.
  7. No proposal path. A validated model answer stops at the attempt; nothing turns it into a composition version. Needs origin: :model, attempt_id, the proposal writer, and the attribution rules of §5a steps 4–5. selection_origin gains :model.
  8. No run, ballot or decision records, and no panel_decision authority (Curation persistence: versioned panels, ballots and reviewed page compositions #196, Persona panel pilot: five versioned editors, fixed quorum and ten-page evaluation #197). Not needed for the MVP; needed for 5b.
  9. The instruction needs its revision and re-run (benchmark condition) and Runtime stage A follow-ups from the #210 audit: stale dispatch refusal, orphaned ollama runners on stop, two vacuous benchmark cells (#195) #221's fixes.
  10. Coverage. Most pages have no concept link, so no quotes, artworks or exemplars reach them: sense-level shelves show on about 1 percent of words. 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 is the parallel track that makes curation worth more than the pilot.
  11. Memory on the Studio. Loading the 9B model under current pressure grew swap. The service stops when idle, which is fine for a pilot; a scheduled bot wants the mini or a third machine (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).
  12. Visit-driven curation refresh: coalesced jobs, fenced leases and atomic publication #198 is reopened as P5's contract (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); Runtime stage A follow-ups from the #210 audit: stale dispatch refusal, orphaned ollama runners on stop, two vacuous benchmark cells (#195) #221 is reopened for P7 (PR Runtime #221 fixes (P7 of #254, PR 1): stale refusals, the retry ceiling, confirmed runner stops, observable adverse counts #260 carries A1, S1 and B1; the instruction revision and the 9B re-run follow in a second PR). Neither is done. Visit-driven curation refresh: coalesced jobs, fenced leases and atomic publication #198's admission, fencing and publication paragraphs are the contract §6a and §7 restate; closing Visit-driven curation refresh: coalesced jobs, fenced leases and atomic publication #198 after a model-free P5 would wrongly imply its crash, recovery and proposal obligations are met.

10. The MVP, stated as acceptance

The MVP is done when all of these are true on the live site:

  • Ten pilot pages (§12 D2) show a curated opening to the public, each with a lead under the lead rule and at least one highlight with a signed reason, or an honest nothing where the record is thin.
  • The owner curated at least three of them by hand on the screen in §6, and withdrew and re-published one, under the owner's public actor label (today "Account Bump image_processing from 1.12.2 to 1.13.0 #2") and with the accepted placeholder reasons on love replaced.
  • The generic editor proposed compositions for at least five of them from real packets. Quality gate: the owner reviewed all five on a fixed form and recorded, per item, relevant to the word, correct meaning, reason useful to a reader, not a duplicate of the lead or another highlight, attribution correct. A proposal is taken unchanged only when it deserves it; the form, not the count, is the deliverable. The five forms are posted on this issue. Panel automatic publication remains off until a held-out comparison with the generic editor using the same quality form and measured reviewer/machine cost (Delivery: curated opening section with five local persona editors and visit-driven refresh #193).
  • At least one highlight on a pilot page is a pinned PoetryDB poem that was not materialized before (checked through the existing poetrydb_poem identity adapter, §9.6), offered by the screen and rendered as a work, and the withhold-and-restore round trip was demonstrated under the supersession rule the owner decides (§9.6).
  • Each pilot page's disclosure shows published, last assessed, checked against the record, inputs last changed and next assessment after as distinct dates; withdrawing a claim makes the page dirty, the next visit admits exactly one run, a second visit admits none, and the owner can watch it in the queue.
  • The §6a rules are demonstrated: an owner-curated page receives a bot proposal without its published version moving; a withdrawn page receives none; a publication during a run yields a superseded_on_arrival proposal.
  • Every number above comes with the row ids and attempt ids; the record restores from a bundle like routing state.

Prerequisites, each with its owner (they block acceptance, not implementation):

Not in the MVP: personas and the panel (5b), public nominations, personal or domain configurations (#201 stages 2–3), the history and method pages (#203, #204), any model other than the pinned 9B, any page outside the allowlist.

11. Parts, each independently reviewable

"One PR per phase" was too coarse; the earlier phase 2 bundled a screen, durable ingestion and a packet change. The work is seven parts, each one PR with precommit, one independent review and a before/after on a real page. The briefs an implementer needs are in #254, one section per part.

Part Delivers Depends on Reopens
P1 Seed, activate, the page reads published compositions, the disclosure's publication rows (the assessment dates are P5's) — #156 (Phase 2)
P2 The owner's screen (§6): queue, per-page editor, publish, withdraw, re-open P1 —
P3 The pin, PoetryDB end to end (§9.6) P1 #164's rule (closed)
P4 work (catalog artwork) and exemplar candidates in the packet; the meaning bound and the trimming policy; runtime-packet-v2 P1 (existing catalog artwork fixtures; P3 only for pinned poems) —
P5 Freshness: the refresh-state table, admission on visit, the worker without a model call, §6a's rules, the disclosure's assessment dates P1 #198
P6 The proposal path: attempt → :model version, attribution, "Take" and edit, the quality form, D8's marks, D9's audit P2, P4, P5, and P7's revised-instruction benchmark go #197 (first half)
P7 Runtime: #221 A1, S1, B1 (A1 including the retry ceiling), the instruction revision, the benchmark re-run on the pinned 9B — #221
∥ QID minting, lexicon-wide (coverage) — #217

P1 goes first and alone; it is small and unblocks everything. P2, P3, P4 and P7 touch disjoint code (web screen; discovery and registry; runtime packet; runtime gateway and bench) and can run in parallel after P1. P5 follows P1 and should land before P6. P6 is last: P6 live inference requires P7's revised-instruction benchmark go, plus P2, P4 and P5, and uses the actual pinned instruction label, not a hard-coded generic-editor v1. The five per-item quality questions and the five forms stay on this issue (§10), owned by P6. Personas and the panel (§5b) come after the MVP and get their own parts then.

12. Decisions, with the pilot's defaults

Each has a default so an implementer never has to infer one; the owner can change any of them on this issue.

  • D1. Highlights: three. Decided.
  • D2. The ten pilot pages. Default: nepotism, greed, war, love, coward, patience, justice, truth, hope, bestiality (the Current words: a trending signal (Wikipedia top-1,000, keyless), a News shelf (Guardian then GDELT) and Urban Dictionary by permission — tested on /define/bestiality, 2026-09-15 #134 test case). The owner may swap any.
  • D3. Fallback lead where Bierce has no entry. Default: Johnson where the page has his entry; otherwise no lead. WordNet/Wiktionary senses are not entry leads or content highlights in the current record, so greed and bestiality have no lead. The owner's open choice: keep this pilot limitation (the default) or authorize a narrow sense-reference extension, with its storage, policy, rendering, validation and recovery work; P4 and P6 alone cannot make a sense into a content item. A duplicate definition is never manufactured to evade the record. The bot may abstain from a lead; the owner's hand overrides on the screen. The lead policy's :manual_fallback statement is reworded for a bot-chosen fallback ("selected by the editor") when P6 lands.
  • D4. Self-publication. Default: off for the whole MVP; the owner keeps the last word. Panel automatic publication remains off until a held-out comparison with the generic editor using the same quality form and measured reviewer/machine cost.
  • D5. Pinned works as public pages. Default: highlight-only. A pinned work has no family address and no public page of its own; its card links to the provider. Routing a pinned work is a later decision under 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's deployment and the routing families.
  • D6. Pin provider for the pilot. PoetryDB. Decided, with the identity contract in §9.6. Open with it, the poem-to-page rule: catalog references can be stored/rendered, but P2 lists only page-matched artwork rows; poems/books need an explicit page-selection relation. The owner decides that relation in the P3/P4 handoff, over the existing attestation and page-selection records, not a guessed "uses the word, so it is about the meaning" claim. Default until then: neither the screen nor the packet offers a catalog poem or book.
  • D7. Refresh windows. 48 hours default, 48 hours for abstentions. Pilot allowlist only during MVP; later expansion disabled.
  • D8. Emphasis. An item the owner marks kept survives into every bot proposal unless unmarked. A page marked mine receives proposals, shown in the queue but never auto-taken. Marks live on the composition with who and when. Storage is P6's.
  • D9. Choice audit. Every editor-chosen item links to its attempt. The attempt records the candidates offered and those omitted by the packet cut. The page-level disclosure remains the one place naming configuration, selector, publisher, panel and the five freshness dates.

Activity

  1. holden commented on Oct 21, 2026

    @holden
    MemberAuthor

    Independent review applied, 21 October 2026

    A review graded this B+ (A− direction, B− implementation readiness) and asked for six sharper contracts. I checked each against main at dc888b7 before changing anything; all six were right.

    Request Checked against Verdict Where it landed
    The refresh clock used publication age; separate assessment, publication and retry #198's clocks (last_completed_at, refresh_after, next_retry_at, refresh_generation) Correct; the draft would have re-run inference after every abstention §7 rewritten; #198's paragraphs carried in
    Owner edits, withdrawal and bot refresh were undefined Publications.publish/3 compares-and-switches; the gateway has fences; nothing said what withdrawal means Correct New §6a with a rule per situation
    The pin needed a contract Discovery.Result has identity and display_allowed; #164's identity rule; nothing on retention or disappearance Correct §9.6 rewritten; PoetryDB first (D6)
    Proposal attribution Compositions.note/2 attributes only to the authenticated human and refuses caller-supplied authorship; origin and selection_origin admit only :manual Correct; "Take" would have relabelled model text §5a steps 4–5; origin: :model, attempt_id, parent links
    A quality gate, not a count — Agreed §10: a fixed form per proposal, posted here; self-publication off (D4)
    Trimming was a silent policy change Packet.build/3 refuses more than 12 ({:too_many_candidates, n}) Correct §5a step 2: quotas, Bierce always in, tie-break, omitted list, no dirty from outside the cut

    Smaller points, all taken: "no new tables" replaced by the storage each part needs; "one PR per phase" replaced by seven independently reviewable parts (§11); "checked today" split into five distinct dates; #198 and #221 named as closed-for-consolidation and reopened by their parts; three highlights kept; every pilot decision given a default (§12).

    The implementation plan with a brief per part is #254.

  2. holden commented on Oct 29, 2026

    @holden
    MemberAuthor

    Recorded owner amendment — #271 sitting 3

    Posting the amendment already specified by #271 §3, before its citation deployment. This records the owner’s existing decision; it does not activate the editor, a model, panels or scheduling.

    Amendment (owner, 27 October 2026).
    (1) The goal is the two layers of #271 §1: layer 1 only chooses from layer 0, and every choice is citable and transparent to readers, for every content type.
    (2) §4c and §9.6: discovery results on a page's shelves are layer-0 candidates for layer 1, cited by their exact stored observation under #271 §3. No owner pin is needed, and the provider's match basis is disclosed as what the choice was based on.
    (3) Sense definitions, and public entities and claims linked to the page, are citable without creating new claims or definitions. D3's lead rule is unchanged.
    (4) D6's poem-to-page relation, for shelf items, is the shelf match itself.
    (5) §7's refresh window per configuration version is implemented (C9).
    (6) §10's acceptance is replaced by #271 §7. Personas and the panel (§5b), unattended scheduling, enrolment beyond the pilot, publication without the owner (D4 stays off), a full restore drill and open-page suppression are deferred to later issues.

    #275 separately owns the later agent/panel work; this foundation amendment does not claim it implemented. Live #271 acceptance remains outstanding.

  3. holden commented on Oct 29, 2026

    @holden
    MemberAuthor

    Owner amendment — 29 October 2026

    The owner explicitly rejects mandatory human review of routine curation. Full correction; current implementation contract: #275 Phase A; updated foundation acceptance: #271 §§4–8.

    The historical manual-only/no-bot-publication and compulsory Take/Edit/quality-form requirements are superseded for an explicitly enabled single_editor_v1 ten-page pilot. A validated accepted attempt may publish through the common composition/citation/receipt path under honestly named single-editor authority, without a human acceptance record. Existing manual policy remains the default and optional owner intervention remains supported. Development inspection measures the experiment; it is not a permanent publication queue. Consensus is later work, not falsely claimed by this first curator.

    Retain all non-conflicting exact-source, citation, retention, runtime budget, concurrency/fencing and provenance requirements. Owner holds/withdrawals must prevent automated overwrite. Invalid/stale/expired/failed results cannot publish. Closure must distinguish superseded manual exercises from proven live autonomous behavior. No deployment, activation or model operation is authorized by this amendment; implementation stops at a tested independently reviewed PR.

  4. holden commented on Oct 29, 2026

    @holden
    MemberAuthor

    Owner-directed whole-entry amendment — 29 October 2026

    The owner has required actual useful, visible curated entries on the existing ten pilots, love first, rather than treating a working pipeline or three highlights as completion. #275 and PR #279 carry this implementation; no broader enrollment or panel is added.

    For explicit whole_entry_v1 only, this supersedes the old three-total-highlight/twelve-candidate presentation limit: one sourced lead plus up to twelve section-limited highlights, chosen in one call from at most 24 candidates/12 meanings. Definitions, examples, quotations and seven discovery shelves use the same exact-record selection/citation/provenance and reader-visible reasons. Sections may remain empty; packet omissions/rotation are recorded, with no claim all records were inspected. Historical packet/instruction/configuration versions retain their old behavior. C1–C9's identity, retention, source-content boundary and transparency rules remain intact; this is a bounded versioned selection/display amendment.

    Routine publication uses the previously recorded single-editor amendment, not human approval. The actual model must publish useful choices on love before the other nine run. Tests and independent review gate pinned deployment. Real expiry/citation survival and next-day unattended assessment remain unproved until observed; no closure is claimed by the code change or fixture tests. The approved bounded trial ends with model/editor stopped. Server restart/reconfiguration, model downloads, consensus and extra pages remain outside it.

  5. holden commented on Nov 1, 2026

    @holden
    MemberAuthor

    Closed 1 November 2026 into #298. §10: nine of ten pilot pages show a curated opening live (versions 20–28); bestiality is still missing (#298 C2). Owner hand-curation, the five forms and the PoetryDB pin were superseded by the owner amendments of 27 and 29 October; the §6a rules are tested and partly superseded by single-editor policy; the bundle restore is pending (#298 A4 and C1). The specification sections remain the reference; carried obligations are in #298 §8.

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

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions