Repository navigation
Description
Owner-approved sequencing update — 3 November 2026 (Central). Sourced, data-only persona dossiers can be drafted and reviewed now under #197. Start from the Hitchens dossier in #302 and apply the same source, rights and editorial-lens criteria to the remaining profiles. Dossier work can proceed independently of #298's trial closeout, offline comparisons and rollout readiness. Live admission and publishing still require the applicable quality and readiness evidence and explicit approval.
Handoff from #298. v7 failed its frozen primary quality gate. Track the next packet-policy experiment here using #303 and the D6 four-arm design. Track the matched generic/Hitchens comparison here as the first persona behavior evaluation. Retain D1 as an optional diagnostic with its disposition recorded. For unattended rollout, retain any unproved outage/deferral evidence and all unresolved #298 §9 obligations. Reconcile the register against the final trial closeout before closing #298. #299 and #300 remain the page and panel delivery tracks.
Tracker pointer, 1 November 2026. Implementation and experiment work for the curation agents now runs in one issue, #298: close out the frozen v6 trial (stop 4 November 02:50:07 UTC), deploy and retest v7 on the same ten pages against frozen scorecard v2, then controlled model and personality comparisons on identical packet hashes, with admission made data-driven. #253, #254, #264, #267, #268, #271, #275 and #198 are closed into it; every obligation they still carried is labelled in #298 §8. This issue stays the programme umbrella. The 1 November status below is accurate as of that date and is superseded by #298 where they differ.
Outcome and scope
Keep existing definitions, examples and discovery shelves intact. Add a colorful, sourced curated opening section above them: a lead definition, Bierce first whenever an applicable Devil's Dictionary entry exists, plus up to three highlights (image/artwork, quotation, poem/passage or another eligible item). Each highlight carries source attribution, intended meaning and a concise attributed editorial reason; readers can inspect permitted dissent. Never manufacture material to fill slots.
This issue is the delivery tracker, extending #97's research and #17's original brainstorm. Development runs on the Mac Studio, with a later measured Mac Mini handoff under #252 (which absorbs #199). No application code, migrations, model downloads or deployment are performed by this planning update.
A saved page composition is presentation selection/order across meanings. An illustrates nomination is a separate attributed claim about a specific meaning, using existing evidence/review. Selecting a sourced picture or poem does not automatically create a claim. Under the owner’s 29 October amendment, #275’s explicitly enabled single-editor policy publishes automatically; no mandatory human review or publication click is required. The owner can intervene or hold a page. Future panel consensus needs its own reviewed decision authority and evaluation; single-editor publication must never be labelled consensus. Linked exemplars retain their independent claim-review gate.
Current delivery status — 1 November 2026
Active route: finish the frozen #275 ten-page experiment and #271 evidence, then improve one curator before adding perspectives or pages. This dated status and the owner amendments on #275 supersede conflicting historical manual-review, three-highlight and sequencing assumptions below.
- Merged/deployed foundation: runtime, proposal writer, exact native/shelf citations and retention-aware reader receipts; autonomous single-editor publication and grouped whole-entry selections. Prepare the ten-page curator quality retest #296’s preflight is merged/deployed at
9a17333d9aa364e1116ecc12ea2b0c1ce0c08738; frozen scorecard v2, read-only coverage report and v6 sense-lead instruction are delivered. No open PR was awaiting merge at this check. Merged implementation is not quality acceptance. - Current experiment: one installed 9B artifact and one sourced Voltaire lens, same ten pages, frozen v6/scorecard v2. The 72-hour window ends 4 November 2026 at 02:50:07 UTC (3 November 8:50:07 p.m. Chicago). Keep its inputs/version policy, failures and stop deadline unchanged while preparing corrections separately. No expansion or extra personas.
- First-cohort quality NO-GO: all ten assessed unattended, nine published and bestiality incorrectly abstained despite supplied native definitions. Across all 73 proposed choices: 43 supported meaning, 27 insufficient evidence, 3 wrong; reasons: 52 grounded, 20 unsupported, 1 policy-withheld/unassessable. These are developer judgments, not independent editorial approval. Do not quote an assessable-only percentage as an overall pass.
- Merged correction, not deployed: PR Prepare evidence-grounded v7 curator candidate and provider-wide regressions #297 at
093ef1534a05fc345dd007c4c2ca105e251fee09contains Agent MVP: one autonomous curator on ten pages, then measured multi-agent consensus — under #193 #275’s immutable v7 evidence/meaning/lead-first candidate and synthetic end-to-end coverage for six native adapters and all twelve durable server shelf providers. 94 focused checks and full precommit (19 doctests / 3,330 tests) pass; independent exact-head CODE GO, required none. It is merged as8d738f11946b5bd2cbf65de5499518393f66ab2cand is not deployed. The current v6 experiment remains on its original deployed pin. GIPHY remains browser-only/transient. No live migration, new model calls or trial mutation. A later separately authorized same-ten-page real-model retest must meet the frozen whole-denominator quality gates before expansion. - Foundation acceptance: 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 retains actual due-day, unchanged/fresh-repeat, hold/offline, news expiry/cleanup/citation survival and final stopped-state evidence. Reconcile proven, superseded and deferred obligations before closing 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/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/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/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/Curation: the next stretch to the agents — #267 and #264 in one audited order, deployment gates, and "ready to prepare the persona plan" (under #254, #253, #193) #268/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. Visit-driven curation refresh: coalesced jobs, fenced leases and atomic publication #198 keeps every unproved recovery/open-page/freshness obligation.
- Next perspectives: Agent MVP: one autonomous curator on ten pages, then measured multi-agent consensus — under #193 #275 Phase B only after one useful curator is measured against the generic control on identical lawful evidence. Reuse one physical inference slot and the same exact-choice records; add minimal participant/ballot/decision authority with preserved dissent. Several roles on the same weights are correlated judgments, not independent corroboration. Consensus cannot repair missing evidence or override expiry.
- Separate backlog: 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 hosting, 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 linking, Word page height and shelf templates: fold the prose, one template per content type, the world above the fold — measured on /define/love, phased #111 presentation, 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 lists and Naming: 100 broader ideas and eight extensions, plus the earlier top 50 #192 naming remain independent; none is silently a prerequisite for this focused quality correction. Delivery: curated opening section with five local persona editors and visit-driven refresh #193 remains the programme umbrella, not another implementation queue.
Folded specifications retained
#196 persistence, #197 personas/evaluation, #201 configuration, #203 transparency and #204 editorial method remain programme specifications, with #275's staged owner-reviewed implementation now the next agent work. #190 still gates any future semantic-claim writes; this MVP only chooses existing records. #189 voting UI, #101 broad assessor, #105 additional exemplars, #222 quote linking and #213 quote-corpus growth are not silently made prerequisites. Wider enrollment, scheduling, ownerless publication and hosting remain separate measured decisions.
Delivery sequence and owners
| Order | Issue | Deliverable / completion evidence |
|---|---|---|
| Early, parallel | #204 | About / Editorial method page centered on Bierce’s satire, expressive insight, checkable evidence and honest current/planned status. Can ship before AI or public history; omit unavailable feature links. |
| 1 | #156 | Editorial output contract and responsive opening section using manually selected, eligible sourced data; Bierce-first lead, up to three highlights, credits/reasons/dissent, existing page below preserved. Merged #159 sketches are the starting point. |
| 2, parallel with 1 | #195 | Persistent Ollama/Req runtime spike on the Studio: pinned pretrained model, explicit storage, readiness/warm loading, one global inference slot, bounded requests and measured latency/memory. |
| 3 | #196 and #190 | Versioned roster/run/ballot/composition persistence, finalized panel-decision publication authority and optional operator review/publication. #190 lands before any bot claim writes, not merely before public launch. |
| 4 | #197 | Ten-page experiment: generic editor versus five-profile panel on identical evidence, held-out evaluation, useful selections and dissent per machine/reviewer time. |
| 5 | #198 | Visit-triggered missing/stale/changed-input checks, coalesced/fenced jobs, retry cooldown, validated replacements and atomic publication on valid panel-decision or optional operator authority. |
| 6 | #252 (absorbs #199) | Measured Mini handoff through the same protocol; no orchestration redesign or speculative hardware purchase. |
#181 is closed on its implemented prototype (examples backbone and reverse views); reuse it. #189's human voting UI/cached tallies are not required for internal panel ballots. #101's broad assessor remains separate and unbuilt: manually reviewed evidence packets unblock this pilot without waiving its provider-spectrum/evaluation gates. #105 owns exemplars. #75 was absorbed by #97; #96 was documentation absorbed by #109, not completed saved composition. #194 owns routing separately; this work uses durable composition/registry identities, never URL spelling as identity.
Explain the method and retain an inspectable history
Two focused delivery children add accountable transparency without expanding #201's future features:
- Curation transparency: equal votes, reproducible selection, profiles and version history #203 — Curation transparency: equal votes, reproducible selection, profiles and version history. Each pilot profile has voting weight 1; four of the fixed five support a recommendation. Missing/abstained responses do not lower the denominator, and mandatory evidence/rights/claim blockers cannot be outvoted.
selection-v1makes highlight ordering reproducible from frozen annotations/ballots; human overrides remain distinct, attributed new versions. History and profile pages project Curation persistence: versioned panels, ballots and reviewed page compositions #196 records with privacy/retention filtering, not raw reasoning or prompt dumps. - Editorial method: declared perspectives, checkable facts and accountable curation #204 — Editorial method: declared perspectives, checkable facts and accountable curation. Contains a cohesive About draft centered on Ambrose Bierce’s satire and the expressive value of humor/art, drawing carefully on Zinn and Gray. Distinguishes satire, exact historical quotation, factual statements and generated commentary; keeps checkable evidence and deployed/planned status explicit. Zinn/Gray are references only, not roster additions. Method/history links integrate with Curated opening section: Bierce-first lead, sourced highlights and inspectable editorial choices #156 under Routing before launch: On overviews, standards-informed namespaces, and stable subject URLs #194's route ownership.
Implement policy/provenance alongside #196/#197, then the public history/profile reader when approved persistence is available; publish the About/method page early with truthful release status and working source links. It need not wait for runtime, persistence or history: omit unavailable links, mark future capabilities planned and describe the correction path actually available without requiring a new tool. #156 Phase 1 (PR #202) is merged as a development-only manual fixture with no panel, and P1 (#256) added the published reader; show “none recorded” rather than inventing votes. These transparency deliverables belong to the public curation experience, not domain selection, personalization, extra generation or unequal weighting.
Configuration boundary: one default now, expert and personal views later
#201 Curation configurations: global default, domain panels and future personal views is a focused architecture child of this tracker. Stage 1 seeds one named system-owned global-default configuration with immutable versions and versioned membership of the existing five profiles. The same default applies to every enabled subject; the pilot is still ten allowlisted pages, not an all-corpus sweep. All offered persona subjects are deceased, checked through sourced profile admission. New roster entries are configuration data, not name-specific code.
Minimum additions belong in #196: stable configuration identity, immutable versions/roster references and configuration-scoped composition uniqueness. A composition belongs to (stable target/scope, configuration ID, language); its configuration version is frozen run/content input, not a new composition identity per tweak. Actual participants, model weights, profile identities and reusable panel/configuration identity remain distinct. Shared source material does not imply shared composition approvals, published pointers or freshness.
#197 consumes the resolved immutable version; #198 scopes cache/admission/invalidation accordingly, while one global inference slot and budget remain shared across configurations. MVP resolves only the system default server-side. Two logical configurations in integrity tests demonstrate independent histories without enabling alternate public panels or accepting arbitrary configuration IDs from URL parameters. #156's in-flight default-only UI can continue, with config/version/resolution provenance added to its data seam when connected.
Later Stage 2 adds subject-domain expert defaults (Monet/Picasso are examples only), with versioned stable-ID evidence/assignments, explicit overlap precedence and global fallback—not URL spelling or public namespace guessing. Later Stage 3 adds user-configured saved views only after ownership/access/cache-isolation/sharing policy is designed and authorized. No domain UI, accounts, sharing, personal ownership controls, fine-tuning or multiple public panels are MVP requirements. Current Bierce-first is the default configuration's versioned editorial rule; mandatory source/rights/claim-integrity rules are non-overridable.
#194 retains page/route/On-body ownership: at most one active default editorial binding per page, without preventing future alternate configuration-scoped compositions. Its merged ADR 0004 needs a narrow wording amendment, recorded in #201/#194; no repository/ADR file is changed by this issue update.
Pilot roster and bounded disagreement
Five versioned, explicitly inspired-by AI profiles: Ambrose Bierce, Voltaire, Kurt Vonnegut, Christopher Hitchens and Ursula K. Le Guin. Le Guin is the recommended fifth dossier to ground in her writing, not a verified stance on modern subjects. Dossiers include interests, original-source passages with edition/rights, principles, positive/negative examples, blind spots and abstention. Personality changes what is chosen, not merely phrasing. No fine-tuning required.
One model serves all profiles sequentially. Freeze actual participants, profile versions and policy per run; a future roster of ten routes a bounded panel through configuration, not name-specific code. Each packet holds up to 12 candidates with exact eligible evidence and intended meanings. The pilot is ten allowlisted pages, with manually bounded meaning scope; compositions may span meanings and do not assume the first sense represents the page.
Two rounds per profile: independent proposals, then challenge/final ballots. Each voting weight is 1; four of five support recommends a candidate if no evidence/policy blocker remains. Optional operator publication is a separate accountable action, not another vote or a routine per-selection prerequisite. Three–two stays disagreement; missing responses/abstentions do not shrink the denominator. Five people do not prevent deadlock. Finish bounded rounds, preserve prior eligible content or an empty slot, and retain concise dissent. Agreement is not independent corroboration and cannot bypass evidence/policy checks or approve a semantic claim. The sourced Bierce-first rule is not up for a model vote.
Ten logical generation calls, at most one additional schema-repair generation per panel (11 physical generation requests maximum). No open-ended debate or autonomous search. Proposed panel highlights are capped at three, with at most three linked illustrates nominations per packet (one unfinished composition plus three pending claims = at most four pending-work units). Private ballots never become human votes or automatic endorsements. Sensitive personal interpretations require specific sources and separate human claim review; do not infer psychiatric diagnoses from public behavior.
Persistence and freshness
#196 specifies named curation_configurations/immutable versions/membership, curator_profiles/immutable versions, pinned local_model_configs, curation_runs with frozen participants/candidates and curation_ballots, nominations through existing claims, and stable editorial_compositions with immutable versions/items and separate presentation decisions. A composition can exist manually before inference. Exact durable source/content/sense revisions and passage locators anchor selections; expiring discovery result IDs alone do not.
Visits return the current eligible approved version immediately. Missing/stale/changed evidence admits at most one background run per composition. Default 48h, configurable 24h/7d; no global sweep. Input fingerprints include meaningful candidate/evidence/sense changes and profile/model/policy versions; coalesce provider updates. Successful assessment freshness, failed-attempt backoff and published-version date are distinct. Empty completed assessments get a cooldown. Failed or offline inference never marks curation successfully fresh.
Proposals remain drafts until a finalized consensus decision or optional operator acceptance authorizes atomic publication of the exact, still-eligible version. An operator rejection remains a veto. Existing readers stay on their version during ordinary refresh; withdrawal/rights expiry suppresses invalid content promptly. No silent substitution or mid-read reshuffling. Changed evidence/senses/policies require renewed validation; rejected nominations do not resurrect on retries. Attribution, retention and rights apply to prompts, logs, ballots and historical versions too.
Operating limits and hardware
Phoenix/Oban/PostgreSQL orchestrate; Req calls private persistent Ollama with structured JSON and server-side validation. Start by benchmarking pinned Qwen3.5 4B/9B as candidates, not asserting their quality. Models and runtime files are already on the mounted external /Volumes/LLM Models/dictionary root, with no silent internal fallback. The dictionary's source, PostgreSQL, worktrees and bulk artifacts moved there too (#209, #211). Native Bumblebee/Nx is unnecessary for HTTP inference; reserve it for separately justified #101 work.
Initial limits: one physical inference globally; 8K context with output room; 1,024 generated tokens/call including thinking; 120-second call and 20-minute panel deadlines; 30 minutes/day inference budget, 30 pending-review units (draft composition plus each separate draft claim counts). #195 defines repairs, retries, reservations and pause behavior. Offline runtime defers work; no reader delay and no silent cloud fallback.
Planning arithmetic only: ten calls at 20s each = 200s/panel, about 18/hour; at 60s = 600s/panel, 6/hour, before overhead/repairs/review. A 30-minute budget fits at most nine or three such panels respectively before overhead. Measure real Studio capacity before changing those limits. Mini feasibility depends on actual chip/RAM, context and workload; #252 measures it. Hugging Face model distribution and hosted inference are separate options, neither a pilot requirement. Current runtime sources and verified project precedents are linked in #195.
End-to-end acceptance
-
loveopens with applicable sourced Bierce text and up to three eligible sourced highlights; original definition rows and discovery remain available below. - Sources, exact passages, generated-reason attribution, profile/model versions, publication authority, optional operator actions and lawful dissent can be inspected without exposing private pending adverse nominations. Curation transparency: equal votes, reproducible selection, profiles and version history #203 supplies version history/diffs, reproducible counts/order and reviewer overrides; Editorial method: declared perspectives, checkable facts and accountable curation #204 explains the method with checked sources and accurate implemented/planned status.
- Fresh repeat causes zero inference; intentionally stale concurrent visits cause one refresh and one reviewable replacement. Offline runtime does not break reads.
- Fixed 4/5 quorum, 3–2 disagreement, missing-response denominator, empty slots and bounded termination are demonstrated.
- One system-default configuration is seeded; two test-only configurations for the same scope retain separate histories/freshness/publication, while one global budget and default-only public resolution are enforced (Curation configurations: global default, domain panels and future personal views #201).
- Schema/publication tests prevent duplicate writes, stale-lease commits, invalid evidence, pending-claim leakage and unauthorized publication; two reviewers cannot overwrite each other silently.
- Held-out generic-versus-panel report demonstrates whether personalities improve useful selections at an acceptable measured machine/reviewer cost before expansion.
- Mini handoff is decided from measured results, using the same protocol.
Decisions still genuinely open
Final visual treatment among #159's sketches, the no-Bierce lead fallback order, and whether transient/query-only items may ever enter a saved composition need product decisions. MVP defaults in #156 are conservative and implementable: manually selected eligible fallback or no lead, and durable verified material for saved highlights. Actual model choice, quiet operating window and future expansion budget await the spike. Automatic publication, hosted compute and hardware purchases remain later decisions, not blockers to this plan.
Updated 26 September 2026 after the planning discussion. This supersedes this issue's initial four-person/eight-call/one-sense/Mini-first proposal; useful sourcing, human review, bounded compute and visit-driven constraints are retained. #193 remains the sole delivery umbrella.
Configuration extension recorded 26 September 2026: #201 stages future domain/personal views; only minimal configuration identity/version/default-resolution foundations enter this MVP.
Cohesion audit — 26 September 2026
Reviewed current bodies of #193, #156, #195–199, #201, #203, #204 and the #194 integration contract. This is specification coverage, not implementation acceptance:
| Contract | Matching issue coverage |
|---|---|
| Five deceased-inspired profiles; weight 1; fixed 4/5; ten normal generations + max one repair | #195, #197, #201, #203; umbrella limits above |
| One global-default MVP; domain panels later; personal views later | #196–198, #201, #203; #156 default-only UI seam |
| Configuration-scoped composition identity, history, freshness and approvals; global inference budget | #195–199, #201, #203 |
| Separate page publication, composition presentation approval and semantic claim review | #194, #196, #190 dependency in #193/#197; #203 publication provenance |
| Current evidence/rights validity, immutable selection history with lawful redaction | #156, #196–198, #203 |
| Early About copy; source quotation versus satire/fact/generated commentary; future-feature status | #204 and #203; #156 attribution/disclosure contract |
Clarified #197 so a timed-out/missing response cannot be recorded as a fabricated abstention. The denominator remains five.
Outstanding repository mismatch: remote main at 5713f8e0960599a8a55f78518d8857a98b7c64e6 still says “at most one active binding per page” in ADR 0004, interface with curation persistence. Apply the already recorded #194/#201 amendment: at most one active default editorial binding per page, allowing future alternate configuration-scoped compositions. No ADR or application file was changed by this audit. Runtime benchmarks, persistence, panel evaluation and public history remain implementation work; agreement between issue bodies does not establish that they work.
Activity
[-]MVP: visit-driven local curation with four competing persona editors and freshness budgets[/-][+]Delivery: curated opening section with five local persona editors and visit-driven refresh[/+]holden commentedon Sep 27, 2026
Migration-readiness checkpoint — revised owner priority, 27 September 2026
The owner clarified the immediate goal: wrap Stage 2 enough to preserve and consolidate all current work, then move dictionary storage to the external 2 TB drive. Completion of every routing Stage 2 deliverable is not a prerequisite for this checkpoint. This supersedes the earlier requirement to finish all Stage 2 before relocation; it does not waive data-preservation or migration verification.
Current position: #206 is merged; #202, #210 and #208 remain open. CodeRabbit checks have completed. #202 has no unresolved review threads; all nine on #210 are resolved; #208 has two unresolved findings plus independent audit findings. Bot success alone is not acceptance.
Bounded work before migration
mix precommitfor the integrated head when capacity permits. If a required check cannot fit, record the exact missing evidence and resource requirement; continue safe work. Do not create internal database copies simply to satisfy the earlier finish-all sequence.Exit and remaining scope
This checkpoint is complete when all current code is accounted for in main plus at most one designated integration branch, local changes and evidence are preserved, integration/safety checks are reported honestly, and the migration has a concrete preservation and rollback plan. The migration remains a separate execution task under #209; do not start it under this closure handoff. The owner's explicit relocation goal covers the dictionary source and PostgreSQL, superseding the older internal-storage preference for this project; unrelated projects/shared PostgreSQL databases must remain protected.
Update #194/#195/#209 with concise relevant status and links. Keep #194 open: full Routing Stage 2 acceptance follows its existing criteria, with unfinished work resumed after relocation. Keep personalities, refresh, reader enablement and publication outside this assignment. Do not self-award A+ or relabel unfinished work complete.
holden commentedon Sep 27, 2026
Consolidation checkpoint for the plan above:
mainc6bd882pluscodex/consolidation-2026-09-27at697ac8a, which holds #202 (accepted, not merged), #210 and #208 (both not accepted). Precommit on that head: 2,453 tests, 0 failures. The migration is not yet safe to start. One session is still active in the #208 worktree, and the owner must decide on the 13:11Z dev-database incident. Handoff: #209. Status: #194, #195. Nothing was merged, relocated or deleted.holden commentedon Oct 21, 2026
Consolidation, 21 October 2026: this issue is the one open tracker for the curation programme
The owner asked for the open issues cut to a handful. The children below are closed and folded here; each keeps its full specification, so reopen one when its work starts. Status as of today:
?opening=fixturedevelopment gate; Phases 2–3 (read saved compositions, show publicly) pending. Routing's page–composition binding (ADR 0004 §4, Stage 4) waits for Phase 2's reader.global-defaultmerged in PR #206. Stages 2–3 (domain panels, personal views) are backlog.MatchErrorinstead of refusing; orphaned Ollama runners on stop; two vacuous benchmark cells. Do before #197.illustratesnomination until acceptedAlso closed as historical: #17 (the 2025 curator-bots brainstorm) and #97 (the research parent), both already pointed here.
holden commentedon Oct 21, 2026
The full MVP specification for the curation agent, the page, the owner's screen, the freshness clock and the persona panel is #253 (21 October 2026). It wins where it differs from the folded children; they keep their detail.
holden commentedon Oct 26, 2026
Status pointer, 26 October 2026: P1, P2, P4, P5 and P7 are merged (P7's benchmark is a GO, #266). The path from here to the persona editors is #268: the runtime made live (#267), the first bot proposing (#264), five owner quality forms, then a persona-phase plan issue. #268 is awaiting its independent audit.
holden commentedon Oct 27, 2026
Status pointer, 27 October 2026: the curation infrastructure is now closed off by one dedicated implementation issue, #269 (linked here as a sub-issue).
It finishes #267 (runtime live), #264 (the first editor bot), a new P3 issue (the PoetryDB pin) and #198's remaining contract. It proves #253 §10 on the live site, tidies the repository and the cluster (99 stale remote branches, 9 local ones, 26 leftover test databases, and a fresh bundle, since none of today's bundles holds a curation record), and then closes #267, #264, P3, #198, #268, #254 and #253 with evidence. Its owner decisions D1–D10 are listed in its §3.
This issue stays open as the persona tracker. Persona implementation starts after #269 closes, which includes opening the persona plan issue and requesting its independent audit (#268 §4).
holden commentedon Oct 29, 2026
Owner clarification, 28 October: #275 now includes an authoritative bounded-autonomy amendment. Owner approval is a development gate, not permanent per-entry approval. Its intended end state includes recorded consensus-authorized publication, due-only unattended reassessment of the ten pilot pages, distinct assessment/publication/source/expiry dates, and durable optional owner holds/overrides. This supersedes conflicting owner-only/shadow-only wording in my earlier status update. These capabilities are not yet built; #271 live foundation acceptance and #275 benchmark/review/autonomous acceptance remain required. No deployment or activation was performed or authorized by this planning update.
holden commentedon Oct 29, 2026
Owner correction, 29 October 2026: routine agent curation must not wait for human review/publication. Human inspection is development evaluation only. The implementation issue #275 and #271's remaining operational acceptance have been rewritten accordingly; the older #253/#264 manual-only requirements have dated amendments, not false completion claims.
PR #278 implements the bounded correction: one explicitly enabled single-editor publication authority, exact common citations/receipts, existing owner hold/withdrawal controls, due checks through Refresh, and password-free private owner curation. No panels, expanded enrollment or competing history/source system. Manual policy remains the default until separately authorized configuration/activation. Model is stopped and live code/schema unchanged by this work. Independent exact-head review is GO conditional on final full precommit; real-model quality and live next-due/expiry evidence remain unproved and assigned to the amended pilot. This is progress toward autonomy, not a claim the programme is complete.
holden commentedon Nov 1, 2026
Status pointer, 1 November 2026: the curation programme now has one open implementation and experiment issue, #298, under this umbrella. It carries the v6 close-out (A1–A4), one reviewed PR (B1–B6: data-driven admission, a non-publishing comparison arm on recorded packet hashes, the comparison report, a parameterised stop guard, committed evaluation records) and the experiments in order (C1 deploy once, C2 v7 retest, C3 installed-4B model arm, C4 generic control, C5 second persona on OD-A). Closed into it with closure notes: #253, #254, #264, #267, #268, #271, #275, #198. Backlog #111, #170, #192, #217 and #252 stay open and separate. Nothing on the Studio, in v2 or in the running v6 trial changed.
holden commentedon Nov 1, 2026
Pointer, 1 November 2026: #299 is the design and MVP issue for curator pages (one public page per AI personality with its dossier sources, model, instruction, audit trail and cost, plus a directory), implementing the profile-page part of #203. It is a read-only projection of existing records, not a prerequisite for #298, and must be deployed only between #298's frozen windows. Open under this umbrella alongside #298.
holden commentedon Nov 1, 2026
Pointer, 1 November 2026: #300 is the design and MVP issue for living, visit-driven multi-persona curation (overlap consensus k of N on one packet, panel_v1 publication authority, support and dissent on receipts, one staleness window now and per-kind windows later, the living-person rule for current events). It keeps the never-built panel rules from #193, #196, #197 and #203; it goes live only after #298's quality gate and uses #299's pages. Open under this umbrella with #298 and #299.
holden commentedon Nov 4, 2026
Pointer, 4 November 2026: #305 is the implementation and design issue for what follows #298's closure, posted on the owner's words in chat ("Post it"). Part 1 carries the experiment the #298 amendment left untracked: the D6 four policy arms offline, the one v8 publishing retest on the same ten pages under the unchanged gate, the matched generic-control and Hitchens (and Bierce) comparisons, the optional 4B model arm, the D4 decision and one read-only audit. Parts 2 to 5 record, in the owner's words, how a dossier is built and earns a seat (with the next two dossiers, Vonnegut and Le Guin), how agents reach consensus with per-seat weights (#300 L1 builds the general weighted-overlap form with equal weights as the default), how a human seat weighs in by standing preferences, and how a second model would be admitted and benchmarked (not built now). Six owner decisions with proposed defaults are in its §11, including where dossier work is tracked (#197 is closed; #302 files its dossiers under #300 L2). No live step follows from that issue's text.
holden commentedon Nov 4, 2026
Handoff from #298: dated register, 4 November 2026 (executor A–B)
This is the register the handoff at the top of this issue's body asks for. It is reconciled with #298's final closeout: dated v7 report, B6, B4 outage. #298 closes as FAIL, DIAGNOSED under its owner-approved scope amendment. The experiments that follow are tracked in #305 (Part 1), under this issue. The deferred obligations below stay in this register.
Result handed over
v7 (
voltaire-whole-entry-v7onqwen3.5-9b, model config 8) failed the frozen primary gate on the ten pilot pages. Scorecard v2 (156ee531…) is unchanged.Next work, tracked in #305
Spec; Curation agents: close out v6, retest v7, then controlled model and personality comparisons on identical packets (under #193) #298 §6 D6 and §7 "Policy comparison validity".
Code: Curation: matched comparison support, policy arms and stop guard (#298 Part C, C1–C9) #303 (C1–C9), merged at
2a04126on 4 November and not yet deployed. The deploy is #305 Stage 1: the first persona measurement, Hitchens and Bierce against Voltaire and the generic control, on the current policy and one pinned model (under #305) #307 Stage 1's, on its own yes, after Curation agents: close out v6, retest v7, then controlled model and personality comparisons on identical packets (under #193) #298 closes.Two decisions are still open before the arms are hashed:
abstain_reasonand each choicereasoninsection_output.ex, which Curation: matched comparison support, policy arms and stop guard (#298 Part C, C1–C9) #303 keeps;Both are in the 3 Nov audit, §4.1 and §4.2.
The result names the arm for the v8 publishing retest on the same ten pages, under the unchanged gate (Personas META: from one curator to a weighted panel — the experiments in stages, the dossier pipeline, consensus with weights, humans outside the gate and a later model bench (under #193) #305 Part 1).
4a0b997: Curation: matched comparison support, policy arms and stop guard (#298 Part C, C1–C9) #303, #300 L2: Hitchens and Bierce dossiers as reviewed data files (nothing admitted) #302 and the pin PR Curation: pin the Hitchens and Bierce comparison instructions (#307 P3b) #308, all merged) only after Curation agents: close out v6, retest v7, then controlled model and personality comparisons on identical packets (under #193) #298 is closed; this executor deploys nothing.generic-whole-entry-v1, both on 9B; only the instruction differs.qwen3.5:4b) is an optional diagnostic and was not run in Curation agents: close out v6, retest v7, then controlled model and personality comparisons on identical packets (under #193) #298. It stays optional in Personas META: from one curator to a weighted panel — the experiments in stages, the dossier pipeline, consensus with weights, humans outside the gate and a later model bench (under #193) #305; run it, or record why not, when the next batch is frozen.Deferred obligations register (from #298 §9, reconciled at close)
no_change; no version was written and the pointers did not move. It recurred in cycle 3: justice, attempt 101.timeout:docs/curation/evaluations/) is in #303, which is not mergedservice_unreachableand doubling backoff; no version was written and the pointers did not change; all eight retried successfully after the restart (attempts 97–104). No attempt row exists, because readiness is checked before admission. The service was restarted at 901 s, 1 s past the 15-minute boundfile-write-createforstop.logon the external volume (kernel log from 09:18:26.307 UTC; reads are allowed), so nothing ran. The editor stayed enabled and cycle 4 ran after the window end (attempts 111–121, publications 61–63, not scored) until the stop by hand at 14:12:07 UTC (B6). B2's "loaded, last exit 0" proved loading onlyops/curation/trial-stop/) rehearsed under launchd itself, with its marker and log written in the run folder on the external volume; if launchd cannot write there, move them to a path it can, and record which (the audit's E6)Scratch inventory, read-only, 4 November 2026. Port 5434 holds
devils_dictionary_v2,postgresand 80 scratch databases: 78devils_dictionary_test_*and 2devils_dictionary_runtime_bench*. Four of them belong to work still open:devils_dictionary_test_c298: Curation: matched comparison support, policy arms and stop guard (#298 Part C, C1–C9) #303's partition;devils_dictionary_test_c299: Curator pages: directory, profile and version pages (#299 M1–M7) #304;devils_dictionary_test_c300: Living curation: three to five personas reach a consensus on each visited page, visit-driven, two layers, per-kind staleness later, current events and living people bounded — design and MVP (under #193; after #298, with #299) #300;devils_dictionary_test_s6: dropping it is the owner's step.On the shared 5432 cluster,
devils_dictionary_ex212is a dictionary leftover. The exact names are inrun/v7-six-hour-2026-11-03/closeout/scratch-inventory.txt. Worktrees and branches are listed in the 3 Nov audit §4.7. No deletion is proposed for a database or worktree that open work still uses.Operational lines proven in #298 (for reference; not carried)
/on/love(×3) and/on/bestiality(×1) at about 20:56 UTC on 3 Nov, while both rows were fresh. No refresh or discovery job was inserted and no attempt ran between 20:45 and 21:20 UTC