Repository navigation
Description
Delivery child of #193. Follows the runtime evidence from #195 and extends the persistence foundation from #196/#201, now merged in PR #206. Reviewed manual evidence packets allow the first experiment without claiming #101's broad assessor is implemented. #190 precedes bot illustrates writes. #189 human voting UI is not a dependency.
Publication contract — authoritative clarification
Routine composition publication requires no per-selection human acceptance. A genuine finalized panel decision on the exact composition version, meeting the configured 4/5 consensus and passing source/evidence/rights/meaning/policy checks, authorizes panel_decision publication. Human operator review is an optional path for manual compositions, overrides/corrections and withdrawal, and an explicit operator rejection remains a veto. Configuration/profile admission and controlled pilot evaluation are separate operator activities; neither creates a human gate on every routine selection.
Implement the decision authority described in the merged docs/curation/persistence-slice-1.md, extending existing records rather than adding a second publication system. Freeze configuration/profile/model/policy versions, exact candidates and source revisions; preserve actual ballots and per-round statuses. Finalize an immutable decision for the exact arrangement/reasons/evidence/meanings, with recomputable counts and selection-v1 trace. Add the receipt's decision reference and accountable pipeline actor, mutually exclusive with operator-review authority. Apply the same transactional pointer, idempotency, scope/configuration/eligibility checks and read-time withholding. Reject missing, forged, stale, cross-run, cross-version or cross-configuration decisions, and reject late responses that would alter a finalized result. Extend the reader to recognize genuine panel authority; do not make it depend on the existence of an operator acceptance.
A panel decision never accepts a semantic claim. Pending/rejected/withdrawn exemplar claims remain subject to #190, independently of editorial selection. Failure to reach consensus preserves previous eligible content or leaves the slot empty. Neither missing votes nor simulated personas count as genuine decision evidence.
The next implementation task is #195 runtime stage A. Sourced dossier research may proceed separately, but panel implementation and activation require that runtime's measured readiness. No pretrained weight training or per-person model installation is implied.
Configurable roster
Pilot: the single system-owned global-default configuration in #201 uses Bierce, Voltaire, Vonnegut, Hitchens and Ursula K. Le Guin, all explicitly inspired-by AI. All offered persona subjects must be deceased, with sourced identity/deceased-status admission checks; new names remain data, not name-specific branches. Le Guin is the recommended fifth profile to ground in her writing, not a historically verified position on modern examples. Suggested lenses to test: Bierce notices hypocrisy; Voltaire challenges dogma; Vonnegut foregrounds human consequences; Hitchens presses evidence and consistency; Le Guin examines power, social alternatives and overlooked perspectives. These lenses are hypotheses for dossier research, not fabricated quotations or a guarantee of ideological independence.
Each versioned dossier contains interests, selection principles, original-source passages with edition/translation/rights, positive and negative selections, blind spots, counterexamples, abstention criteria and voice guidance. Allowed quotations must be verified; rights-restricted passages are not copied into unrestricted logs. Profiles affect which material is selected, not just prose. Use short retrieved feedback, not unbounded chat memory; improvements require reviewed new versions, not automatic self-training or fine-tuning.
Registry-driven orchestration resolves a named reusable curation_configuration and its immutable version/roster before freezing actual run participants. #196 stores configuration versions/membership separately from profile identity/version, actual participation and model weights. The MVP resolver selects only global-default for every enabled subject, records reason=global_default and the resolution-policy version, and never interprets a URL parameter as an alternate panel request. Freeze configuration version, actual profile versions, model/template/policy and evidence inputs per run; changing any current pointer mid-run cannot alter the electorate.
A future roster of ten can be represented in configuration tests without hardcoded names or loading one model per person. Domain-based routing and multiple exposed panels belong to #201 Stage 2, and personal preferences/saved views to Stage 3; neither ships in this pilot. Future routing is based on versioned stable-ID evidence/assignments, with deterministic overlap handling/global fallback, not public namespaces or URL spelling. Historical persons and bot actors are separate; endpoint/model configuration remains independent from dossiers. The configured pilot stays exactly five participants.
Bounded protocol
A packet targets a stable composition/page, with at most 12 eligible existing API/corpus candidates, exact source/sense revisions and current decisions. It can span meanings; every selected candidate names its intended sense/concept. The ten-page pilot manually bounds these meanings; it does not silently force the whole page to its first sense. Bierce-first lead selection is the current system-default configuration's versioned editorial rule from #156, not something a majority can vote away. Future explicit policy versions may differ in editorial preference; mandatory source/evidence/rights/claim gates cannot be overridden.
- Round 1: each of five profiles independently assesses the same packet, proposes highlights/optional exemplar nominations and records support/oppose/abstain plus concise evidence-based reasons.
- Freeze the candidate set; merge proposals deterministically. Round 2: each profile sees concise competing reasons/objections and returns final ballots for the same bounded candidates. No third debate round or autonomous search.
- Voting weight = 1 for each of the five. Four of five support recommends a candidate, provided no unresolved evidence/policy blocker remains. Count valid final-round support only; round 1 is history, not extra votes. Eligibility, ordering preferences and votes are distinct, with no composite truth score or confidence percentage. An optional operator decision is separate from panel consensus, not a sixth vote or a routine publication prerequisite. Three–two is disagreement; abstentions, missing/invalid responses and offline profiles never silently shrink the denominator. Five voices can still deadlock. Four valid supports with one missing response can qualify under the frozen rule; missing is disclosed. Ballots remain editorial recommendations, not independent evidence or human votes.
- On ties/disagreement, keep previous eligible published content or leave the slot empty; the bounded run completes with its recorded outcome. Invalid evidence fails server checks regardless of unanimity. Deadline expiry ends incomplete rounds as deferred, with missing/invalid response status recorded separately; only an actual valid abstention ballot is an abstention. Never invent ballots.
- Apply Curation transparency: equal votes, reproducible selection, profiles and version history #203's declared
selection-v1to eligible 4/5 recommendations: repeatedly choose highest final support; within that group lowest frozenrelation_basisorder (direct_target=0,located_usage=1,reviewed_interpretation=2); then prefer a content type not already selected; then lexicographically smallest canonical candidate key using fixed byte ordering. Append and remove duplicatedisplay_identity_keyalternatives; stop at three or no candidates. The MVP packet preparer human-reviews the relation category and evidence before voting; it is an editorial annotation, not a calibrated model score. Unknown/unvalidated category is held for review. Store policy/version, inputs and selection trace so reordering packet input cannot change the result. Curation transparency: equal votes, reproducible selection, profiles and version history #203 defines category/evidence boundaries and dedup keys. - A reviewer may create a new composition version with a required reason and explicit
human_override/human_addedorigin; display its difference from the panel recommendation. A manually added item has no panel support unless actually voted on. Overrides still obey source/rights/claim gates and the active default's mandatory lead policy. Reported generated reasons are concise explanations, not exhaustive internal cognition; never generate historical quotations, debate transcripts or hidden chain-of-thought for display.
Five profiles × two rounds = ten logical generation calls. At most one repair generation across the panel gives 11 physical calls maximum. Budgets, timeout and retry rules are #195's. At most three highlight proposals plus one composed lead; at most three linked illustrates nominations may be proposed per packet, reserving at most four pending-work units including the unfinished composition, as defined in #195; completed consensus publication releases the composition unit without human sign-off. Don't create an exemplar claim simply to feature a sourced picture or passage.
Sensitive real-person interpretations require specific evidence and explicit human claim review; no model diagnosis from public behavior or consensus-as-fact. Tests use fictional adverse subjects. Routine composition publication follows the finalized panel-decision authority above; rejected nominations stay rejected until explicit human reopening.
Experiment and competition
Ten pilot pages including love, war, bank, nepotism, grief, justice, mountain, situationship, logomachy, and a sparse/fictional adverse-case fixture. Prepare source-supported eligible packets, verifying actual availability rather than inventing filler. Use five pages for dossier tuning, five sealed for evaluation; keep related passages/near-duplicates out of both splits. Label intended meanings and rights before inference.
Compare a generic editor and the five-profile panel on identical packets, model version and output contract. Record that panel compute is higher; optionally add an equal-compute generic baseline. Randomize/blind reviewer presentation where practical. Score useful selections, source-supported relevance, wrong-sense rate, duplicates, type diversity, fabricated attribution, disagreement/abstention, reviewer effort and accepted useful highlights per machine-hour. Measure whether individual profiles select differently and whether the panel improves judgment, not merely more colorful language. Count failed attempts and refusals in denominators.
Pre-register pilot expansion gates: zero fabricated quotations or public unaccepted claims; all published items pass source/rights checks; at least 90% reviewed proposals support the intended meaning; panel preferred on at least three of five held-out pages, with ties not wins. Publish absolute counts, uncertainty and total machine/reviewer time. These small-sample gates support another bounded experiment, not a general accuracy claim. If the panel costs more without useful improvement, keep manual/generic composition and revise the dossiers before expanding.
Acceptance
- Five grounded dossiers and ten packets reviewed/versioned; one generic baseline and panel comparison report including held-out results.
- Named configuration/version/membership resolves to frozen actual participants with an inspectable default-resolution reason. A five-to-ten roster representation in tests needs no orchestration rewrite; the production pilot remains five and global-default only, with one shared cost cap.
- Tests cover equal weight 1, 4–1 recommendation, 3–2 disagreement, abstention/missing denominator, blocker veto even at 5–0, exhausted round, one-repair cap, stable previous/empty output, category/diversity/key ties and duplicate display identities. Permuting frozen candidate input leaves order unchanged (Curation transparency: equal votes, reproducible selection, profiles and version history #203).
- Show a useful consensus selection, inspectable dissent and honest abstention; genuine panel decisions can authorize routine publication, while bots cannot impersonate an operator, manufacture votes or accept semantic claims. Prove missing, forged, stale, cross-configuration and operator-vetoed authorities are refused at service and database boundaries.
- Published pilot composition retains source/AI attribution and respects Curated opening section: Bierce-first lead, sourced highlights and inspectable editorial choices #156/Gate every illustrates nomination until accepted before bots or the public can nominate (#181 decision 6, deferred) #190.
mix precommiton implementation.
Configuration integration (#201): reader inspection may show which inspired-by profiles actually selected material and why, plus config/version provenance; it is not a domain/personal panel picker. Keep ten allowlisted pages, 4/5 support, ten normal calls and at most one repair unchanged. Source corpora and semantic claims may be shared underneath, but run ballots/presentation results never cross configuration boundaries. No fine-tuning, accounts or new models per configuration.
Accountable output (#203/#204)
Keep each actual proposer's identity when merging duplicate proposals; record response status per participant per round (including missing/invalid), safe declared reasons, evidence/category annotations and computed counts/selection trace. #196 owns immutable records; #203 exposes an authorized version-specific projection with human changes and publication timestamps, not an unrestricted log. Profiles show sourced influences, selection rules/examples/counterexamples, blind spots, changes and permitted selection history. #204 explains the method and its limitations; neither Zinn nor Gray joins this pilot.
Unequal persona-prestige weights are out of MVP. Any later proposal needs explicit reviewed configuration version, actor/change reason, exact formula and quorum/abstention semantics plus evaluation; one shared model is not independent corroboration. Limits remain five participants, ten normal calls + max one repair and one global compute budget.
Activity
Repair the curation audit findings and name every publication's autho…
holden commentedon Oct 21, 2026
Folded into #193, the curation programme's one open tracker, on the owner's word (21 October 2026). The specification here stands; reopen when the work starts. Status is in the consolidation comment on #193.
holden commentedon Nov 4, 2026
Reopened 4 November 2026 on the owner's words in chat (#305 OD-1, on an independent critique's recommendation): this issue tracks persona dossier research and review only, S1–S3 of #305 §4.3, starting with Vonnegut and Le Guin as data-only PRs from the #302 template. Completion for each dossier: every passage verified as an exact substring of its source, rights recorded, an independent exact-head review with nothing REQUIRED, merged as files with nothing admitted. Measurement, admission and seats stay in #305 and #300. The specification below stands as written on 21 October.