Skip to content

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

Open

Description

@holden

Implementation and design issue, 4 November 2026, under #193. Posted on the owner's words in chat, 4 November 2026 ("Post it"). #298 closes today as FAIL, DIAGNOSED under its owner-approved amendment, which moved Part D, the v8 run and the post-comparison checkpoint to "follow-up work under #193" without an issue to carry them. This issue carries them (Part 1, build now). It also records, in the owner's words of 4 November, how the persona programme is meant to work from here: how a dossier is built and earns a seat (Part 2), how agents reach consensus and how seats can be weighted (Part 3), how humans weigh in (Part 4), and how a second model would be benchmarked against the first (Part 5). Parts 3 to 5 are the thought process, written down once so #300 and later issues build the general form; they are not a brainstorm and only their named first steps are built here. No live step follows from this text.

Revision 1, 4 November 2026, after an independent read-only critique (B− as a design brief, C+ for implementation readiness) and the owner's clarification on human intervention. Four corrections applied: one finish line (§§9–10: this issue closes on the experiment reports, the v8 disposition and one read-only audit; Parts 3–5 stay design references and get their own issues when selected); a promotion rule decided before D6 (§3.1, OD-7), with a terminal state for offline evidence that does not justify v8; the persona comparison isolates the lens (E4 rewritten); an executable budget, stop and admission plan (§3.2). Part 4 is rewritten: humans are not a gate. The loop publishes without approval; the owner's intervention is occasional, per entry and time-limited; reader up-and-down votes are a later, bounded layer, never a veto. Phases A–C below replace the earlier flat list.

Revision 2, 4 November 2026, after the re-audit (A− design, B+ readiness). Runbook clarifications only, no new sub-issues: promotion needs all ten complete matched cases and an incomplete batch is INCONCLUSIVE, distinct from NOT JUSTIFIED (§3.1); the promotion order is total, with cost defined and a declared-order fallback, and the bad-reason floor counts every origin, a wrong rule statement being a disqualifying tooling defect (§3.1); the love gate requires a useful page and an early stop there is classified with its evidence (§3.2 step 6); re-templated comparison instructions get fresh immutable names, shas and model configs, and the Voltaire dossier version is resolved before the Phase C freeze (E4). The first D6 call waits for these and for #298's closure and #303's deployment.

Revision 3, 4 November 2026 (owner's words in chat): this is the persona META issue, and its experiments run as numbered stages, each a dedicated implementation issue. Stage 1 is the first persona measurement, #307, on the policy live today (v7's), run first because the lens question does not depend on the policy question; Stage 2 is the four policy arms offline, #306; Stage 3 is the conditional v8 publishing run, under this issue until a Stage 2 arm qualifies. "Phase A, B, C" in the revision notes above are Stages 2, 3 and 1.

0. The owner's words (4 November 2026)

  • On models: "right now we just have the one, it'd be good to have another one"; "we don't need to build it now, but just a thought process of eventually we would like to benchmark LLMs to see how they compete".
  • On consensus and weight: "how we reach consensus between different agents"; "a way that we can weight different agents separately, meaning different personas separately, so we could perhaps decide to give more weight to a particular persona because it's done well in the past, or … to me as the neutral persona".
  • On what curation is: it "is always building on top of what we have … taking the API layer, which is the layer zero, and then building on top of it to shape it in terms of the information presented, but never the information necessarily itself, but to suggest examples of things, to elevate a particular book, poem, whatever it is".

1. Outcome, and the one invariant

Outcome. One Voltaire-inspired curator that passes the frozen quality gate on the ten pilot pages (the v8 run), measured against a generic control and against a second persona on identical evidence; then a panel whose seats can carry different weights, deciding by weighted overlap on one frozen packet, with humans outside the decision (occasional, time-limited intervention now; a bounded reader-vote layer later); and a recorded way to admit and benchmark a second model. Every step keeps the record that already exists and adds ids to it.

The invariant (layer 0 and layer 1). Layer 0 is the objective record: the definitions, senses, quotations, works, artworks, films, news items and their provenance that the sources and APIs give us. Nothing in this issue writes to layer 0. Layer 1 is subjective and keeps changing: which of those records a page shows first, which examples are elevated, and why, signed by the lens that chose them. Personas, panels, weights, humans and models all live in layer 1. A persona never writes a definition, a quotation, a claim or an entity; it chooses among candidates the packet already holds, names the meaning each choice speaks to, and gives a reason a reader can judge. Expiry, rights, meaning and eligibility are layer-0 facts and cannot be outvoted.

2. Where things stand (verified read-only on the Studio, 4 November 2026, 03:30 UTC)

3. Part 1, the experiments in three stages (build now; this is what #298 left untracked)

Prerequisites, each on the owner's yes: #298 closed; #303 merged and deployed (migration 20261103202412 plus fast-forward, no restart); #302 merged (files only). The editor stays disabled for every comparison batch.

Stages: Stage 1 = E4, the first persona measurement (#307), first; Stage 2 = E1–E2, the policy arms offline (#306); Stage 3 = E3, the conditional v8 run; then E5, E6 and the single audit close this issue.

  • E1. D0 batch manifest and guard. Before any call: code SHA, arm identities (instruction and dossier versions with hashes, model config ids and hashes, generation settings), the ten pages, the case identity scheme (BATCH:SLUG:n), the evidence-fingerprint rule, scorecard v2 sha 156ee53110b7c8ecad98291a8087c91b831e2a07c1c47f19aca4e269ee7cd1a9 unchanged, the batch deadline and attempt cutoff. Load the C6 guard from ops/curation/trial-stop/ after its dry run. Frozen file in run/d6-<date>/.
  • E2. D6, the four policy arms. mix dd.runtime compare --batch d6-<date> --pilot --arms A0=…v7,P=…v7:text,L=…v7-fixed-lead,PL=…v7-fixed-lead:text (dry run first). Forty calls, about ten minutes on the 9B. Score every proposed choice of every arm within 20 hours with scorecard v2 from mix dd.curation.coverage --comparison d6-<date> --record PATH. Report the three measures separately, never combined: useful pages (of 10, bestiality included), supported meaning across all proposed choices, unsupported or misattributed reasons (model-written counted apart from rule statements); plus lead correctness and highlights by kind. Unmatched cases reported, not dropped. Decision by the rule in §3.1, recorded in one comment before anything else runs. This ends Stage 2 (#305 Stage 2: the four policy arms offline (D6), runbook and execution (under #305) #306).
  • E3. Stage 3, the v8 publishing run on the same ten pages (only if a Stage 2 arm qualified). The chosen arm becomes one immutable instruction voltaire-whole-entry-v8 (its packet policy named in the model config), one profile version (dossier v9 if the lens text changes, else the v8 file re-admitted under the new template), one model config on the unchanged 9B artifact. Manifest frozen before calls with cutoff equal to the max attempt id; fresh C6 guard with the deadline; love gate while the other nine are held; release; window 18 hours, three 6-hour cycles, fixed stop (OD-2). The first assessment per page is primary; repairs separate; cycles 2 and 3 are operational evidence only. Judgments within 20 hours of each attempt; developer minutes measured. Stop verified (STOPPED marker, editor NULL, listener down, final attempt id), external bundle captured and verified, restore-checked; the internal second copy as the owner decides for this run (it was waived for v7 only). Verdict on the unchanged §7 gate of Curation agents: close out v6, retest v7, then controlled model and personality comparisons on identical packets (under #193) #298. A failed cohort stands; never rerun to replace it.
  • E4. Stage 1, the first persona measurement (#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): Voltaire, the generic control and Hitchens, with Bierce because it holds a seat in LP-1. Runs first, on the policy live today (v7's evidence policy, under which #300 L2: Hitchens and Bierce dossiers as reviewed data files (nothing admitted) #302's Hitchens and Bierce instructions were written), so no template is edited and no stage waits on another. The treatment is the editorial lens and nothing else. Every arm shares, pinned in the D0 manifest: v7's policy and packet cut, model-selected leads for every arm under the same rule, the output schema, the model artifact and the generation settings. Each arm is one named instruction/dossier pair: voltaire-whole-entry-v7 with Voltaire profile version 7; generic-whole-entry-v1, no dossier; hitchens-whole-entry-v1 with Hitchens v1; bierce-whole-entry-v1 with Bierce v1; a v1 name is never edited. The baseline is the Voltaire arm. One snapshot per page, equal fingerprints across all four arms, run within minutes, nothing published, every proposed choice of every arm scored within 20 hours. Report per arm the three measures separately and lead outcomes apart from highlights. Forty calls: the baseline runs inside each case so every arm shares its snapshot. Bierce's hypothesis is written in #300 L2: Hitchens and Bierce dossiers as reviewed data files (nothing admitted) #302's bierce-v1.md; that, and its LP-1 seat, are why it runs now. If Stage 2 later changes the policy, the seat recommendation is re-checked on the new policy before any seat is created (fresh v2 template names with their own shas and model configs in one small reviewed PR, then a 40-call repeat); Stage 1's record stands as the first lens measurement either way.
  • E5. D1, the model arm (optional diagnostic, per the Delivery: curated opening section with five local persona editors and visit-driven refresh #193 handoff). Register qwen3.5:4b under the baseline instruction; matched comparison on the ten pages, twenty calls (ten 4B plus the ten matched baseline calls). Run only if 15 minutes of the day's budget remain after E4; otherwise record "not run because …". Its result feeds Part 5, never E3.
  • E6. D4 decision and close. One comment: the model config and instruction for the next publishing experiment; whether a second persona is proposed for admission (OD-B in Curation agents: close out v6, retest v7, then controlled model and personality comparisons on identical packets (under #193) #298: data-driven lookup of admitted profile versions and instructions, not another literal list); the disposition of D1. Then the single read-only audit from the owner's session (were the manifests frozen before calls, cases matched by fingerprint, arms scored on all proposed choices, the decision supported by the counts) and closure.

3.1 The promotion rule, frozen before D6's first call (OD-7)

An arm qualifies for the v8 publishing run only if, on the D6 batch, it meets all three floors offline: useful pages ≥ 7 of 10 (the publishing gate is 10 of 10; an offline batch has no repairs and bestiality may lawfully abstain), supported meaning ≥ 90 percent of all its proposed choices, and zero unsupported or misattributed reasons of any origin (model-written reasons and rule statements are reported apart, and both count, as the live gate counts every published explanation). A wrong rule statement (a fixed lead that is not the applicable priority candidate, or a statement that misdescribes it) is a tooling defect: that arm is disqualified for the batch and the defect is fixed before any new batch. Completeness: promotion needs all ten cases complete and matched, four valid arms each. A page retried once as BATCH:SLUG:2, all four arms on a new snapshot, enters the scored set as that page's case; its first, unmatched attempt stays recorded and unscored. A batch with fewer than ten complete matched cases after the one retry is INCONCLUSIVE: no decision, the figures kept, the cause fixed, a new batch id run; missing evidence is never recorded as NOT JUSTIFIED. Among qualifying arms the order is total: more useful pages, then fewer bad reasons, then higher supported rate, then higher lead correctness, then lower cost, where cost is the charged milliseconds of the arm's ten attempts from the inference ledger, then fewer prompt tokens, then the manifest's declared order A0, P, L, PL (least change first). A0 wins any tie it is part of, since a change must earn its run; if A0 fails the floors the declared order settles the last tie. v7 shows why the order matters: 9 of 9 supported, one bad reason, one useful page, and it would not qualify. If no arm qualifies the issue takes the terminal state NOT JUSTIFIED: no v8 run, the batch's figures recorded, and B5's remaining choices (a narrower instruction, or a model change informed by E5) decided in the next dated issue. The rule is cited by the D0 manifest and cannot change after the batch's first call.

3.2 Budget, stop and admission: the executable plan, written before the first call

  • Budget. Observed calls are 6–14 s, so a 40-call batch is 4–10 minutes; the worst case, 40 × 120 s = 80 minutes, exceeds the 30-minute day. Rules: a batch starts only when the day's charged time is under 5 minutes; the compare task charges the shared budget and halts at exhaustion with the remaining pages not_run; a case is atomic (every arm of a page back to back), so a halt inside a case leaves it incomplete and unmatched, and the full page-and-arm set is retried once on a new snapshot the next day as BATCH:SLUG:2; Stage 1 (#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) and Stage 2 (#305 Stage 2: the four policy arms offline (D6), runbook and execution (under #305) #306) run on different days; E5 runs only with 15 minutes left after E4. An uncertain call halts the batch; nothing is retried blind.
  • Expiry inside a case. The snapshot records the earliest shelf expiry; a case whose expiry passes before its last arm is unmatched and retried once. News (23 h) is the usual cause, so a batch starts early in a news window and never within an hour of one.
  • Stop. Every batch runs under the C6 guard, loaded with the batch deadline after --dry-run wrote DRY_RUN_OK; the editor is disabled throughout; the batch report records the guard's final state and the last attempt id (D5).
  • Before v8 (Stage 3), each step on the owner's yes, in this order: (1) the instruction voltaire-whole-entry-v8 admitted into the runtime allowlists and into the database CHECK and author guard, in one small reviewed PR with one migration (or OD-B's lookup if it has landed first); (2) the model config registered on the unchanged artifact; (3) the profile version admitted; (4) the manifest frozen with cutoff = max attempt id; (5) the guard loaded and dry-run; (6) hold 2–10, enable the editor, love gate: one primary attempt on love, judged under scorecard v2 before anything is released; the gate passes only if love is scored useful (a supported lead and at least one supported highlight, every reason grounded); a lead-only love fails it, which is the v7 failure mode and must not release the other nine pages. An early stop at the gate is recorded as FAIL AT THE GATE: the editor is disabled at once, compositions 2–10 are never released, the stop is verified as at any end (STOPPED marker, editor NULL, listener down, last attempt id), the love attempt's judgment is posted as the run's report, the external bundle is captured because a version was published, and the diagnosis obligation of FAIL, DIAGNOSED applies; (7) release 2–10. (8) After the stop: external bundle captured and verified, restore list, schema-only scratch restore created and dropped, each its own yes; the internal second copy is its own decision for this run.

Gates (copied from #298 §7, unchanged). Quality GO for a publishing run: supported / all proposed choices ≥ 90 percent with wrong and insufficient in the denominator; zero unsupported or misattributed published explanations; ten useful entries including bestiality on supplied sense definitions; every refusal, abstention and repair retained; costs recorded. Matched comparison validity: same page, one snapshot per case, equal fingerprints, arms differ in exactly one of instruction or model config, seed and temperature fixed, all proposed choices scored, no arm publishes, unmatched pages reported. Policy comparison validity: one frozen source snapshot per batch, arms differ only in cut and lead rule, usefulness, supported meaning and bad reasons reported separately.

Live steps, each on the owner's named yes in chat, one at a time: deploy #303; register model configs; admit profile versions; load or unload a guard; enable the editor; stop a run by hand; capture a bundle; drop a scratch database; delete anything.

Terminal states, one of which closes this issue. SUCCESS: an arm qualified, v8 PASSED on the unchanged gate, E4 reported on matched cases, D4 names the next publishing configuration. FAIL, DIAGNOSED: an arm qualified and v8 FAILED; one bounded diagnosis (by kind and section, lead against supplementary, "metadata cannot support a claim" against "model ignored evidence") and one decision, opened as the next dated issue. NOT JUSTIFIED: no D6 arm met the §3.1 floors; the D6 report and decision posted; Stage 1's lens measurement stands; the next dated issue opened for B5's remaining choices. STOPPED: the owner stops; every finished part's evidence posted, every unfinished part marked not run. Stage 1 runs in every path except STOPPED. INCONCLUSIVE is a batch disposition, not a terminal state: the batch is run again under a new id once its cause is fixed. A stop at the love gate (§3.2 step 6) is a FAIL of the v8 run and takes the FAIL, DIAGNOSED path.

4. Part 2, the dossier pipeline: how a persona is made and earns a seat

4.1 Who, and the rule for adding anyone

The roster was decided in #97 and #201 and confirmed in #275 §4: Bierce, Voltaire, Vonnegut, Hitchens, Le Guin, all deceased, all labelled "inspired by", none claimed to be the person. Selection hypotheses to test, not opinions attributed to anyone: Bierce, hypocrisy, self-interest and satirical contrast; Voltaire, dogma, intolerance and pretension; Vonnegut, human consequences, absurdity and compassion; Hitchens, evidential support and consistency; Le Guin, power, social alternatives and overlooked perspectives. Names discussed and not on the roster: Zinn and Gray are references for the method page (#204), Johnson is a source of definitions that can lead a page (#253 D3), Monet and Picasso were examples of later domain panels (#201). Decided elsewhere and kept: OD-A, Hitchens is the second persona.

Rule for any new name: deceased, with an authority record in deceased_evidence_refs; original-source passages with edition, locator and rights; a lens that changes what is chosen, not how it is phrased; a documented blind spot; an abstention rule; a reason rule for living people (quote or name the supplied source, never characterise the person). A living person is never a persona. A name is data in a dossier file, never a branch in code.

4.2 Anatomy of a dossier (the shape the v8 Voltaire file and #302 already use)

priv/curation/agents/personas/<slug>/v<n>/profile.json
┌───────────────────────────────────────────────────────────────────────┐
│ slug                 "hitchens"                                       │
│ template_version     "hitchens-whole-entry-v1"   (the instruction)    │
│ dossier                                                               │
│   disclosure         one sentence: an AI lens inspired by …, not his  │
│                      voice or his opinions on modern events           │
│   selection          the lens in one paragraph: what to prefer, why   │
│   principles[]       2–4 short rules the lens applies                 │
│   blind_spots[]      what this lens tends to miss or overdo           │
│   positive_example   one supplied-candidate case it should choose     │
│   negative_example   one it should refuse, and why                    │
│   lead clause        "the lead is set by the lead rule, not by this   │
│                      lens; apply the lens to highlights only"         │
│   living-person rule "quote or name the supplied source; never        │
│                      characterise the person"                         │
│ source_refs[]        url · work · edition · locators · rights · basis │
│                      · passages (exact substrings, checked by script) │
│ deceased_evidence_refs[]  url · authority · field · value             │
└───────────────────────────────────────────────────────────────────────┘
docs/curation/personas/<slug>-v<n>.md   one page: sources, rights, what is
                                        documented vs our interpretation,
                                        the hypothesis to test

Short compiled passages go into the prompt; the full dossier never does. Rights-restricted passages are referenced, not copied into unrestricted logs. A dossier edit is a new immutable version; old runs keep the version they used.

4.3 The seven stages, and the evidence each leaves

Stage What happens Evidence Who
S1 Research Data-only PR from the #302 template; every passage checked as an exact substring of its source; rights recorded PR, check script output executor
S2 Review Independent exact-head review (nothing REQUIRED), CodeRabbit; the critique checks lens, rights, blind spots, lead clause review comment id reviewer session
S3 Merge Files only; nothing admitted, registered or bound merge SHA owner's yes
S4 Measured Comparison arm against the baseline on matched evidence (Part 1 E4 pattern); scored on all proposed choices comparison batch id, report executor
S5 Admitted Immutable profile version, instruction and model config; admission by lookup (OD-B), not a literal-list migration profile version id, config hash owner's yes
S6 Seated A seat in a new roster version with a weight (default 1) configuration version id, roster hash owner's yes
S7 Shown /curators/<slug> shows the dossier, the measurement and every attempt #299 page automatic

Next dossiers: Vonnegut and Le Guin, S1–S3 now, in two data-only PRs from the #302 template; S4 after #303 is deployed. Home for this work: OD-1.

4.4 Wireframe W1: what one seat sees and returns (one call)

┌─ system instruction (immutable, sha recorded) ─────────────────────────┐
│ whole-entry policy v7+: sections and caps (lead; up to 2 examples,     │
│ 2 quotations, 1 each of the rest), evidence rule (a title alone is    │
│ not evidence; sense definitions are leads), lead rule (Bierce first), │
│ abstain when nothing fits, JSON contract selection-output-v4          │
├─ compiled lens (from the dossier version, sha recorded) ───────────────┤
│ disclosure · selection · principles · blind spots · 2 examples ·      │
│ 1–3 short sourced passages (rights permitting)                        │
├─ packet (frozen; packet hash + evidence fingerprint recorded) ─────────┤
│ page: greed  · meanings m0..m11 (ids, labels)                          │
│ c1  definition  Bierce "GREED, n. …"            m0   basis: entry      │
│ c2  quotation   "… (Johnson, Dictionary 1755)"  m0   basis: sense      │
│ c3  work        Dante, Inferno, canto VII       m2   basis: catalog    │
│ …  (24 candidates; omitted ones listed by reason on the attempt)       │
└────────────────────────────────────────────────────────────────────────┘
                         │  one call, ≤ 120 s, 1,024 tokens
                         ▼
┌─ answer (validated field by field; any failure = refused) ─────────────┐
│ { "decision": "select",                                                │
│   "lead": {"candidate": "c1", "meaning": "m0", "reason": "…"},         │
│   "highlights": [ {"candidate": "c3", "meaning": "m2", "reason": "…"}, │
│                   … ],                                                  │
│   "abstain_reason": null }                                             │
└────────────────────────────────────────────────────────────────────────┘
→ inference_attempts row: seat, profile version, model config, packet
  hash, fingerprint, prompt sha, outcome, cost. A single editor publishes
  from it; a comparison arm never does; a panel seat feeds the decision.

4.5 Wireframe W2: the persona comparison report (S4, from mix dd.curation.coverage --comparison)

batch p-2026-11-06 · snapshot per page · scorecard v2 156ee531… · editor disabled
page       arm        attempt  matched  proposed  supported  insufficient  wrong  bad reasons  useful  s/call
greed      voltaire-v8  121     yes      3         3          0             0      0            yes     8.9
greed      generic-v1   122     yes      2         1          1             0      1            no      7.4
greed      hitchens-v1  123     yes      3         2          1             0      0            yes     9.3
greed      bierce-v1    124     yes      2         2          0             0      0            no*     8.1
…
totals     voltaire-v8           10/10   27        24 (89%)   3             0      1            7/10
           generic-v1            10/10   19        13 (68%)   5             1      3            3/10
           hitchens-v1           10/10   26        22 (85%)   4             0      1            6/10
           bierce-v1             10/10   21        19 (90%)   2             0      0            5/10
* lead-only pages are not useful; bierce's lead is the rule's, counted apart
Figures above are illustrative shapes, not results.

4.6 Wireframe W3: the "measured" block on a curator page (#299 addition)

├──────────────────────────────────────────────────────────────────────────┤
│ MEASURED BEFORE ADMISSION                                                │
│ Compared with Inspired by Voltaire v8 and the generic editor on the     │
│ same ten pages and identical evidence (batch p-2026-11-06):             │
│   supported meaning 22 of 26 proposed · 1 unsupported reason ·          │
│   6 useful pages · 9.3 s per call                                        │
│ Admitted 8 Nov 2026 as version 1 · seat weight 1 (roster v11)           │
│ Every figure can be reproduced by the read-only query in PR #3xx.       │
├──────────────────────────────────────────────────────────────────────────┤

5. Part 3, consensus between agents, with weights

Design reference, not acceptance for this issue. Built in #300: L1 proves equal-weight overlap on fixtures first; non-default weights are a later #300 phase, after L4's report and a seat performance report. This section fixes the general form once so L1 is not rewritten later.

5.1 Start from #300 §4 and generalise it once

#300's MVP is overlap, not debate: one round, every admitted seat answers the same frozen packet independently, a highlight is published when at least k of N seats chose it, the lead is set by the rule and never voted, missing or refused answers are zero and never shrink N, and a blocker (eligibility, expiry, meaning, rights) vetoes anything. Keep all of that. Generalise the count into a weighted sum so that the equal-vote case is the default and nothing else changes:

  • every seat s in a roster version has an integer weight w_s ≥ 1;
  • support(c) = Σ w_s over the seats whose validated answer selected candidate c for the same meaning;
  • c is recommended when support(c) ≥ T and no blocker remains; T is set in the same roster version;
  • missing, refused, failed and abstained seats contribute 0; T never moves within a run;
  • with every w_s = 1 and T = k this is exactly 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's rule (k = 2 of 3, 3 of 4, 4 of 5), so L1 builds the general form and ships with the special case;
  • a decision needs support from at least two distinct seats, whatever their weights (solo: false, the default); letting one named seat carry a decision alone is a later owner decision (OD-3);
  • order among recommended highlights: support, then the existing selection-v1 categories and section caps;
  • every supporter's reason is kept; seats below T are shown as dissent with their reasons;
  • one model playing several lenses is correlated judgment, and the receipt says so.

Weights, the threshold and the solo rule live in the configuration version's manifest and in curation_configuration_members.voting_weight, which already exists. They change only by a new roster version with a named actor and reason, as #203 requires ("exact formula, threshold, abstention semantics, named change actor and reason, before/after evaluation and public explanation").

5.2 Earned weight: a report proposes, the owner decides, nothing is automatic

"Done well in the past" is measurable with what exists: every scored attempt of a seat under scorecard v2 gives its supported-meaning rate, its bad-reason count and its useful-page count. A seat performance report (an extension of the coverage report, read-only) lists, per seat and per instruction version, those figures over the last N scored attempts and proposes nothing more than the figures. The owner may then issue a new roster version raising or lowering a weight, with the report linked as the reason. No code changes a weight on its own; a persona cannot vote on weights; a weight is not a judgment of a writer's worth, only of a lens version's measured record.

5.3 Arithmetic, so the rule has no surprises

Roster version Seats and weights T What publishes
LP-1 (#300) Voltaire 1, Hitchens 1, Bierce 1 2 any two of three
Five equal five seats at 1 4 #203's 4 of 5
Earned Voltaire 2, Hitchens 1, Bierce 1, Vonnegut 1, Le Guin 1 4 Voltaire plus any two, or any four of the others
Readers layer (later, §6) the five above plus readers capped at 1 4 readers can complete a decision three seats support; never alone, never against a blocker
Missing seat Hitchens missing (refused or offline) 4 counts as 0; T stays 4; the rest decide or no_consensus

5.4 Wireframe W4: a receipt under weighted overlap

┌ /on/greed/receipt/9 ─────────────────────────────────────────────────────┐
│ Published 9 Nov 2026 14:02 UTC by panel decision (panel_v1)              │
│ roster v11: Voltaire (w2) · Hitchens (w1) · Bierce (w1) · threshold 3    │
│ at least two distinct seats · readers layer off · no human approval      │
│ lead  Bierce, GREED, n. — by the lead rule, not voted                    │
│ ★ Inferno, canto VII (Dante)      support 3 of 4: Voltaire 2, Bierce 1   │
│   meaning m2 · work                   reasons ▸                          │
│   not chosen by: Hitchens — "the Gutenberg text is a translation;        │
│   prefer a sourced passage with its translator named" (dissent, kept)    │
│ ★ "Avarice …" (Johnson, 1755)     support 4 of 4 · reasons ▸             │
│ proposed below threshold (dissent, not published):                       │
│   news item 212 — Hitchens only (1 of 4): "reported by …, 8 Nov"         │
│ one model, three lenses: correlated judgments, not three witnesses       │
│ assessed 14:01 · inputs changed 13:40 · next after 10 Nov 14:01          │
└──────────────────────────────────────────────────────────────────────────┘

5.5 Wireframe W5: the roster manifest (configuration version)

{
  "version": 11, "lead_policy": "bierce_first_v1",
  "selection_scope": "whole_entry_v1", "editor_publication": "panel_v1",
  "decision": { "rule": "weighted_overlap_v1", "threshold": 3, "solo": false },
  "seats": [
    { "slot": 1, "kind": "persona", "profile": "voltaire", "profile_version": 9,
      "model_config": "qwen3.5-9b-whole-voltaire-v8", "weight": 2,
      "weight_reason": "seat report 2026-11-20: 91% supported over 40 attempts" },
    { "slot": 2, "kind": "persona", "profile": "hitchens", "profile_version": 1,
      "model_config": "qwen3.5-9b-whole-hitchens-v1", "weight": 1 },
    { "slot": 3, "kind": "persona", "profile": "bierce", "profile_version": 1,
      "model_config": "qwen3.5-9b-whole-bierce-v1", "weight": 1 }
  ],
  "readers": { "enabled": false, "weight_cap": 1 },
  "refresh_window_seconds": 86400, "change_reason": "…", "created_by_actor_id": 7
}

6. Part 4, humans: not a gate (design reference; rewritten 4 November after the owner's clarification)

The owner's words, 4 November 2026: this is an autonomous system; human intervention "could not conceivably keep up with the pace of it once it is running"; "we're not building any sort of gate to approve or disapprove"; if a thumb goes on the wheel at all "it will be just a random entry"; the owner may "from time to time" come in "as a dictator" on an individual entry, and even that is time-dependent; "long, long, long term" other people may vote entries up or down as "a more limited layer". This matches the 29 October correction on #193 (routine curation never waits for human review) and replaces the human seat drafted earlier today, which is withdrawn.

Three human mechanisms, none an approval step, in the order they exist:

  1. Built, and the whole of the owner's role for now: occasional intervention on one entry. The hold (mine) stops automatic curation of one page; a manual version with a required reason (human_override or human_added, Curation transparency: equal votes, reproducible selection, profiles and version history #203) replaces what the agents chose; withdrawal removes an item; an operator rejection is a veto on a claim, not on curation. None of these is a vote and no decision rule counts them. Add one thing: time. An intervention carries an until (a date, or a number of refresh windows, OD-8); when it lapses the page returns to automatic curation at its next visit, and the receipt says "owner's choice until …". Without an expiry a hand-curated page silently leaves the living loop, which is the opposite of what the owner wants. Holds end the same way.
  2. Decided as design, not built: no human ballot inside the decision rule. No seat waits for a person; no run has an approver; the owner is not a seat. "Kept" marks (Curation agent MVP: what an editor does, what a curated entry contains, how the page looks, how the owner and the bots share one record, and how it stays current (under #193) #253 D8) are not migrated into recurring votes: a mark was never consent to vote at every future decision. The standing-preference seat drafted on 4 November is withdrawn.
  3. Later and bounded: reader votes as a limited layer (OD-9). When the owner selects it: signed-in readers vote an item up or down with an optional reason, on the assertion_votes shape that already exists for exemplars (value of −1 or 1 per actor, humans and bots split by actor kind, exemplars already ordered by net votes; Exemplar votes: assertion_votes.reason, cached assertion_scores, the vote control and endorse-on-duplicate (#181 build 3) #189's reason column and cached tallies are the natural extension). All readers' votes enter a decision as one aggregated input with a capped weight set in the roster version (default: at most one seat's weight). Net support can promote or demote among candidates that are already eligible and already supported by at least one lens; it can never veto, never publish an ineligible or unsupported item, never move the lead, never create a claim or an entity. Votes show on the receipt as counts with reasons, apart from the lenses' reasons. Human opposition is therefore negative support, bounded by the cap, never dissent in a seat's name and never a veto.

6.1 Wireframe W6, revised: the two receipt lines humans can add (to W4, when either applies)

│ owner's choice until 1 Dec 2026 — "…" — automatic curation resumes after │
│ readers: 12 up · 3 down (capped weight 1 of threshold 3) · reasons ▸     │

The owner's screen gains an until field on hold and override, and nothing else. No ballot screen is built.

7. Part 5, a second model and how models would be benchmarked (thought process; not built now)

Design reference, not acceptance for this issue. Its own issue when the owner selects it, with its own quality, latency and resource criteria (the W7 columns).

What exists. One artifact family, Qwen3.5 9B Q4_K_M, pinned by manifest and weights digest in every model config; the 4B tag installed and unused; a readiness benchmark harness and its acceptance criteria from #254 P7 (the 26 October record); the D1 design (same instruction, matched evidence, the model config as the only difference); scorecard v2; per-attempt wall, prompt and output tokens and charged milliseconds.

The one path to a second model, when the owner wants it. (1) Choose one challenger: the installed 4B costs nothing and is E5; the natural download is Qwen3.5 27B dense from #97's shortlist (17 GB, fits 64 GB with the stated 16–20 GB headroom; latency unknown; a cross-family challenger comes after, one at a time). (2) Download to the volume, pin manifest digest, weights digest, quantization, runtime version. (3) Readiness benchmark on the P7 fixtures under the current packet version; record memory, swap and seconds per call. (4) Register an immutable model config under the same instruction as the baseline. (5) D1-style matched comparison on the ten pages; score on all proposed choices. (6) Enter the league table.

The league table (W7) is the comparison report grouped by model config, per instruction, nothing new to store:

instruction voltaire-whole-entry-v8 · ten pages · scorecard v2 · matched cases only
model config               params  quant   supported  bad reasons  useful  abstain  s/call  prompt tok  peak mem
qwen3.5-9b-whole-v8         9.7B   Q4_K_M  24/27      1            7/10    2        8.9     2,650       ~7 GB
qwen3.5-4b-whole-v8         4B     Q4_K_M  11/18      4            2/10    6        4.1     2,650       ~3 GB
qwen3.5-27b-whole-v8        27B    Q4_K_M  —          —            —       —        —       —           —  (not run)
Figures are illustrative shapes, not results. Same instruction, same evidence; a row differs from another in the model only.

What the bench may and may not decide. It chooses the lightest model that meets the quality gate at an acceptable seconds-per-call; a better model does not change the policy or the dossiers, which are measured separately. A seat's model config is already per seat in the roster, so a later roster may give one persona the 27B and another the 9B, only after each seat's own matched comparison. Not in scope at any time: cloud inference, fine-tuning, a model per persona by default, or choosing a model by a general benchmark score.

8. What does not change

  • Layer 0 is never written by an agent: no definitions, quotations, claims (illustrates stays behind Gate every illustrates nomination until accepted before bots or the public can nominate (#181 decision 6, deferred) #190 and human claim review) or entities.
  • Exact citations, expiry (news 23 h), rights, meaning and eligibility are blockers no vote can pass.
  • The lead is set by the configured rule, never voted.
  • One physical slot, 8K context, 1,024 output tokens, 120-second call, 30 minutes a day; readers never wait; offline runtime defers.
  • Every run freezes its packet, instruction, dossier version, model config and roster version; a later edit never rewrites an old decision.
  • No cohort is rerun to replace a failed one; no window is extended.
  • Receipts show who chose what, with what weight, and the dissent.

9. Build order, in phases

10. Acceptance (the one finish line)

  • D0 manifests frozen before any call, cited in every report; the §3.1 rule recorded before D6's first call.
  • D6 report posted with the three measures separate, lead correctness and highlights by kind, on ten complete matched cases (or the INCONCLUSIVE record and the new batch); the decision posted before anything else ran.
  • Either the v8 report (pass, fail, or fail at the love gate, on the unchanged gate, with stop and backup evidence) or the NOT JUSTIFIED record.
  • E4 report with per-case fingerprints, matched or unmatched flags, per-arm counts and lead outcomes apart; E5 run or "not run because …".
  • D4 decision: the next publishing configuration; the admission recommendation for the Hitchens and Bierce seats, with the counts; D1's disposition.
  • One read-only audit from the owner's session (manifests before calls, cases matched by fingerprint, all proposed choices scored, decisions supported by the counts) posted satisfactory; then this issue closes in one of the four terminal states.

11. Owner decisions (defaults proposed; the owner can change any on this issue)

12. Not in this issue

Non-default weights (a later #300 phase); the reader-vote layer and intervention expiry (their own issue when selected); the model bench (its own issue when selected); live panel operation (#300 L4–L6; this issue only fixes the rule's general form); curator pages (#299); consensus on claims or entities (#190); wider enrollment beyond the ten pages; cloud inference; fine-tuning; prompt mutation; hosting (#252); and any human approval or rejection step inside the loop, which is never built.

Activity

  1. changed the title [-]Personas: from one curator to a weighted panel — the v8 experiment, the dossier pipeline, consensus with weights, human seats and a later model bench (under #193)[/-] [+]Personas: from one curator to a weighted panel — the v8 experiment, the dossier pipeline, consensus with weights, humans outside the gate and a later model bench (under #193)[/+] on Nov 4, 2026
  2. holden commented on Nov 4, 2026

    @holden
    MemberAuthor

    Revision 1, 4 November 2026 (owner's words in chat: "look at these suggestions … update the issue if needed")

    Applied after an independent read-only critique (B− as a design brief, C+ for implementation readiness) and the owner's clarification that humans are not a gate. The body now has: one finish line (§§9–10: this issue closes on the D6 report, the v8 disposition, the first persona measurement and one read-only audit; Parts 3–5 are design references for their own issues); the promotion rule frozen before D6, with floors and order (§3.1, OD-7) and a NOT JUSTIFIED terminal state; the persona measurement isolating the lens with every other factor pinned and lead outcomes reported apart (E4); the executable budget, stop and admission plan (§3.2); Part 4 rewritten as three human mechanisms, none an approval step: time-limited intervention per entry (OD-8), no human ballot in the rule, a later bounded reader-vote layer on the existing assertion_votes shape (OD-9). The human seat drafted this morning is withdrawn (OD-4). Phases A–C replace the flat list. Title changed to match. #197 is reopened for the Vonnegut and Le Guin dossiers (OD-1).

  3. changed the title [-]Personas: from one curator to a weighted panel — the v8 experiment, the dossier pipeline, consensus with weights, humans outside the gate and a later model bench (under #193)[/-] [+]Personas: from one curator to a weighted panel — the v8 experiment, the dossier pipeline, consensus with weights, human governance of autonomous curation and a later model bench (under #193)[/+] on Nov 4, 2026
  4. holden commented on Nov 4, 2026

    @holden
    MemberAuthor

    Revision 2, 4 November 2026 (owner's words in chat: "look at these suggestions, update if they're good")

    Applied after the re-audit (A− as a design brief, B+ for implementation readiness). Runbook clarifications only, no sub-issues: (1) promotion needs all ten complete matched cases; a full-page retry on a new snapshot enters the scored set as that page's case; fewer than ten after the one retry is INCONCLUSIVE, a batch disposition distinct from NOT JUSTIFIED (§3.1); (2) the promotion order is total: usefulness, bad reasons, supported rate, lead correctness, charged milliseconds from the ledger, prompt tokens, then the declared order A0, P, L, PL; A0 wins its ties; the bad-reason floor counts every origin, and a wrong rule statement is a disqualifying tooling defect (§3.1, OD-7); (3) the love gate requires love to be scored useful under scorecard v2, and a stop at the gate is FAIL AT THE GATE with its stop, report and bundle evidence named (§3.2 step 6); (4) re-templated comparison instructions get fresh immutable names, shas and model configs, and the Voltaire dossier version is resolved before the Phase C freeze (E4). The first D6 call waits for these and for #298's closure and #303's deployment.

  5. holden commented on Nov 4, 2026

    @holden
    MemberAuthor

    Pointer, 4 November 2026: Phase A runs as its own narrow implementation issue, #306 (the four policy arms offline, D6: prerequisites P1–P7, preparation R1–R6 with the dry run, the run, scoring within 20 hours, and the decision comment posted here under §3.1). Posted on the owner's words in chat ("Post it"). Phases B and C stay in this issue; dossiers are #197; the fixture-only consensus engine is #300 L1.

  6. changed the title [-]Personas: from one curator to a weighted panel — the v8 experiment, the dossier pipeline, consensus with weights, human governance of autonomous curation and a later model bench (under #193)[/-] [+]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)[/+] on Nov 4, 2026
  7. holden commented on Nov 4, 2026

    @holden
    MemberAuthor

    Revision 3, 4 November 2026 (owner's words in chat): the persona META issue, with its experiments as numbered stages

    This issue is the persona META issue. Its experiments run as dedicated implementation issues, one per stage, in this order: Stage 1, #307, the first persona measurement (Hitchens and Bierce against Voltaire and the generic control) on the policy live today, run first because the lens question does not depend on the policy question; Stage 2, #306, the four policy arms offline (D6); Stage 3, the conditional v8 publishing run, under this issue until a Stage 2 arm qualifies. E4 is rewritten for the current policy (no template edited; a re-check on the new policy precedes any seat if Stage 2 changes it), §9 is in stage order, and the title says META. The earlier "Phase A, B, C" labels are Stages 2, 3 and 1.

  8. holden commented on Nov 4, 2026

    @holden
    MemberAuthor

    Amendment, 4 November 2026, 16:10 UTC: Stage 2 runs today (owner's words in chat: "Yes, run Stage 2 today")

    Posted by the audit session ("MVP progress audit and timeline") on the owner's decision. It amends §3.2's one-batch-a-day timing and the sentence "Stage 1 (#307) and Stage 2 (#306) run on different days", for 4 November 2026 and after. Nothing else in this issue changes.

    1. A second comparison batch may start on the same day when all of the following hold at its start: at least 15 minutes of the day's 30-minute budget remain (the condition E5 already uses); the earlier batch has ended with its guard unloaded, gateway paused and service stopped; and the new snapshot's earliest shelf expiry lies beyond the new batch's deadline (start + 3 hours). The "under 5 minutes charged" condition applies to the first batch of a day only.
    2. Stage 2 (#305 Stage 2: the four policy arms offline (D6), runbook and execution (under #305) #306) may run on the same day as Stage 1 (#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), after Stage 1's batch has ended. Stage 1's scoring continues in parallel; its 20-hour clock and its evidence-expiry rule are unchanged.
    3. Unchanged: atomic cases; halt on budget exhaustion or an uncertain call; one full-page retry the next day as :2; INCONCLUSIVE below ten complete matched cases; the §3.1 floors and order; the editor NULL throughout; no publication; every live step on the owner's typed yes, one at a time, in the executor's chat.

    Numbers at the amendment. 4 November charged 485,318 ms (8.1 minutes) after Stage 1's batch s1-2026-11-04 (attempts 122–161, 292,978 ms, ten cases matched, ended 15:59:12 UTC, verified read-only by this session); 21.9 minutes remain. Stage 1's earliest expiry was love at 21:46:02 UTC; Stage 2's own dry run reads its snapshot's expiry, and its batch must start at least three hours before it.

    Executor. #306 was written for another session on another day. Today it runs in the Stage 1 executor session ("Stage 1 persona measurement"), which holds the runtime state, the proven launchd guard layout (R3b: run dir, stop script and log on the internal disk) and the owner's chat for the typed yeses. #306's P6 and P7 read accordingly (P7 is already met: the v6 and v7 guards were unloaded under #307 S4). Required before the first call, unchanged: P1–P5 verified with ids, R1 dry run, R2 manifest frozen and sha-logged with the §3.1 rule text, R3 guard rehearsed, R4–R6 in the run folder.

  9. holden commented on Nov 4, 2026

    @holden
    MemberAuthor

    Stage 1 report (#307): the first persona measurement, batch s1-2026-11-04, 4 November 2026

    What ran: one matched, non-publishing comparison batch.

    Records (sha256):

    • D0 manifest frozen-manifest.json: c4f1de95d82977a890560d6d55597efb2c1561f84ce52b74388efe1a7fd262ac (frozen 15:52:40 UTC, cutoff 121).
    • Batch record record/batch.json: 20b9d3aa2ec9….
    • Coverage coverage.json: 6117b8e9edca….
    • Judgments judgments.json: dd723185e9e79fe665d63abae6e3136d83f98b898d6a10cfaccc3823b937935a.
    • Step log: run/s1-2026-11-05/steps.txt.
    • Scorecard v2: 156ee531… (unchanged).

    How it was scored:

    • Every proposed choice of every arm was scored: 77 in all (27 leads, 50 highlights). Abstentions count as zero-choice outcomes.
    • Scoring began 16:03:50 UTC, under 5 minutes after the batch, well inside the 20-hour window.
    • Evidence was inspected before any shelf expiry. At 16:02:42 UTC all 40 packets were rebuilt read-only, and each hash equals its attempt's packet_hash. The first expiry was love at 21:46 UTC.
    • The scorer was blind. A fresh scorer saw each page's four outputs under codes shuffled per page. The mapping was kept apart and applied only after scoring.
    • Judgments keep only ids, digests and the scorer's own short bases. All 77 reasons are model-written; there are 0 rule statements.
    • The D4 margin was set in steps.txt before any score existed: 15 percentage points.

    Per page (W2 form; outcome per scorecard v2; s/call is wall time)

    batch s1-2026-11-04 · snapshot 2cd9dae3… · scorecard v2 156ee531… · editor disabled
    page       arm          attempt  matched proposed supported insuff wrong  bad  outcome              s/call
    nepotism   voltaire-v7      122  yes            0         0      0     0    0  incorrect-abstention    9.7
    nepotism   generic-v1       123  yes            4         4      0     0    0  useful                  8.3
    nepotism   hitchens-v1      124  yes            5         5      0     0    1  useful                  9.3
    nepotism   bierce-v1        125  yes            4         4      0     0    0  useful                  8.5
    greed      voltaire-v7      126  yes            0         0      0     0    0  incorrect-abstention    6.1
    greed      generic-v1       127  yes            0         0      0     0    0  incorrect-abstention    5.9
    greed      hitchens-v1      128  yes            0         0      0     0    0  incorrect-abstention    6.2
    greed      bierce-v1        129  yes            0         0      0     0    0  incorrect-abstention    6.2
    war        voltaire-v7      130  yes            5         5      0     0    1  useful                  9.7
    war        generic-v1       131  yes            5         5      0     0    1  useful                  9.6
    war        hitchens-v1      132  yes            5         5      0     0    1  useful                 10.1
    war        bierce-v1        133  yes            5         5      0     0    2  useful                  9.7
    love       voltaire-v7      134  yes            5         4      1     0    1  useful                  9.9
    love       generic-v1       135  yes            6         5      1     0    1  useful                  9.8
    love       hitchens-v1      136  yes            1         1      0     0    0  lead-only               6.5
    love       bierce-v1        137  yes            1         1      0     0    0  lead-only               6.5
    coward     voltaire-v7      138  yes            1         1      0     0    0  lead-only               6.3
    coward     generic-v1       139  yes            1         1      0     0    0  lead-only               6.1
    coward     hitchens-v1      140  yes            1         1      0     0    0  lead-only               6.3
    coward     bierce-v1        141  yes            3         3      0     0    0  useful                  8.1
    patience   voltaire-v7      142  yes            0         0      0     0    0  refused                 6.1
    patience   generic-v1       143  yes            0         0      0     0    0  refused                 6.1
    patience   hitchens-v1      144  yes            4         4      0     0    0  useful                  8.4
    patience   bierce-v1        145  yes            0         0      0     0    0  incorrect-abstention    5.8
    justice    voltaire-v7      146  yes            5         5      0     0    2  useful                  9.2
    justice    generic-v1       147  yes            4         4      0     0    0  useful                  8.1
    justice    hitchens-v1      148  yes            1         1      0     0    0  lead-only               6.4
    justice    bierce-v1        149  yes            1         1      0     0    0  lead-only               6.3
    truth      voltaire-v7      150  yes            0         0      0     0    0  incorrect-abstention    5.8
    truth      generic-v1       151  yes            4         4      0     0    1  useful                  8.1
    truth      hitchens-v1      152  yes            1         1      0     0    0  lead-only               6.3
    truth      bierce-v1        153  yes            1         1      0     0    0  lead-only               6.3
    hope       voltaire-v7      154  yes            1         1      0     0    0  lead-only               6.5
    hope       generic-v1       155  yes            1         1      0     0    0  lead-only               6.3
    hope       hitchens-v1      156  yes            1         1      0     0    0  lead-only               6.5
    hope       bierce-v1        157  yes            1         1      0     0    0  lead-only               6.5
    bestiality voltaire-v7      158  yes            0         0      0     0    0  incorrect-abstention    5.7
    bestiality generic-v1       159  yes            0         0      0     0    0  incorrect-abstention    5.4
    bestiality hitchens-v1      160  yes            0         0      0     0    0  incorrect-abstention    6.0
    bestiality bierce-v1        161  yes            0         0      0     0    0  incorrect-abstention    5.8
    
    
    totals     voltaire-v7           10/10         17  16 (94%)      1     0    4  3/10 useful          
    totals     generic-v1            10/10         25  24 (96%)      1     0    3  5/10 useful          
    totals     hitchens-v1           10/10         19 19 (100%)      0     0    2  3/10 useful          
    totals     bierce-v1             10/10         16 16 (100%)      0     0    2  3/10 useful          
    lead-only pages are not useful (scorer's rule as in #298 cycle 1)
    

    Per arm

    arm complete, matched proposed supported insufficient wrong bad reasons (model / rule) useful pages lead correct / none highlights by section charged ms prompt tokens
    V voltaire-v7 10/10 17 16 (94%) 1 0 4 (4 / 0): 3 unsupported, 1 misattributed; 1 unassessable (no reason) 3 5 / 5 quotes 6, definitions 2, examples 2, images 2 75,473 32,006
    G generic-v1 10/10 25 24 (96%) 1 0 3 (3 / 0): 2 unsupported, 1 misattributed 5 7 / 3 quotes 10, definitions 4, examples 3, images 1 74,370 30,816
    H hitchens-v1 10/10 19 19 (100%) 0 0 2 (2 / 0): 2 unsupported 3 8 / 2 quotes 6, definitions 2, examples 2, news 1 72,683 33,136
    B bierce-v1 10/10 16 16 (100%) 0 0 2 (2 / 0): 2 unsupported 3 7 / 3 quotes 6, examples 2, definitions 1 70,452 33,086

    Page outcomes over the 40 arm-pages: 14 useful, 13 lead-only, 11 incorrect abstentions, 2 refused (decision_inconsistent), 0 justified empty. Every lead given was correct: 27 of 27.

    What the measurement shows

    1. Every proposed choice is well supported; the problem is how little gets proposed.
      • 75 of 77 choices are supported. Both insufficient ones are a title-only artwork on love.
      • But only 14 of the 40 arm-pages are useful, and no arm reaches more than 5 of 10.
    2. Abstention comes from the shared policy, not from any lens.
      • All four arms abstained on greed (noun sense definitions c7–c9 offered) and on bestiality (eligible sense definitions c4–c7 offered). This is the v6/v7 regression, unchanged by any persona.
      • patience was refused as decision_inconsistent by V and G, and abstained by B; H made it useful.
    3. Most leads have nothing beside them. 13 arm-pages are lead-only: a supported Bierce lead with every optional section empty, although supported definitions and quotations were offered.
    4. Bad reasons appear in every arm.
      • The Bierce lead on war is misread in all four arms.
      • Others are creators named without support in the fields (justice, truth), a quotation called a headline, and depicted content claimed for title-only artworks.
      • The scorer notes that reasons are cut near 100 characters, often mid-word. The cap is kept in the runtime contract, and some reasons may be weaker only because they are truncated.
    5. The personas trade volume for precision.
      • H and B propose fewer choices (19 and 16) than G (25), and every one of theirs is supported.
      • G proposes the most supported choices (24) and has the most useful pages (5).
      • V proposes the fewest choices (17), has the most bad reasons (4), and also refused patience.

    Small-sample limits.

    • Ten pages give 16–25 choices per arm.
    • The supported rates differ by one or two choices: 94% against 100% is one insufficient choice.
    • Wilson 95% intervals overlap widely (V 73–99%, G 80–99%, H 83–100%, B 81–100%), so the supported rates do not separate the arms.
    • Bad-reason counts of 2–4 are too small to rank arms.
    • These are iterative-retest pages, already observed and tuned; none of this generalises beyond them.

    D4 recommendation (#307 §7, the rule and margin fixed before scoring)

    test H hitchens-v1 B bierce-v1
    supported rate not below V's (94.1%) by more than 15 points 100%: pass 100%: pass
    bad-reason count not above V's (4) 2: pass 2: pass
    choices differ from V's on ≥ 1 page with a grounded, supported choice nepotism, patience, truth: pass nepotism, coward, truth: pass
    D4 admit to a seat admit to a seat
    adds supported choices over G? no: 19 against G's 24; useful pages 3 against 5 no: 16 against 24; useful pages 3 against 5

    Recommendation:

    E5, the 4B model arm: not run, because it runs only on the owner's word and none was given.

    Nothing was published, admitted or deleted. The editor stays NULL, the gateway is paused, the service is stopped and the guard is unloaded.

  10. holden commented on Nov 4, 2026

    @holden
    MemberAuthor

    Stage 2 decision (#306, D6): NOT JUSTIFIED, batch d6-2026-11-04, 4 November 2026

    What ran: one offline, matched, non-publishing batch of the four policy arms on the ten pilot pages, on one artifact (qwen3.5:9b, config 8's generation) and one snapshot (2cd9dae3…):

    • A0, current v7, config 8;
    • P, v7 on the text cut;
    • L, the fixed Bierce lead, config 12;
    • PL, both.

    It ran 16:18:19–16:23:20 UTC, attempts #162–#201, halted nil, ten cases complete and matched, nothing published. Run evidence: #306.

    Records (sha256):

    • D0 manifest frozen-manifest.json: c58b7b795a5f785a41fe2797dff18c379896ac4bc970f1f7e3fc718862f63f05 (frozen 16:17:42 UTC, cutoff 161). It embeds the §3.1 rule verbatim, sha 8ddd79d5c3b2a76ad74dbfb66ff2a74f31504daba0eb2627ae6ebec441380a8e, before the first call.
    • Coverage coverage.json: e1c6057b0168….
    • Batch record: 26357923d532….
    • Judgments judgments.json: a7689872af609bdaf83976facb19acb7b105e9a36c15c774333334b81657db66.
    • Step log: run/d6-2026-11-04/steps.txt.
    • Scorecard v2: 156ee531…, unchanged.

    How it was scored:

    • Every proposed choice of every arm, 87 in all, judged against the arm's own candidate set.
    • The review ran 16:24:30–16:31 UTC, inside the 20-hour window.
    • The evidence was rebuilt read-only at 16:24:12 UTC, before any shelf expiry; every packet hash equals its attempt's.
    • Two fresh scorers worked blind: arms were shuffled to codes on each page, and the two cut sets were named X and Y at random. The mapping was applied only after scoring.
    • Rule statements (the fixed leads) and model-written reasons were judged and counted apart.
    • Inter-scorer check: A0's outputs are identical to Stage 1's V arm (same choices and reason hashes, same packets, temperature 0, seed 195). The two independent blind scorings agree on 17/17 choice judgments and 10/10 page outcomes.

    Per page

    batch d6-2026-11-04 · snapshot 2cd9dae3… · scorecard v2 156ee531… · editor disabled · bad = unsupported + misattributed reasons
    page       arm attempt  proposed supported insuff  bad  outcome              lead     s/call
    nepotism   A0      162         0         0      0    0  incorrect-abstention none        9.5
    nepotism   P       163         4         4      0    0  useful               correct     6.6
    nepotism   L       164         1         1      0    0  lead-only            correct     5.2
    nepotism   PL      165         3         3      0    0  useful               correct     5.2
    greed      A0      166         0         0      0    0  incorrect-abstention none        6.1
    greed      P       167         0         0      0    0  incorrect-abstention none        5.9
    greed      L       168         4         4      0    0  useful               correct     8.4
    greed      PL      169         4         4      0    0  useful               correct     8.3
    war        A0      170         5         5      0    1  useful               correct    10.0
    war        P       171         5         5      0    2  useful               correct     9.1
    war        L       172         1         1      0    0  lead-only            correct     5.0
    war        PL      173         1         1      0    0  lead-only            correct     5.3
    love       A0      174         5         4      1    1  useful               correct    10.4
    love       P       175         6         5      1    0  useful               correct    10.8
    love       L       176         1         1      0    0  lead-only            correct     5.2
    love       PL      177         3         3      0    0  useful               correct     6.6
    coward     A0      178         1         1      0    0  lead-only            correct     6.6
    coward     P       179         3         3      0    0  useful               correct     8.1
    coward     L       180         1         1      0    0  lead-only            correct     4.9
    coward     PL      181         3         3      0    0  useful               correct     6.5
    patience   A0      182         0         0      0    0  refused              none        6.5
    patience   P       183         3         3      0    0  useful               correct     6.6
    patience   L       184         1         1      0    0  lead-only            correct     4.7
    patience   PL      185         1         1      0    0  lead-only            correct     3.3
    justice    A0      186         5         5      0    2  useful               correct     9.6
    justice    P       187         3         3      0    0  useful               correct     7.3
    justice    L       188         1         1      0    0  lead-only            correct     4.7
    justice    PL      189         1         1      0    0  lead-only            correct     3.9
    truth      A0      190         0         0      0    0  incorrect-abstention none        6.1
    truth      P       191         0         0      0    0  incorrect-abstention none        5.3
    truth      L       192         1         1      0    0  lead-only            correct     4.9
    truth      PL      193         1         1      0    0  lead-only            correct     4.2
    hope       A0      194         1         1      0    0  lead-only            correct     5.8
    hope       P       195         5         4      1    0  useful               correct     9.0
    hope       L       196         1         1      0    0  lead-only            correct     4.9
    hope       PL      197         3         3      0    1  useful               correct     6.2
    bestiality A0      198         0         0      0    0  incorrect-abstention none        5.4
    bestiality P       199         0         0      0    0  incorrect-abstention none        4.6
    bestiality L       200         4         4      0    1  useful               correct     9.1
    bestiality PL      201         5         4      1    0  useful               correct     8.8
    lead-only pages are not useful
    

    Per arm (#306 §6 form)

    arm complete, matched proposed supported insufficient wrong bad reasons (model / rule) useful pages lead correct (of pages with a priority candidate) highlights by section charged ms prompt tokens
    A0 10/10 17 16 (94.1%) 1 0 4 (4 / 0) 3 5/8 quotes 6, examples 2, definitions 2, images 2 76,318 32,006
    P 10/10 29 27 (93.1%) 2 0 2 (2 / 0) 7 7/8 quotes 14, examples 6, definitions 2 74,056 26,940
    L 10/10 16 16 (100.0%) 0 0 1 (1 / 0) 2 8/8 quotes 4, examples 2 57,542 32,914
    PL 10/10 25 24 (96.0%) 1 0 1 (1 / 0) 6 8/8 quotes 11, examples 4 58,868 27,848

    L and PL each carry 8 rule statements (the fixed Bierce leads, c1 on the eight pages with a priority candidate). All 16 rule statements are grounded: each fixed lead is the applicable priority candidate, so there is no tooling defect.

    Floors (§3.1) and the order

    arm useful ≥ 7 supported ≥ 90% zero bad reasons ten complete, matched qualifies
    A0 3: no 94.1%: yes 4: no yes no
    P 7: yes 93.1%: yes 2: no yes no
    L 2: no 100%: yes 1: no yes no
    PL 6: no 96.0%: yes 1: no yes no

    No arm qualifies, so there is no total order to apply.

    Decision: NOT JUSTIFIED

    What the batch shows (for that next issue; figures only, small sample)

    1. The text cut (P) raises usefulness.
      • Useful pages go from 3 to 7 and highlights from 12 to 22, and the patience refusal and the nepotism abstention disappear.
      • But it keeps v7's abstentions on greed, truth and bestiality.
      • Its two bad reasons are both on war: the Bierce lead misread, and an instance claim's reason going beyond its fields.
    2. The fixed lead (L) removes abstention and lead error.
      • L selects on all ten pages, with every lead correct (8/8 rule leads, plus correct model leads on greed and bestiality).
      • L and PL both recognise bestiality's eligible WordNet sense definition, the regression v6 and v7 never fixed.
      • But once a lead is fixed it chooses almost nothing else: 8 of 10 pages are lead-only.
    3. Both together (PL) come nearest to qualifying.
      • 6 useful pages, 96.0% supported, 1 bad reason (a quotation's reason describing a different quote, on hope).
      • L and PL are the cheapest arms: 57,542 and 58,868 ms against A0's 76,318.
      • It misses the floors by one useful page and one reason.
    4. The war lead misreading. Every model-written war lead reason misreads the Bierce entry, in Stage 1 and here. The rule statement removes that error class in L and PL.
    5. Reasons are cut near 100 characters, often mid-word (the runtime contract's cap, unchanged in Curation: matched comparison support, policy arms and stop guard (#298 Part C, C1–C9) #303). Both scorers note it, and some unsupported verdicts may come from truncation.

    Small-sample limits.

    • Ten iterative-retest pages, already observed and tuned; 16–29 choices per arm.
    • Wilson 95% intervals for the supported rate overlap widely: A0 73–99%, P 78–98%, L 81–100%, PL 80–99%.
    • The floors are discrete counts, so one page or one reason decides PL's verdict.
    • This justifies another bounded experiment, not autonomy.

    Nothing was published, admitted or deleted. The editor stays NULL, the gateway is paused, the service is stopped, and the guard is unloaded. Today's charge is 752,102 ms of 1,800,000.

  11. added a commit that references this issue on Nov 4, 2026
  12. holden commented on Nov 5, 2026

    @holden
    MemberAuthor

    D7 decision (#311): NOT JUSTIFIED, batch d7-2026-11-05, 5 November 2026

    What ran: one offline, matched, non-publishing batch of four arms on the ten pilot pages. All four used one artifact (qwen3.5:9b, config 8's generation) and one snapshot (83e0f1ac…). The arms:

    • A0: current v7, config 8.
    • PL: the text cut with the fixed Bierce lead, config 12. This is D6's nearest arm, run again.
    • PLC: PL with the reason cap raised from 100 to 200, config 13 (selection-output-v5).
    • PLCN: PLC plus one paragraph against lead-only pages, config 14 (selection-output-v5).

    It ran 02:19:09–02:23:56 UTC as attempts #202–#241: halted nil, ten cases complete and matched, nothing published. The run evidence is on #311.

    Records (sha256):

    • Frozen manifest frozen-manifest.json: 3dd8ed8920a0b54b3599eee25d81a73be2fda2b8973569dc2f61bc8ca07c5c23, frozen 02:18:11 UTC with cutoff 201. It embeds the §3.1 rule verbatim, sha 8ddd79d5c3b2a76ad74dbfb66ff2a74f31504daba0eb2627ae6ebec441380a8e, before the first call.
    • Coverage coverage.json: 3f636289f93b….
    • Batch record: d74e14dc904a….
    • Judgments judgments.json: 18313707e6e9a1732f4069b76308e2cad2c2b0c48ee8b7293a51a867dfed47fa.
    • Step log: run/d7-2026-11-05/steps.txt.
    • Scorecard v2: 156ee531…, unchanged.

    How it was scored:

    • All 95 proposed choices, across every arm, were judged against each arm's own candidate set.
    • The review ran 02:25:30–02:29:27 UTC, inside the 20-hour window.
    • The evidence was rebuilt read-only at 02:24:56 UTC, before any shelf expiry. Every packet hash equals its attempt's.
    • Two fresh scorers worked blind. On each page the arms were shuffled to codes and the two cut sets were named X and Y at random. The mapping was applied only after scoring.
    • Rule statements (the fixed leads) and model-written reasons were judged and counted apart.
    • Inter-scorer check:
      • PL's prompt sha and evidence fingerprint equal D6's PL on all ten pages, and its outputs are identical (temperature 0, seed 195). So the repeat measures the scorers, not the model.
      • The D7 scorers agree with D6's on 25/25 meaning judgments, 24/25 reason judgments and 10/10 page outcomes. The one difference is hope c18, judged unsupported in D6 and misattributed here, so it counts as a bad reason both times.

    Per page

    batch d7-2026-11-05 · snapshot 83e0f1ac… · scorecard v2 156ee531… · editor disabled · bad = unsupported + misattributed reasons
    page       arm  attempt  proposed supported insuff  bad  outcome              lead     s/call
    nepotism   A0       202         0         0      0    0  incorrect-abstention none      11.5
    nepotism   PL       203         3         3      0    0  useful               correct    5.1
    nepotism   PLC      204         3         3      0    0  useful               correct    4.4
    nepotism   PLCN     205         3         3      0    0  useful               correct    4.6
    greed      A0       206         0         0      0    0  incorrect-abstention none       6.2
    greed      PL       207         4         4      0    0  useful               correct    7.5
    greed      PLC      208         4         4      0    0  useful               correct    7.8
    greed      PLCN     209         4         4      0    0  useful               correct    7.9
    war        A0       210         5         5      0    1  useful               correct    9.8
    war        PL       211         1         1      0    0  lead-only            correct    4.9
    war        PLC      212         1         1      0    0  lead-only            correct    4.8
    war        PLCN     213         3         3      0    0  useful               correct    6.5
    love       A0       214         1         1      0    0  lead-only            correct    6.3
    love       PL       215         3         3      0    0  useful               correct    6.2
    love       PLC      216         3         3      0    0  useful               correct    6.3
    love       PLCN     217         5         4      1    1  useful               correct    7.9
    coward     A0       218         1         1      0    0  lead-only            correct    6.3
    coward     PL       219         3         3      0    0  useful               correct    6.2
    coward     PLC      220         3         3      0    0  useful               correct    6.4
    coward     PLCN     221         3         3      0    0  useful               correct    6.4
    patience   A0       222         0         0      0    0  incorrect-abstention none       5.5
    patience   PL       223         1         1      0    0  lead-only            correct    3.1
    patience   PLC      224         1         1      0    0  lead-only            correct    3.1
    patience   PLCN     225         1         1      0    0  lead-only            correct    3.3
    justice    A0       226         6         5      1    1  useful               correct    9.8
    justice    PL       227         1         1      0    0  lead-only            correct    3.7
    justice    PLC      228         1         1      0    0  lead-only            correct    3.7
    justice    PLCN     229         1         1      0    0  lead-only            correct    3.8
    truth      A0       230         0         0      0    0  incorrect-abstention none       5.7
    truth      PL       231         1         1      0    0  lead-only            correct    3.8
    truth      PLC      232         1         1      0    0  lead-only            correct    4.1
    truth      PLCN     233         3         3      0    1  useful               correct    5.8
    hope       A0       234         1         1      0    0  lead-only            correct    6.5
    hope       PL       235         3         3      0    1  useful               correct    5.9
    hope       PLC      236         1         1      0    0  lead-only            correct    4.4
    hope       PLCN     237         5         4      1    1  useful               correct    7.9
    bestiality A0       238         0         0      0    0  incorrect-abstention none       5.8
    bestiality PL       239         5         4      1    0  useful               correct    7.4
    bestiality PLC      240         5         4      1    0  useful               correct    7.8
    bestiality PLCN     241         5         4      1    0  useful               correct    7.8
    lead-only pages are not useful
    

    Per arm (#306 §6 form)

    arm complete, matched proposed supported insufficient wrong bad reasons (model / rule) useful pages lead correct (of pages with a priority candidate) highlights by section charged ms prompt tokens
    A0 10/10 14 13 (92.9%) 1 0 2 (2 / 0) 2 5/8 quotes 4, examples 2, definitions 1, images 1, news 1 73,798 31,911
    PL 10/10 25 24 (96.0%) 1 0 1 (1 / 0) 6 8/8 quotes 11, examples 4 54,233 27,848
    PLC 10/10 23 22 (95.7%) 1 0 0 5 8/8 quotes 9, examples 4 53,450 27,848
    PLCN 10/10 33 30 (90.9%) 3 0 3 (3 / 0) 8 8/8 quotes 16, examples 7 62,400 29,078

    PL, PLC and PLCN each carry 8 rule statements: the fixed Bierce lead c1, on the eight pages with a priority candidate. All 24 are grounded. No answer was refused, so the decision_inconsistent confound recorded for PLCN did not arise.

    Floors (§3.1) and the order

    arm useful ≥ 7 supported ≥ 90% zero bad reasons ten complete, matched qualifies
    A0 2: no 92.9%: yes 2: no yes no
    PL 6: no 96.0%: yes 1: no yes no
    PLC 5: no 95.7%: yes 0: yes yes no
    PLCN 8: yes 90.9%: yes 3: no yes no

    No arm qualifies, so there is no total order to apply.

    Decision: NOT JUSTIFIED

    What the batch shows (figures only, small sample)

    1. PL is deterministic.
      • Same prompts, same outputs and the same score as D6: 6 useful pages, 96.0% supported, 1 bad reason (hope c18).
      • Its text-cut evidence did not change overnight. A0's current-cut prompt changed on 8 of 10 pages; PL's changed on none.
      • Re-running an arm on the same text-cut evidence adds no information. Only a change of evidence, instruction or model can.
    2. The cap alone (PLC) chose PL's ids on 9 of 10 pages.
      • On hope it kept only the lead, dropping PL's two hope highlights, including the misattributed c18. That gives zero bad reasons, but one useful page fewer (5).
      • Model-written reasons grew, but not to the cap: the median is 76 characters against PL's 55. Only 1 of 15 is over 100 characters, the longest is 112, and none is near 200.
      • The 100 cap did not bind PL here, whose longest model reason is 76. Only A0's reasons reach it: 3 of 13 are at 95–100 characters.
    3. The nudge (PLCN) is the first arm to clear the useful floor, with 8 pages: against PLC, it turns war, truth and hope from lead-only into useful. It has 23 highlights against PLC's 13, at 90.9% supported, just above the floor. Its 25 model reasons have a median of 81 characters, and 5 are over 100 (the longest is 114). It fails only on bad reasons, 3 of them:
      • love c13 and hope c13: claim records (source wikipedia) placed under examples. Each labels the headword as a lexeme, and its reason asserts more than the label shows. Both were judged insufficient in meaning.
      • truth c3: a quotation whose reason names a poet the candidate does not supply.
    4. Lexeme claim records (kind claim, under examples, labelling the headword only) are insufficient on every arm that chose them: bestiality c13 in PL, PLC and PLCN, and love c13 and hope c13 in PLCN. They carry 2 of PLCN's 3 bad reasons.
    5. A0 on the day's snapshot:
      • 2 useful pages and 5 incorrect abstentions (nepotism, greed, patience, truth, bestiality).
      • Its prompt equals D6's only on coward and bestiality.
      • The v7 baseline stays far from the floors, as in D6 (3 useful, 4 bad).

    Small-sample limits:

    • Ten iterative-retest pages, already observed and tuned; 14 to 33 choices per arm.
    • The Wilson 95% intervals for the supported rate overlap widely: A0 69–99%, PL 80–99%, PLC 79–99%, PLCN 76–97%.
    • The floors are discrete counts, so one page or one reason decides each verdict.
    • This justifies at most another bounded experiment, not autonomy.

    Nothing was published, admitted or deleted. The editor stays NULL, the gateway is paused, the service is stopped and the guard is unloaded. Today's charge is 243,881 ms of 1,800,000. #306, #307 and #311 stay open for the owner.

  13. holden commented on Nov 5, 2026

    @holden
    MemberAuthor

    D8 is #316 (5 November 2026, prepared on the owner's words in the audit session's chat: "Yes, prepare D8"). After the D7 decision above: a 2 × 2 of forty non-publishing calls on the ten pilot pages, measuring apart (a) a text cut whole-entry-text-cut-v2 that no longer offers claim candidates bound to the lexeme rather than a sense (the lexeme_entity_candidate path; exactly three such candidates were offered across D7's ten PLCN packets, the three judged insufficient) and (b) one attribution paragraph in the nudge instruction. Floors and order unchanged, scored blind from scratch, benefit stated as bounded, the attribution rule counted on its own. Executor: the Stage 1/2/D7 session; every live step on the owner's typed yes there.

  14. added a commit that references this issue on Nov 6, 2026
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