Skip to content

#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

Closed

Description

@holden

Implementation issue, 4 November 2026, Stage 1 of the persona META issue #305, under #193. Posted on the owner's words in chat, 4 November 2026 ("this is the implementation that we're gonna do first"). Narrow by design: one matched, non-publishing batch of forty calls on the ten pilot pages, four arms that differ in the editorial lens alone, on the policy that is live today (v7's evidence policy, under which #302's Hitchens and Bierce instructions were written), scored within 20 hours, ending in the admission recommendation for the two persona seats of the first roster (LP-1: Voltaire, Hitchens, Bierce). One model, pinned below. No v1 template is edited, but one small pin PR comes first (P3b): #303 carries only the persona-by-slug seam, so the Hitchens and Bierce instructions must be pinned in Packet and named in WholeEntry.comparison_instructions/0 before an arm can run. No stage waits on another. The policy experiment is Stage 2 (#306), on another day; the conditional v8 publishing run is Stage 3. No publication; the editor disabled throughout. No live step follows from this text.

1. Outcome

One report and one recommendation on #305: whether the Hitchens and Bierce lenses, on identical evidence and under the same policy as the Voltaire baseline, choose differently and at least as well (supported meaning, bad reasons, usefulness, each reported apart), and whether either should be admitted to a seat. Secondary: whether the personality adds supported choices over the generic control. This issue closes on that comment.

2. The model, pinned (the only model this issue uses)

Field Value (verified read-only on v2, 4 November 2026)
Model qwen3.5:9b, 9.7B parameters, Q4_K_M
Runtime Ollama 0.34.4 on the volume, private service 127.0.0.1:11435, one slot
Artifact manifest digest 6488c96fa5fa… (full value in every model config)
Weights digest dec52a44569a…
Generation settings {"seed": 195, "temperature": 0, "think": false, "num_ctx": 8192, "num_predict": 1024, "keep_alive": "10m"}
Baseline model config qwen3.5-9b-whole-voltaire-v7 (config 8, hash 81a12ded2792…), the v7 trial's own configuration
Optional model arm (E5) qwen3.5:4b, installed, no model config yet; registered under the baseline instruction only if E5 runs, its own manifest and weights digests recorded then

No other model is downloaded, registered or run in this issue.

3. The four arms (one lens each, everything else identical)

Arm Instruction (used as it is; a v1 name is never edited) Dossier Model config
V, baseline voltaire-whole-entry-v7 (sha a3219a5c…) Voltaire, profile version 7 (dossier file v8, manifest 7db26817…) config 8, exists
G, control generic-whole-entry-v1 (#303 C3: v7's evidence policy and section rules without a dossier) none qwen3.5-9b-whole-generic-v1, to register
H hitchens-whole-entry-v1 (#302: v7's policy plus the Hitchens lens) Hitchens v1, priv/curation/agents/personas/hitchens/v1/profile.json qwen3.5-9b-whole-hitchens-v1, to register
B bierce-whole-entry-v1 (#302: v7's policy plus the Bierce lens) Bierce v1, priv/curation/agents/personas/bierce/v1/profile.json qwen3.5-9b-whole-bierce-v1, to register

Pinned across all four and recorded in the manifest: v7's evidence policy and packet cut (24 candidates, 12 meanings); model-selected leads for every arm under the same lead rule; runtime-packet-v4 and selection-output-v4; the artifact and generation settings above. Within the batch every arm of a page must show the same evidence fingerprint, which the dry run proves.

Why on the current policy, and first. The lens question (does a different lens choose differently and at least as well on the same packet) is independent of the packet-policy question Stage 2 asks. #302's dossiers were written for v7's policy and #303's seam composes v7's text with a lens (Packet.grounded_with_lens/1), so no re-templating is needed; only the pin PR (P3b) precedes the batch. 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); this issue's record stands as the first lens measurement either way.

Why Bierce now: it holds a seat in the owner's first roster (LP-1, #300), and #302's bierce-v1.md states its hypothesis (satirical contrast chooses differently from evidential support on the same packet). The lead is not voted and is model-selected under the same rule in every arm, so the Bierce lens is not confounded by Bierce-first leads; lead outcomes are reported apart from highlights regardless.

4. Prerequisites, verified with ids before the first call

5. Preparation, no model calls (can start once P2 and P3 hold)

  • R1. Dry run on v2. mix dd.runtime compare --batch s1-<date> --pilot --dry-run --arms V=qwen3.5-9b-whole-voltaire-v7,G=qwen3.5-9b-whole-generic-v1,H=qwen3.5-9b-whole-hitchens-v1,B=qwen3.5-9b-whole-bierce-v1. Expected: ten cases, one evidence fingerprint shared by all four arms per page, four distinct packet hashes per page (the instruction version is in the packet JSON), omissions by reason, earliest shelf expiry per case, and no attempt row written (max attempt id unchanged, verified read-only). Pass --timeout-ms 300000 if the read-only transaction overruns.
  • R2. The D0 manifest run/s1-<date>/frozen-manifest.json, written and sha-logged before the first call: code SHA (serving = main), batch id, the four arm identities (instruction names and shas, dossier versions with manifest hashes, model config ids and hashes, generation settings), the ten pages with composition ids, the case scheme s1-<date>:SLUG:n, the fingerprint rule (C1), scorecard v2 sha 156ee53110b7c8ecad98291a8087c91b831e2a07c1c47f19aca4e269ee7cd1a9, the deadline (start + 3 hours) and the attempt cutoff (max attempt id at freeze).
  • R3. The guard, rehearsed: ops/curation/trial-stop/trial-stop.sh --deadline <epoch> --run-dir run/s1-<date> --dry-run writes DRY_RUN_OK; the real load with the deadline is a live step on the owner's yes immediately before the run.
  • R4. Scoring record template (the Curation agents: close out v6, retest v7, then controlled model and personality comparisons on identical packets (under #193) #298 B3 shape): per case, arm and proposed choice: candidate id, meaning id, kind, section, meaning judgment (supported, insufficient, wrong), reason judgment (grounded, unsupported, misattributed), reason origin; per page and arm the outcome (useful, lead-only, abstained, incorrect abstention, refused) and the lead outcome kept apart; the review clock; the file's sha when frozen. No provider text.
  • R5. RUNBOOK.md in the run folder with 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 §3.2's rules: start under 5 minutes charged; atomic cases; halt on budget exhaustion or an uncertain call with remaining pages not_run; expiry inside a case makes it unmatched; one full-page retry the next day as :2; INCONCLUSIVE below ten complete matched cases.
  • R6. A step log (steps.txt), every step timestamped in UTC with exit codes and the owner's words for each live step.

6. The run (on the owner's yes: "run batch s1-")

  1. P1–P5 re-checked and logged; R1 repeated the same day; R2 frozen; guard loaded (owner's yes); editor NULL confirmed.
  2. mix dd.runtime compare --batch s1-<date> --pilot --arms … (the R1 arms, without --dry-run). Forty calls, about ten minutes at the observed 6–14 seconds each; each case's four arms back to back on the one slot and the shared budget.
  3. Verify read-only: forty purpose = comparison attempts (or the recorded halt), each with arm, comparison_case, comparison_snapshot, evidence_fingerprint; every case complete and matched; no composition version, proposal or publication written; gateway available; budget charged recorded from the ledger.
  4. mix dd.curation.coverage --comparison s1-<date> --record run/s1-<date>/coverage.json (read-only; every case listed); sha logged.
  5. Guard unloaded (owner's yes); final state recorded: editor NULL, last attempt id, charged budget (D5).
  6. If a case is incomplete or unmatched: the full page-and-arm set once more the next day as s1-<date>:SLUG:2 on a new snapshot; never one arm alone; never combined across snapshots.
  7. E5, optional, a separate smaller batch on another day: register qwen3.5-4b-whole-voltaire-v7 on the owner's yes; twenty calls (ten 4B plus ten matched baseline); same scoring; its result goes to Personas META: from one curator to a weighted panel — the experiments in stages, the dossier pipeline, consensus with weights, humans outside the gate and a later model bench (under #193) #305 Part 5's record, never to a seat decision.

7. Scoring and the report

Scorecard v2, unchanged: every proposed choice of every arm, within 20 hours; abstentions as zero-choice outcomes; model-written reasons and rule statements apart; evidence inspected before its expiry; only ids, digests and short bases retained; the judgments file frozen with its sha.

The report, one comment on #305: per arm, cases complete and matched (of 10), proposed choices, supported, insufficient, wrong, bad reasons by origin, useful pages, lead outcomes (apart), highlights by kind, charged milliseconds, prompt tokens; paired per page; unmatched pages reported, not dropped; small-sample limits stated. Then the D4 recommendation: for each of H and B, admit to a seat (its supported rate not below V's by more than the stated small-sample margin, its bad-reason count not above V's, its choices differing from V's on at least one page with a grounded reason) or not, with the counts; and whether the personality adds supported choices over G. None of this authorises publication by itself; a seat is #300's step, after #298's quality gate has been passed by a curator.

8. Acceptance (closes this issue)

9. Live steps, each on the owner's named yes in chat, one at a time

Register three model configs (four with E5); unload the v6 and v7 guards; load and unload the C6 guard; run the batch, the retry and E5. Nothing published, nothing admitted into the publication guards, nothing deleted.

10. Not in this issue

Stage 2, the policy arms offline (#306); Stage 3, the conditional v8 publishing run (#305); seats, rosters and the live panel (#300); the Vonnegut and Le Guin dossiers (#197); any model beyond the installed 4B (#305 Part 5's own issue).

Activity

  1. holden commented on Nov 4, 2026

    @holden
    MemberAuthor

    Amendment, 4 November 2026: P3b, the pin PR, added to §4 and the top note. #303's C4 is the persona-by-slug seam only (its PR body: "when a dossier lands, its instruction is pinned in Packet and named once in WholeEntry.comparison_instructions/0"), and #302 carries the dossier files, so the Hitchens and Bierce instructions need one small reviewed PR before an arm can run: both pinned as grounded_with_lens of each dossier's lens, both named in the comparison list, public summaries added, no migration, publication guards untouched. A dedicated executor session for this issue is to be opened by the owner from the desktop app with the prepared brief.

  2. holden commented on Nov 4, 2026

    @holden
    MemberAuthor

    S1, 4 November 2026: exact-head confirmations and merges (#303, #302)

    Executor of this issue; ids and hashes only, no provider text.

    Step When (UTC) Evidence
    #303 exact-head review of 3449553 (range 44d25dd..3449553, N1–N3) 04:30 c6622597532: fresh subagent, partition _s1a, 42 focused + 3 runbook tests, 0 failures; nothing REQUIRED; 2 RECOMMENDED (unescaped --out glob; record search limited to --out), left as follow-ups, neither touches this issue
    #302 exact-head review of 2919d1a (range 00c8979..2919d1a, RECOMMENDED 1–3) 04:29 c6622581655: fresh subagent, partition _s1b, 14 tests, 0 failures; nothing REQUIRED; 2 RECOMMENDED (test-only), carried into the pin PR
    Merge #303 04:36:40 owner's words in chat: "Yes, merge #303". gh pr merge 303 --merge --match-head-commit 3449553…, merge commit 2a041267aa55241cd6c9065a05788747c8d7850e (parents d33a655, 3449553)
    Merge #302 04:37:48 owner's words in chat: "Yes, merge #302". gh pr merge 302 --merge --match-head-commit 2919d1a…, merge commit f63775a59e4bdce00b269eb3737e7429061f03c2 (parents 2a04126, 2919d1a)

    Main is f63775a, tree e0b570f: the main+303+302 tree tested whole on 4 November (00:21–00:29 UTC, 3,400 tests, merge assessment c6618975256). Nothing is deployed: the serving checkout is still d33a655, no migration applied, #298 is open. P2 and P3 stay unticked until the S3 deploy.

    Next: the pin PR (P3b) rebased onto f63775a, full precommit and an independent exact-head review; the deploy (S3) only after #298 is closed, on the owner's yes.

  3. holden commented on Nov 4, 2026

    @holden
    MemberAuthor

    S2, 4 November 2026: the pin PR (P3b) is up, PR #308

    Executor of this issue; ids and hashes only.

    • PR #308, branch claude/307-persona-pins, head e011ca6f21862940e3a24028ae53e041641c52ea on main f63775a. MERGEABLE, CLEAN. 7 files.
    • Pinned in Runtime.Packet: each instruction is v7's grounded body and lens preface, followed by that dossier's lens. The dossier is read through Agents.Persona and compiled in, so the policy sections are v7's byte for byte.
      • hitchens-whole-entry-v1: sha256 3463bce9960a3919fbe8d9ef79aea83a06e8cfc3d15800338ee1cfdf73abe9ab
      • bierce-whole-entry-v1: sha256 b02355a9a294f4f4c5deca7d07325ae705fda6dec045ab2b64e7ac92efcb57f0
      • voltaire-whole-entry-v7 stays at a3219a5c… and generic-whole-entry-v1 at 0423863d….
    • Allowlists: both are named in WholeEntry.comparison_instructions/0, so the section output, the contract (selection-output-v4) and the evidence guidance admit them. publishing_instructions/0 is unchanged.
    • Public summaries under priv/curation/agents/instructions/: no run of eight consecutive words of either instruction appears in its summary.
    • Not in the diff: no migration, no config/*.exs and no mix.lock change. The configuration CHECK, the author guard, coherent?/1, Persona.prepare/3 and admission are untouched.
    • Tests: full mix precommit at e011ca6, partition _s1f, 05:03:45–05:11:53 UTC: 19 doctests, 3,402 tests, 0 failures. The earlier runs, all explained, are listed in the PR body:
      • one genuine failure, the literal pinned list in packet_test.exs, fixed in f90494d;
      • a scratch-database leftover from a run I had stopped;
      • a latency-timed scorecard row that flipped under load.
    • Independent exact-head reviews: c6623158573 at f90494d, nothing REQUIRED. Its RECOMMENDED 1 (summary wording) is applied as e011ca6; 2 and 3 predate this PR and are follow-ups. c6623158909 at e011ca6, nothing REQUIRED or RECOMMENDED.
    • CodeRabbit CLI: not run. It reports that its Git provider access needs reauthorization (sign out and in again).

    Two corrections to this issue's text, for the runbook:

    1. mix dd.runtime compare has no --timeout-ms switch (R1).
    2. mix dd.curation.coverage --record is the input, the batch record compare wrote, not an output path (§6 step 4). The working form is mix dd.curation.coverage --comparison s1-<date> --record run/s1-<date>/batch.json > run/s1-<date>/coverage.json.

    The merge needs the owner's yes. The deploy (S3) comes after #298 is closed.

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

    @holden
    MemberAuthor

    S2, 4 November 2026: the pin PR merged

  6. holden commented on Nov 4, 2026

    @holden
    MemberAuthor

    S3: comparison attempts deployed (migration and fast-forward), 4 November 2026 (#307 executor)

    Owner's yes, in chat on 4 November 2026: "Yes to S3, S4 and R3". It covers BUNDLE.md (sha e514c638) S3, S4 and R3, run in order. Any failure halts the bundle; the only planned exceptions are R3a failing into R3b, and R3b into R3c. The run folder is /Volumes/LLM Models/dictionary/run/s1-2026-11-05/, and the step log is steps.txt there.

    Precondition: #298 CLOSED (completed) at 14:25:51 UTC: closure, B6, audit verdict SATISFACTORY.

    P, preflight (read-only, 15:37:31 UTC): all 17 checks PASS (preflight.log).

    K1, migration (15:37:46–47 UTC):

    • Run from worktree 307-deploy at 4a0b997 with MIX_ENV=dev DD_NO_OBAN=1 mix run --no-start run/s1-2026-11-05/migrate.exs (sha cccac920).
    • The script refuses unless all of these hold: the target is v2 on 5434 in this cluster, the editor is disabled, and the only pending migration is 20261103202412.
    • Applied [20261103202412]. schema_migrations is now 42, newest 20261103202412.
    • 0 comparison rows, 0 fingerprints, max attempt 121.

    C, old-code compatibility (read-only, before the fast-forward): all 11 checks PASS (compat.log).

    • d33a655 still served on the new schema: / 200, /health 302, pid 1661.
    • The four new columns (arm, comparison_case, comparison_snapshot, evidence_fingerprint) are nullable, and no row uses them.
    • inference_attempts_shape and inference_attempts_comparison_shape are both validated.
    • The identity trigger is present, and comparison_attempt_guard is on curation_item_bases and editorial_composition_versions.
    • The d33a655 Attempt schema names none of the new columns and has no :comparison purpose.

    K2, fast-forward (15:38:04 UTC):

    V, verify (read-only, 15:38:09 UTC): all 7 checks PASS (postdeploy.log).

    • HEAD 4a0b997, tree clean.
    • / 200, /health 302.
    • No restart: 4007 is still pid 1661, and it rewrote 14 modules in its own build (_build/published) at 15:38:08 UTC, including Runtime.Attempt and Runtime.Compare.
    • 42 migrations, editor NULL, max attempt 121.

    Not touched: 4007 was not restarted; no /healthz, /on/* or /define/*; nothing seeded, admitted or bound; no model call; nothing deleted.

    Record shas (first 12): preflight.log 8d38f0125b1d, migration.log e2d699d45616, compat.log da87a958ae96, fetch.log 669a962bfbb3, deploy.log 965596f76069, postdeploy.log ff3ae46b1859.

    Next, under the same yes:

    • S4: register the three model configs, put the service down again, boot out the v6 and v7 guards.
    • R3: the launchd guard rehearsal.
  7. holden commented on Nov 4, 2026

    @holden
    MemberAuthor

    S4 and R3: three model configs, old guards out, guard proven under launchd, 4 November 2026 (#307 executor)

    These ran under the same yes as S3: the owner's "Yes to S3, S4 and R3" in chat, covering BUNDLE.md (e514c638). The step log is run/s1-2026-11-05/steps.txt.

    S4: model configs and the old guards

    K3, private service up for setup only (15:39:46 UTC): runtime.exs service start launched ollama serve as pid 16154 on 127.0.0.1:11435, runtime 0.34.4, authority devils_dictionary_v2.

    K4–K6, setup (15:40:04–05 UTC): runtime.exs setup --model qwen3.5:9b --slug … --instruction … --accept-license, recorded by actor 18. No --pull.

    id slug instruction config_hash
    9 qwen3.5-9b-whole-generic-v1 generic-whole-entry-v1 3acaeb4e44e452184f6f61d8c6e1a98bc4e2de6cabbf17008101df6d648843a5
    10 qwen3.5-9b-whole-hitchens-v1 hitchens-whole-entry-v1 baa0ce4fe68482965f15c0c96ba5a0840455ae1dfc315ed8ace61301678526c9
    11 qwen3.5-9b-whole-bierce-v1 bierce-whole-entry-v1 e0420dbe6d791dd9ebac956ed9de9299ecf1fae554f41b403642796b700d8666

    K6c, check (read-only): all 3 PASS.

    • Each config has its own instruction.
    • Model, manifest, weights, template, quantization, runtime version, licence and generation all equal the V arm's config 8 (qwen3.5-9b-whole-voltaire-v7, 81a12ded…).
    • There are 11 configs.

    K6s, service down (15:40:20 UTC): stopped pid 16154. runners: []: no model was loaded, so nothing was inferred. 11435 is closed, and the gateway stays paused until S6.

    No attempt and no charge: max attempt is still 121. The 4 Nov charge is still 192,340 ms over 25 entries (the last at 11:22:44), against S6's 5-minute rule.

    K8, old guards out (15:40:46 UTC):

    • launchctl bootout of io.wordhoard.v6-trial-stop (5,085 runs, last exit 0) and io.wordhoard.v7-trial-stop (1,460 runs, last exit 0): rc 0 each.
    • launchctl list now shows no wordhoard job.
    • Their plists stay in ~/Library/LaunchAgents, so they would load again at the next login. That is harmless, since both deadlines have passed and both trials have STOPPED. Deleting the plists is a separate yes.

    R3: the guard under launchd (#298 E6, c6628851056)

    I, install: trial-stop.sh was installed into ~/Library/Application Support/DictionaryTrial/ ($GUARD). Its sha 65ad002f… equals the copy at 4a0b997.

    Each stage, run by r3-stage.sh (dc99008a):

    • renders a --dry-run plist into $GUARD (never LaunchAgents), with its deadline 2 h ahead;
    • bootstraps it, and boots it out as soon as its first run has exited;
    • reports launchctl print and the kernel's System Policy lines.

    A dry run only reads and hashes the stop script; it never runs it.

    R3a, run folder on the external volume (15:41:06–08 UTC): FAIL, as expected.

    • io.wordhoard.s1-stop-dry-run-a: runs 1, last exit 1.
    • Kernel: System Policy: zsh(16955) deny(1) file-write-create …/run/s1-2026-11-05/stop.log. This is the same cause as v7's guard.
    • No DRY_RUN_OK.

    R3b, log and marker on the internal disk (15:41:19–23 UTC): PASS.

    • io.wordhoard.s1-stop-dry-run-b: runs 1, last exit 0.
    • A fresh $GUARD/s1-2026-11-05/DRY_RUN_OK at 15:41:22 UTC: mix started on the serving checkout on the external volume; target v2:5434, cluster 7693849764459364596; editor nil; service not alive; stop script f870cce6… read, not run; nothing stopped.
    • No kernel deny was logged for the volume.

    R3c was not run. Both test jobs are unloaded.

    Which guard path S1 uses, and why: the launchd guard in the R3b layout.

    • On this host, launchd's /bin/zsh may read the external volume, so mix loads the checkout and its build. Writes there are denied.
    • So the run directory (stop.log, DRY_RUN_OK, STOPPED), the stop-script copy (sha f870cce6…, the same as the frozen manifest's) and the launchd log all live under $GUARD on the internal disk.
    • For Curation: matched comparison support, policy arms and stop guard (#298 Part C, C1–C9) #303's C6, this gives one correction: the guard is proven under launchd, but not with the run folder on the external volume.

    Two cautions are recorded in the runbook for S6:

    1. The real stop's only write on the volume is File.rm of run/ollama.pid, whose result ServiceProcess.stop ignores. A denial would leave a stale pid file; the stop is still proven (editor off, port closed, STOPPED written internally).
    2. _build/dev must be current before the guard loads, because launchd's mix cannot compile onto the volume. It is current at 4a0b997 now.

    Not touched: no restart of 4007; no /healthz, /on/* or /define/*; nothing seeded, admitted or bound; no gateway resume; no model call; nothing deleted. The serving tree is clean at 4a0b997.

    Record shas (first 12):

    • service-start.log 72553587ad1b
    • setup-generic.log a66af6de7fca
    • setup-hitchens.log 8b5de26c133b
    • setup-bierce.log d66df07c42c0
    • configs-check.log 35fa3adb9c71
    • service-stop.log 6da582a6841c
    • r3a.log 634aea25484c
    • r3b.log 9fd956b02fe1

    Next, no yes needed: R1, the read-only dry run (compare --dry-run). It gives the earliest shelf expiry per case, which sets S6's planned start inside a news window.

    Then S6, the batch, as its own yes: the start time, the day's charged time, the batch id, and the guard's load and unload.

  8. holden commented on Nov 4, 2026

    @holden
    MemberAuthor

    S6: batch s1-2026-11-04 run, 4 November 2026, 15:53:45–15:59:12 UTC (#307 executor)

    Owner's yes, in chat on 4 November 2026: "Run batch s1-2026-11-04". It covers:

    • the rename, the rechecks, the repeated dry run and the freeze;
    • service start, gateway resume and readiness (no smoke);
    • the guard load;
    • the batch;
    • verification and coverage;
    • then gateway pause, service stop and guard unload.

    The run is on 4 Nov rather than 5 Nov because the dry run's earliest shelf expiry (love, 21:46:02 UTC) left room for a start before 18:45 UTC. The run folder stays run/s1-2026-11-05/; the step log is steps.txt there.

    Before the first call

    • Batch id: renamed to s1-2026-11-04 in manifest.json and judgments-template.json, so this batch id has never been used.
    • First freeze refused: "freeze refused: pilot pages", with nothing written. My prepared manifest.json listed love first; the check compares the list in order. It now follows Pilot.autonomous_pages() order: the same ten pages.
    • Rechecks (s6-precheck.sh, 15:51:44 UTC): 15/15 PASS.
      • Cluster ok; 42 migrations; editor NULL.
      • 0 open attempts, 0 comparison attempts, 0 queued curation jobs besides the dispatcher; 11 model configs.
      • 11435 closed; serving checkout clean at 4a0b997; 4007 pid 1661; / 200, /health 302.
      • No wordhoard launchd job.
      • Today's charge was 192,340 ms, under the 5-minute start rule.
    • Dry run repeated: "would run yes: editor disabled, batch id unused". The cases are identical to R1, and the earliest expiry is unchanged.
    • R2 freeze: frozen-manifest.json sha256 c4f1de95d82977a890560d6d55597efb2c1561f84ce52b74388efe1a7fd262ac.
      • Frozen at 15:52:40 UTC; attempt cutoff 121; deadline 18:52:40 UTC (epoch 1793818360); code 4a0b997; stop script f870cce6….
      • Arms: V config 8 81a12ded… (instruction a3219a5c), G 9 3acaeb4e… (0423863d), H 10 baa0ce4f… (3463bce9), B 11 e0420dbe… (b02355a9).
      • Compositions: nepotism 6, greed 7, war 2, love 1, coward 5, patience 8, justice 9, truth 4, hope 3, bestiality 10.
    • Service and gateway:
      • service start: pid 23461, bound to v2.
      • resume --model-config qwen3.5-9b-whole-voltaire-v7: available, fence 121.
      • ready: every check ok, served digest 6488c96f….
      • Max attempt was still 121, so readiness made no call.
    • Guard: io.wordhoard.s1-2026-11-04-stop in the R3b layout.
      • Plist in $GUARD (ee6d4a62…), with the freeze's deadline.
      • Run dir $GUARD/s1-2026-11-04, holding stop.exs (f870cce6…) and physical-stop.exs (fcb3fdb6…, which trial-stop.sh runs from beside the stop script if a stop fails).
      • First tick exit 0.

    The batch

    compare --batch s1-2026-11-04 --pilot --window-minutes 60 --arms V=…voltaire-v7,G=…generic-v1,H=…hitchens-v1,B=…bierce-v1 --out run/s1-2026-11-05/record/batch.json:

    • Exit 0; halted nil.
    • Its own freeze at 15:53:46 UTC gave the same snapshot 2cd9dae3….

    All ten cases are matched, with attempts #122–#161 in order V, G, H, B per case. The outcomes are the runtime's attempt outcomes, not scores:

    case V G H B
    nepotism #122 abstained #123 accepted #124 accepted #125 accepted
    greed #126 abstained #127 abstained #128 abstained #129 abstained
    war #130 accepted #131 accepted #132 accepted #133 accepted
    love #134 accepted #135 accepted #136 accepted #137 accepted
    coward #138 accepted #139 accepted #140 accepted #141 accepted
    patience #142 refused #143 refused #144 accepted #145 abstained
    justice #146 accepted #147 accepted #148 accepted #149 accepted
    truth #150 abstained #151 accepted #152 accepted #153 accepted
    hope #154 accepted #155 accepted #156 accepted #157 accepted
    bestiality #158 abstained #159 abstained #160 abstained #161 abstained

    Verification (read-only, s6-verify.sh, 15:59:26 UTC): all PASS

    Coverage (mix dd.curation.coverage --comparison s1-2026-11-04 --record record/batch.json)

    Report curation-comparison-v2: 10 cases matched. Totals per arm (counts only, before scoring):

    arm select / abstain / refused highlights lead is priority (true/false/nil) model-written reasons charged ms
    V 5 / 4 / 1 12 5 / 3 / 2 16 75,473
    G 7 / 2 / 1 18 7 / 1 / 2 25 74,370
    H 8 / 2 / 0 11 8 / 0 / 2 19 72,683
    B 7 / 3 / 0 9 7 / 1 / 2 16 70,452

    Stopped and unloaded (15:59:50–51 UTC)

    Final state (16:00:01 UTC, s6-final.log):

    • Editor NULL; gateway paused.
    • Last attempt 161, 0 open; today's charge 485,318 ms.
    • 11435 closed.
    • Serving checkout 4a0b997, clean; 4007 pid 1661 (no restart); / 200, /health 302.

    Not touched: no /healthz, /on/* or /define/*; nothing seeded, admitted, bound or published; no smoke call; nothing deleted.

    Record shas (first 12):

    • record/batch.json 20b9d3aa2ec9
    • coverage.json 6117b8e9edca
    • frozen-manifest.json c4f1de95d829
    • run.log fb2acc1dbdee
    • s6-verify.log 2994779fa50d
    • s6-final.log c89d1e137382

    Next: S7, scoring under scorecard v2 (156ee531…), within 20 hours, so by 2026-11-05 15:59 UTC. It fills judgments-template.json per case, arm and choice. Then comes S8: the report and the D4 seat recommendation on #305.

  9. holden commented on Nov 4, 2026

    @holden
    MemberAuthor

    S7 and S8: scored, reported, D4 given, 4 November 2026 (#307 executor)

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