Skip to content

#305 Stage 2: the four policy arms offline (D6), runbook and execution (under #305) #306

Closed

Description

@holden

Implementation issue, 4 November 2026, Stage 2 of the persona META issue #305, under #193. Posted on the owner's words in chat, 4 November 2026 ("Post it"). Narrow by design: one offline batch of forty non-publishing calls on the ten pilot pages, scored within 20 hours, ending in one decision comment on #305 under its §3.1 rule. No code beyond what PR #303 carries; no publication; no new instruction; the editor disabled throughout. Everything before the first model call can be prepared now. No live step follows from this text.

1. Outcome

One D6 report and one decision on #305: which of A0 (current v7), P (text cut), L (fixed Bierce lead) and PL (both) earns the v8 publishing run; or NOT JUSTIFIED (no arm met the floors on ten complete matched cases); or INCONCLUSIVE (fewer than ten complete matched cases after the one retry, with the cause and the next batch id). This issue closes on that comment.

2. Prerequisites, each verified with ids before the first call

  • P1. Curation agents: close out v6, retest v7, then controlled model and personality comparisons on identical packets (under #193) #298 closed, its closure comment linked here.
  • P2. PR Curation: matched comparison support, policy arms and stop guard (#298 Part C, C1–C9) #303 merged at its reviewed exact head and deployed. Verify: the merge commit's second parent equals the reviewed head (3449553 at this writing; re-verify at merge time); migration 20261103202412_add_comparison_attempts applied (v2's migration count rises by one from 41); the serving checkout fast-forwarded to GitHub main; no config/*.exs or mix.lock change in the deployed range (git diff --stat of the range), so no restart; /healthz and one existing receipt page answer as before (never /on/* or /define/* on 4007 or the public host: a visit admits a run).
  • P3. Editor disabled and staying so: editor_enabled_version_id NULL on global-default (read-only query, cluster 7693849764459364596, port 5434).
  • P4. Runtime ready: the private service on 127.0.0.1:11435 with the 9B artifact; mix dd.runtime ready --model-config qwen3.5-9b-whole-voltaire-v7 all checks ok; gateway available, not paused or quarantined; no attempt admitted, dispatched or uncertain.
  • P5. Model configs for the four arms. A0 and P share qwen3.5-9b-whole-voltaire-v7 (config 8, hash 81a12ded…), P as its :text cut. L and PL need one new immutable model config on the unchanged artifact under instruction voltaire-whole-entry-v7-fixed-lead-v1 (sha recorded), same generation settings (seed 195, temperature 0, 8K context, 1,024 output). Registering it is a live step on the owner's yes; its hash goes into the manifest. Within a cut the fingerprints must match (A0 = L, P = PL), which the dry run proves.
  • P6. Day budget: charged time for the day under 5 minutes at start (ledger query); no other batch that day; start early in the news window (after the 02:15 UTC cleanup, never within an hour of a news expiry) so the earliest shelf expiry lies beyond the batch deadline.
  • P7. Old guards. The v6 and v7 launchd stop jobs are still loaded and exit harmlessly on their STOPPED markers; unload both on the owner's yes before loading the C6 guard, so only one guard exists.

3. Preparation, no model calls (can start now in a worktree; R1 after P2)

  • R1. Dry run on v2. mix dd.runtime compare --batch d6-<date> --pilot --dry-run --arms A0=qwen3.5-9b-whole-voltaire-v7,P=qwen3.5-9b-whole-voltaire-v7:text,L=qwen3.5-9b-whole-voltaire-v7-fixed-lead,PL=qwen3.5-9b-whole-voltaire-v7-fixed-lead:text. Expected: ten cases listed with per-arm packet hashes and evidence fingerprints (A0 = L and P = PL per page; A0 ≠ P), omissions by reason (text_cut_excluded under P and PL), fixed_lead set for L and PL on every page with an applicable Bierce entry and null elsewhere (bestiality is never hard-coded), 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/d6-<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 version 7 for Voltaire with its manifest hash 7db26817…, model config ids and hashes, generation settings, cut and lead rule per arm), the ten pages with composition ids, the case identity scheme d6-<date>:SLUG:n, the fingerprint rule (C1), scorecard v2 sha 156ee53110b7c8ecad98291a8087c91b831e2a07c1c47f19aca4e269ee7cd1a9, the §3.1 rule text of 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 copied verbatim with its sha, the batch 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/d6-<date> --dry-run writes DRY_RUN_OK (launchd can start mix, reach v2 read-only, hash the stop script, stop nothing). The real load, with the deadline, is a live step on the owner's yes immediately before the run; C6's stop.exs accepts the already-disabled editor.
  • R4. Scoring record template, the 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 (model, rule, withheld); per page and arm the outcome (useful, lead-only, abstained, incorrect abstention, refused); the review clock; the file's sha when frozen. No provider text.
  • R5. Timing and halt rules written into the run folder's RUNBOOK.md, copied from 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: 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.

4. The run (on the owner's yes: "run D6 batch d6-")

  1. P1–P7 re-checked and logged; R1 repeated the same day; R2 frozen; guard loaded (owner's yes); editor_enabled_version_id NULL confirmed.
  2. mix dd.runtime compare --batch d6-<date> --pilot --arms … (the R1 arms, without --dry-run). Forty calls, about ten minutes at the observed 6–14 seconds each; the task runs 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 d6-<date> --record run/d6-<date>/coverage.json (read-only; every case listed, including any with no attempts); 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 d6-<date>:SLUG:2 on a new snapshot, under a re-rehearsed guard; never one arm alone; never combined across snapshots.

5. Scoring, within 20 hours of the batch

Developer-scored on scorecard v2, unchanged: every proposed choice of every arm, abstentions as zero-choice outcomes, lead correctness where the page has a priority candidate, highlights by kind, model-written reasons and rule statements judged and counted apart. Evidence inspected before its expiry; only ids, digests and short bases retained; the judgments file frozen with its sha.

6. The report and the decision (one comment on #305)

Per arm: cases complete and matched (of 10), proposed choices, supported, insufficient, wrong, bad reasons by origin, useful pages, lead correct (of pages with a priority candidate), highlights by kind, charged milliseconds, prompt tokens. Then the floors per arm (7 useful, 90 percent supported, zero bad reasons of any origin), the total order from §3.1, and one of: "Arm X qualifies and is the v8 policy", NOT JUSTIFIED, or INCONCLUSIVE with the cause and the next batch id. Linked: the manifest sha, the coverage record sha, the judgments sha, the attempt id range, the step log. The decision is posted before anything else runs.

7. Acceptance (closes this issue)

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

Register the fixed-lead model config; unload the v6 and v7 guards; load and unload the C6 guard; run the batch (and the one retry); nothing deleted, nothing enabled, nothing published.

9. Not in this issue

Stage 1 (#307, the persona measurement, which runs first on the current policy) and Stage 3 (the admission PR for voltaire-whole-entry-v8, the love gate, the publishing run, in #305) are not here; dossiers are #197; the fixture-only consensus engine is #300 L1.

Activity

  1. changed the title [-]#305 Phase A: the four policy arms offline (D6), runbook and execution (under #305)[/-] [+]#305 Stage 2: the four policy arms offline (D6), runbook and execution (under #305)[/+] on Nov 4, 2026
  2. holden commented on Nov 4, 2026

    @holden
    MemberAuthor

    Renamed 4 November 2026 on the owner's words in chat: Phase A is Stage 2 of the persona META issue #305. Stage 1 is #307, the first persona measurement, which runs first on the current policy; the two batches run on different days for the budget. Nothing else in this issue changes.

  3. holden commented on Nov 4, 2026

    @holden
    MemberAuthor

    Note, 4 November 2026, 16:10 UTC: runs today, in the Stage 1 executor session

    On the owner's words in chat ("Yes, run Stage 2 today"), #305 is amended (#305 (comment)): a second batch may start the same day with at least 15 minutes of budget remaining, the earlier batch ended, and the new snapshot's earliest expiry beyond the new deadline. P6 reads accordingly; P7 is already met (v6 and v7 guards unloaded under #307 S4, c6635001973). Executor: the "Stage 1 persona measurement" session, under the same live-step rules. 4 November charged 485,318 ms at the amendment; 21.9 minutes remain.

  4. holden commented on Nov 4, 2026

    @holden
    MemberAuthor

    D6 run: batch d6-2026-11-04, 4 November 2026, 16:18:19–16:23:20 UTC (executor in the #307 session)

    Owner's yes, in chat on 4 November 2026: "Yes to Stage 2". It covers the run folder's RUNBOOK.md steps 1–10 as put:

    • register the fixed-lead config;
    • the four-arm dry run, the freeze, resume and readiness;
    • the guard's dry tick and load;
    • the batch, verification and coverage;
    • then pause, service stop and guard unload.

    The owner's decision first reached this session relayed through the audit session; the live steps ran only on the words typed here. It runs under #305's amendment c6635520789: a second batch the same day. The run folder is run/d6-2026-11-04/; the step log is steps.txt there.

    Prerequisites P1–P7 (d6-precheck.sh, read-only, 16:16:44 UTC): 20 PASS

    Before the first call

    • Service: started, pid 35702, bound to v2.
    • P5, register (16:16:58 UTC): config 12 qwen3.5-9b-whole-voltaire-v7-fixed-lead, instruction voltaire-whole-entry-v7-fixed-lead-v1 (sha 2af1ca1a…), config_hash 4a4508bb922d0150a3fd276d23f68cb1d0e7c8a93faf45a2a00680460082946d.
      • Model, manifest, weights, template, quantization, runtime, licence and generation all equal config 8's.
      • The freeze checks by code that the instruction is v7 with only the fixed-lead clause (e1849ae0…) inserted before v7's lead header.
    • R1, dry run with the four arms:
      • Before registration it refused (model_config_unknown), as expected.
      • After registration it reports "would run yes" (16:17:12 UTC). Per page, A0 = L and P = PL, A0 ≠ P, with four distinct packets.
      • fixed_lead is c1 on the eight pages with a Bierce priority candidate, and nil on greed and bestiality (never hard-coded).
      • P and PL omit by text_cut_excluded.
      • Snapshot 2cd9dae3…. A0's packets are byte-identical to Stage 1's V packets.
      • No attempt was written.
    • R2, D0 manifest: frozen-manifest.json sha256 c58b7b795a5f785a41fe2797dff18c379896ac4bc970f1f7e3fc718862f63f05.
    • Gateway: resume, then available (fence 161). ready: 7 checks ok, with no smoke call.
    • R3, guard (#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's proven R3b layout: run dir, stop-script copies and logs under ~/Library/Application Support/DictionaryTrial):
      • Dry tick: DRY_RUN_OK at 16:18:02 UTC, no kernel deny, booted out.
      • Then io.wordhoard.d6-2026-11-04-stop was loaded with the freeze's deadline. Its first tick exited 0.

    The batch

    compare --batch d6-2026-11-04 --pilot --window-minutes 60 --arms A0=…voltaire-v7,P=…voltaire-v7:text,L=…voltaire-v7-fixed-lead,PL=…voltaire-v7-fixed-lead:text --out run/d6-2026-11-04/record/batch.json: exit 0, halted nil, 10 cases matched, attempts #162–#201.

    The outcomes below are attempt outcomes, not scores:

    case A0 P L PL
    nepotism #162 abstained #163 accepted #164 accepted #165 accepted
    greed #166 abstained #167 abstained #168 accepted #169 accepted
    war #170 accepted #171 accepted #172 accepted #173 accepted
    love #174 accepted #175 accepted #176 accepted #177 accepted
    coward #178 accepted #179 accepted #180 accepted #181 accepted
    patience #182 refused #183 accepted #184 accepted #185 accepted
    justice #186 accepted #187 accepted #188 accepted #189 accepted
    truth #190 abstained #191 abstained #192 accepted #193 accepted
    hope #194 accepted #195 accepted #196 accepted #197 accepted
    bestiality #198 abstained #199 abstained #200 accepted #201 accepted

    A0's ten outcomes repeat Stage 1's V arm exactly: the same packets, temperature 0 and seed 195.

    Verification (d6-verify.sh, read-only, 16:23:36 UTC): all PASS

    • 40 comparison attempts after the cutoff, 10 per arm, every identity column set, none open. All are completed: 32 accepted, 7 abstained, 1 refused.
    • 10 cases have four arms with A0 = L and P = PL.
    • Nothing was written since the freeze to proposals, composition versions, compositions or publications. curation_item_bases is unchanged (397), and no version or basis names a comparison attempt.
    • Editor NULL.
    • Charged: 266,784 ms for the batch. Today's total is 752,102 ms of 1,800,000.

    Coverage (curation-comparison-v2, sha e1c6057b0168…): 10 matched

    Counts only, before scoring:

    arm select / abstain / refused highlights model reasons rule statements charged ms prompt tokens
    A0 5 / 4 / 1 12 16 0 76,318 32,006
    P 7 / 3 / 0 22 29 0 74,056 26,940
    L 10 / 0 / 0 6 8 8 57,542 32,914
    PL 10 / 0 / 0 15 17 8 58,868 27,848

    Stopped (16:23:50–54 UTC) and final state

    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 26357923d532
    • coverage.json e1c6057b0168
    • run.log 8b4368ed1826
    • verify.log 375c601ddb02
    • final.log 1639a4976e8d

    Scoring: started 16:24 UTC, within the 20-hour window. The evidence was rebuilt read-only at 16:24:12 UTC, and every packet hash equals its attempt's. Two fresh scorers are working blind. The §3.1 decision goes on #305 when scoring is done.

  5. holden commented on Nov 4, 2026

    @holden
    MemberAuthor

    Scored and decided: NOT JUSTIFIED, 4 November 2026 (executor in the #307 session)

    The decision is on #305: #305 (comment)

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