EV-007 main run: 3-arm pre-registered A/B on the learning loop (bare / harness-no-learning / harness-learning)

[shojikumaru]

Why

EV-007 P0 (2026-08-21) proved the short-cycle distillation loop wires up — and simultaneously showed the content it moves is near-worthless progress chatter (REPORT-P0 §4.3), the intake gate is advisory (§4.2), and pending flushes could be silently clobbered (§4.5). The #144 adjudication (2026-08-24) ruled: the main-run pre-registration is written against the new learning-loop design (#147 raw layer / #148 nightly cross-model review / #145 Last session pointer / #170 VF-GR caps), all of which have since merged (main = v0.19.0). Separately, EV-008 completed the overflow-sentinel effect measurement and the owner declared default-on GO (v0.14.1) — so the original #159 D5 arm definition ("bare / always-harness / sentinel") would now re-measure what EV-008 already answered.

This issue is the home for the EV-007 main run: a 3-arm pre-registered A/B measuring the README core claim — the harness learns from raw experience and the next job gets better — on the sealed EV-006 rig.

Clarify writeback (owner decisions, 2026-08-27)

#	Question	Decision
C1	Arm definition	Redefine on the learning-loop axis: bare / harness with learning loop disabled / harness with learning loop enabled (raw layer + nightly review + apply). Sentinel stays default-on in both harness arms per the EV-008 GO. This supersedes #159 clarify D5; a superseding comment is posted on #159.
C2	Experiment pin (HARNESS_DIR)	Update the pinned worktree to v0.19.0 (current pin @3eab97a is a v0.12.1-era side branch without the new learning loop; incompatible with the #144 ruling). Pin SHA sealed in the prereg.
C3	apply-step ordering	apply implementation lands first (own issue, size L). Prereg design proceeds in parallel, but before sealing: one light full-loop wiring re-check (P0-style smoke incl. apply, on the updated pin), then a second owner confirmation. Owner's words: "ライトに確実にいこう".
C4	Review (judge) model	Kimi K3 — the only path validated by the 34-day real-corpus replay (5/6 windows, fabrication guard exercised); invocation via the decoration-stripping wrapper with measured timeouts. Sealed in the prereg.

Scope of this lane

1. Write experiments/ev007/PREREGISTRATION-MAIN.md against the new design (arms per C1, pin per C2, judge per C4; metrics, pass/fail lines, budget, empty-run rules — P0 decisions carried forward where still valid).
2. Observation targets carried in from open advisories: handoff files actually containing the 4 declared fields (PR last-session: session index with host-side fold and handoff files (#145) #172 Opus advisory — model-contract only, never verified); status of P0 findings §4.1–§4.5 under the new design (4.5 is claimed solved by learning loop: formalise the never-deleted raw layer and cut it weekly (splits 'where lessons accumulate' from 'what rides every prompt') #147/learning loop: weekly cross-model review over the raw layer — replace hash-equality k>=2 with theme-level judgement (and implement the 'reviewed by a different AI' claim for the first time) #148 — the prereg treats that as a guarantee to verify, not assume).
3. Pin worktree update to v0.19.0 (C2) before the wiring re-check.
4. Prereg passes heterogeneous 3-seat upstream review (L1-9; experiment definition = requirements-class work) before sealing.
5. Seal gate (C3): apply issue merged → light wiring re-check → owner GO → seal → run.

Done when

• PREREGISTRATION-MAIN.md drafted against v0.19.0 design (arms C1 / pin C2 / judge C4) with metrics, pass lines, budget cap, and deviation rules
• 3-seat heterogeneous upstream review of the prereg: all blocking findings resolved
• Pin worktree updated to v0.19.0 and recorded
• Light full-loop wiring re-check executed after the apply issue merges (result reported to the owner)
• Owner GO obtained → prereg sealed (immutable record, EV-006/EV-008 sealing conventions)
• Main run executed per the sealed prereg; report published with the same correction-ledger discipline as EV-008

Files to touch (prediction)

• experiments/ev007/PREREGISTRATION-MAIN.md, experiments/ev007/runner/* — outside git (same posture as P0/EV-006)
• pinned worktree caty-agent-harness-ev-pin (git worktree move to v0.19.0 — no source edits)
• No caty-agent-harness source files are modified by this lane. Repo-side artifacts (report/docs) will be declared in a separate PR lane when the run completes.

Refs: #144 (adjudication comment 2026-08-24), #159 (D5 — superseded by C1), #148 #147 #145 #170, apply issue #201, PR #172 advisory, EV-008 record in #159 / docs/benchmark.md#ev-008.

[shojikumaru]

WIP

• agent: Alpha (Claude Code / Fable 5, MBP session 2026-08-27)
• date: 2026-08-27
• Files to touch (confirmed declaration): experiments/ev007/PREREGISTRATION-MAIN.md + experiments/ev007/runner/* (both outside git), pinned worktree caty-agent-harness-ev-pin (worktree checkout move only). Zero caty-agent-harness tracked files — no branch is created for this phase.
• Branch: none (no repo writes in this phase; a branch will be declared in a fresh WIP re-declaration before any repo-side artifact lands)

Parallel-GO note: declared file set is disjoint from every open lane by construction (no tracked files). The prereg-design phase of this lane runs read-only against the repo.

[shojikumaru]

Progress (2026-08-27): PREREGISTRATION-MAIN.md DRAFT v0.1 staged at experiments/ev007/PREREGISTRATION-MAIN.md (outside git, per the declared file set).

Three v0.19.0 facts measured today (file:line survey at 4221b8a) that shaped the draft:

1. No single learning-loop off-switch exists. The pause marker disables the whole harness (task-runner and sentinel included — activation-manifest.tsv), so arm B is defined by construction: Stop/PreCompact hooks unregistered + flush-intake.sh unscheduled + review.conf unwired. Host-side STATE.md is then provably static (the only host writer is flush-intake).
2. The intake gate is still advisory — the task model holds Write on STATE.md and the bootstrap requires direct STATE writes at CHECKPOINT (P0 §4.2 contamination path is current). The draft pre-declares a per-job contamination diff + count for arm B instead of pretending the arm is clean; threshold for "degraded evidence" is fixed at seal.
3. The sentinel ships opt-in (OVF_SENTINEL unset = off — spawn_step.sh:46-49). Both harness arms will export an identical pinned env block (active, 80000/50), matching the EV-008 GO posture; sealed verbatim in §8.

Draft structure: arms per C1 / pin per C2 (v0.19.0) / judge per C4 (Kimi K3, reviewer_timeout_s=900); metrics carry P0 Decisions 1–2 forward; new learning contrast = B-vs-C on the final measurement block; observation targets include the PR #172 handoff 4-field check and a deliberate two-jobs-before-intake window to verify (not assume) the #147 no-clobber guarantee. Series length, pass lines, contamination threshold, and budget cap are deliberately left to seal time — they depend on the #201 apply contract and the light wiring re-check (owner decision C3).

Next in order: #201 design doc → L1-9 3-seat upstream review → implementation + merge → pin update to v0.19.0 → seal-time values filled into this draft → prereg 3-seat upstream review (on the seal candidate) → light full-loop wiring re-check → owner confirmation → seal → run.

[shojikumaru]

Pilot P0.5 "oracle seeding" — SIGNAL (2026-08-27). Owner asked for a light feasibility probe before building the apply step: if promotion worked perfectly (ideal distilled rules hand-seeded into ## General rules), does the next job's primary metric move? Design + prediction were fixed before job 1 in experiments/ev007/PILOT-ORACLE.md; 4 jobs ran (2 per arm, sealed p3-M-i1, haiku, fresh v0.19.0 worktree caty-agent-harness-ev019; arms differ only by 3 seeded rules; no learning machinery ran in either arm).

pooled (2 jobs × 20 items)	control B	seeded L
quote NG	9/40 (22.5%)	3/40 (7.5%)
…true misquote	9	0
…unknown/abstain	0	3
correct	10/40	8/40
attempts / charge tokens	5 / 3.74M	10 / 4.79M

Both pre-registered signal criteria met (gap ≥10pt; L true-misquote 0 vs B ≥3). L's only NGs are abstentions — the exact trade the rules ask for. Caveats recorded with equal weight: n=2/arm, contrast driven by B-j2, correctness unmoved, and a real cost signal — the "re-verify your quote" rule doubled attempts (donecheck retries) and +28% tokens.

Carried into the lane: (a) the main-run prereg §5 gains attempts/tokens as a pre-registered secondary outcome; (b) #201 design must treat promoted-rule wording as first-class (a rule can be effective and expensive at once); (c) sample-size section assumes P0-scale per-job variance.

Total pilot cost: 8.53M charge tokens, 4 jobs, ~35 min wall. This probe was directional only; it authorizes build-out, not effect-size claims.

[shojikumaru]

C5 — owner design decision (2026-08-27, after pilot P0.5): primary analysis reframed from endpoint A/B to longitudinal before/after.

Owner's adjudication of the pilot readout, recorded verbatim in intent:

1. The pilot's cost signal (2× attempts in the seeded arm) is expected, not surprising — learning cost is paid up front; distillation's essential power is that the second encounter with the same mistake doesn't happen. The asset compounds; leverage arrives later, while the bare arm keeps repeating the same mistakes indefinitely.
2. A same-start endpoint race ("よーいスタート勝負") therefore measures the wrong time axis. The claim under test is never repeating the same mistake across future jobs = getting smarter over time, so the verification method must show a trajectory (before/after), not a finish-line snapshot.

Prereg draft updated (§1 framing + §5 metrics):

• Headline: repeat-mistake rate per arm — of the mistake classes seen in an arm's before-block, the fraction recurring in its after-block. Prediction: arm C → 0, arms A/B flat. The series must re-present the same trap classes by construction (fixed at seal).
• Slope of the primary metric over job index per arm; the C-vs-B slope difference is the effect. Endpoint contrasts demoted to secondary anchors (EV-006 comparability).
• The 3-arm structure (C1) is retained — the no-learning arms are the flat baseline that makes the slope attributable to the loop rather than task familiarity.
• Cost becomes a trajectory question too: does the per-job verification cost amortize/decline/recur over the series — pre-registered, reported alongside quality, never netted away.

PILOT-ORACLE.md carries a dated owner-adjudication appendix superseding its "unexpected" framing (append-only; original text left in place).

[shojikumaru]

Pilot P0.6 "before/after repeat-mistake" — SIGNAL; the C5 metric survives its own pilot (2026-08-27). Full record: experiments/ev007/PILOT-BA.md (design + prediction fixed before any job; 7 new jobs, ≈17.4M charge tokens, haiku, v0.19.0 worktree).

Design: arm C ran an unseeded before-block (3 jobs — one pre-registered variance-guard extension fired), then ONE fair learning event (rules distilled ONLY from C's own observed misquotes + corpus, no gold, recorded verbatim before the after-block), then 2 after-jobs. Control B: before-block reused from P0.5, 2 new after-jobs, no learning.

repeat-misquote rate (before-block mistakes recurring in after-block)
C (learning)	0/3 (0%) — after-block has zero true misquotes anywhere
B (control)	4/9 (44%) — plus 12 total new/old misquotes in the after-block

Both pre-registered signal criteria met. Three findings carried forward:

1. The owner's framing held empirically: the learning arm stopped repeating its own mistakes; the control arm kept repeating them indefinitely — the compounding-asset story, measured.
2. Cost inverted vs P0.5: self-derived rules made C cheaper after learning (avg attempts 5.0→4.5, tokens 2.68M→2.01M per job). Rule wording drives cost; rules grounded in real failure modes beat imagined ideals on both quality and cost. First-class input to learning loop: apply step — reviewed candidates → durable-tier writeback (completes the promotion layer for EV-007) #201 apply design.
3. Goodhart warning, pre-registered into the main prereg: C's after-block answered less (abstains up, correct 4/20 vs ≈8.3 before). §5 now carries a correctness no-harm guard — quote improvement achieved mainly by mass abstention is declared as such, never counted as a win alone.

Lane consequence: proceed to #201 design as planned; both pilots (P0.5 endpoint, P0.6 trajectory) now exist as calibration references for the main-run sample-size and guard thresholds.

[shojikumaru]

L0-9 resume checklist (Alpha, new session 2026-08-27, resuming the seal sequence)

1. Lock — mine: WIP declared by Alpha 2026-08-27 (same lane, same agent), < 72h, not stale. No RELEASE/HANDED-OFF/TAKEOVER since.
2. File scope — re-measured gh issue list + gh pr list (2026-08-27): open PRs repo-state: add generation stamp + status.json + caller workflow (family-os#127, pin v0.13.1) #209 (repo-state caller) and task-runner: sentinel ledger confluence + task-level task_end (#187) #206 (task-runner ledger) touch tracked repo files only; this lane's declared set (experiments/ev007/* outside git + caty-agent-harness-ev-pin worktree checkout move, zero tracked files) remains disjoint from every open lane by construction. Parallel-GO holds.
3. Branch — none declared for this phase (no repo writes); origin/main fetched clean @ 0babfef (= v0.20.0 tag commit).
4. Done when — issue body unchanged since handoff (updatedAt 2026-08-26T20:01:11Z); all 6 items as declared. Note: body says "pin to v0.19.0" (C2 wording at decision time); operational target is v0.20.0 since learning loop: apply step — reviewed candidates → durable-tier writeback (completes the promotion layer for EV-007) #201 apply merged after C2 — already noted in the learning loop: apply step — reviewed candidates → durable-tier writeback (completes the promotion layer for EV-007) #201 MERGED comment, and will be explicitly re-confirmed with the owner at the §8 seal gate (C3).

Next in order: pin move → §8 seal-value fill → prereg 3-seat review → light full-loop re-check → owner confirmation → seal.

[shojikumaru]

Progress (2026-08-27, seal sequence steps 1–2):

Pin worktree updated (Done-when item 3, executed against the retarget). caty-agent-harness-ev-pin moved 3eab97a (v0.12.1-era) → 0babfef870c404c754f1dde044f619ef98a94dc6 = git describe → v0.20.0 (exact tag), working tree clean before and after. Rationale: C2 said v0.19.0 at decision time, but #201 apply merged after C2 and an arm-C run without the apply step would not test the claim; the retarget is flagged for explicit owner re-confirmation at the C3 second confirmation (as pre-noted in the #201 MERGED comment). caty-agent-harness-ev019 (v0.19.0) and the main clone untouched.

PREREGISTRATION-MAIN.md → DRAFT v0.2 (outside git, per declared file set). §8 now carries the values determinable today, verbatim: pin SHA/describe; sentinel env block (EV-008 GO posture; TR_STEP_TIMEOUT_S value frozen at seal after the wiring re-check); arm-C review.conf (producer=claude-haiku-4-5, Kimi K3 via the durable wrapper path experiments/ev007/tools/review-wiring/kimi-reviewer-wrapper.sh, timeout 900s, other keys shipped defaults; wrapper sha256 recorded at seal); arm-C apply invocation (APPLY_MAX_AUTO=10 … --auto-capability-facts, no --approve during the run — awaiting-approval counted, not promoted; autonomy is part of the claim); apply contract reference (#201 = merge 0babfef = release v0.20.0, follow-ups → #207). Series / pass-lines / thresholds / caps remain seal-time lines. §3 pin text retargeted with a dated changelog note.

Cross-check on the new pin: the two v0.19.0 file:line facts the draft leans on still hold at v0.20.0 — spawn_step.sh:46-49 (OVF_SENTINEL unset = off) and the flush-intake.sh append-never-replace archive path (cat-append when destination exists).

Next: prereg 3-seat upstream review (seats via loom-seats, --absent opus-5 since the prereg is Alpha-drafted) → light full-loop wiring re-check on this pin → owner re-confirmation (pin retarget + auto-capability-facts reading) → seal.

[shojikumaru]

Prereg 3-seat upstream review (L1-9) — record + adjudication (2026-08-27)

Panel (loom-seats, size L, --absent opus-5 — prereg is Alpha-drafted, Opus is Alpha's lineage; no same-family seat, no exception record needed; writer≠every seat):

seat	requested	actual	liveness	delivery	verdict
Kimi K3	kimi-k3	kimi-k3 (CLI, -m kimi-code/k3)	PING-OK	file, no truncation	GO (blocking fixes: CR-1, CR-2, M1–M6)
Grok 4.6	grok-4.6	grok-4.6 (CLI, --effort high)	PING-OK	file	NO-GO (C1–C3 CRITICAL)
GLM 5.3	glm-5.3	glm-5.3 (CLI, isolated cwd, hooks off)	PING-OK	file (stdout intact)	GO (blocking: F1, F2 + fix list)

Blind, fresh-context, read-only; same brief; per-seat isolated output dirs. All three seats independently fact-checked the draft against the v0.20.0 pin (their evidence tables agree with each other and with my own survey). Review outputs retained at the session scratchpad prereg-review/seat-*/review.md (verbatim copies will ride the sealed record).

Adjudication is by evidence, not verdict count (a GO seat does not negate another seat's finding). Convergence map and rulings:

Adopted — 3-seat convergent (CRITICAL/blocking):

1. Headline metric had no operational mistake-class identity (Grok C1 = Kimi CR-1 = GLM F2). Fixed: measurement blocks are P0.6-style re-presented trap items; identity = question id + the scorer's own mechanical true-misquote check; question-id→class map becomes a §8 seal line.
2. Before/after boundary rule undefined (Grok C1 = Kimi CR-2 = GLM F4). Fixed and frozen now: boundary = first apply with ≥1 decision=promoted; same indices split all arms (cross-arm pairing); headline effect = difference-in-differences; <3-mistakes fallback carried from P0.6.
3. §4.5 clobber cite pointed at the wrong mechanism, no fail consequence, cadence contradiction (Grok M4 = Kimi MAJOR-3 = GLM F7). Fixed: probe re-cited to the model-side pending-file Write path; two jobs in one UTC day as a declared cadence exception; pass/fail + degraded-evidence consequence pre-registered.
4. Arm-B hook disable was not an executable recipe on user-global registration (Grok M1 = Kimi MINOR-3 = GLM F5). Fixed: user-global entries absent for the run; arm C gets hooks via workspace-local .claude/settings.json (snippet sealed at seal; wiring re-check must show fires-in-C/silent-in-B).
5. Contamination predicate unscoped/asymmetric (Grok M2 = Kimi MAJOR-2 = GLM F9). Fixed: scope = 4 learning-signal sections (Last session/handoffs excluded but reported); baseline = consecutive end-of-job snapshots; both arms same rule; source-trailer attribution in C.
6. Arm B is not "no channel to improve" (Kimi MAJOR-1 = GLM F8; Grok FM3). Fixed: §1 rewritten — B keeps the episodic channel (Last session/handoffs/residue) open by design and turns off only the lessons/rules channel; residue-access measured per job.

Adopted — 2-seat convergent (CRITICAL):
7. Compressed clock vs the k≥2 distinct-ISO-week apply gate: a same-week run promotes nothing by construction (Grok C2 = GLM F1; Kimi MINOR-4 adjacent). Fixed: §3 clock protocol — raw content must span ≥2 ISO weeks before apply-1 (≥3 for two independently-promotable turns); job→UTC-rollover→intake→review→apply ordering; window-coverage check; the run is a multi-week calendar run (schedule sealed). Wiring re-check gains "one real cross-week promotion end-to-end".

Adopted — single-seat, source-verified:
8. TR_SPAWN_STEP sealed as relative would kill every harness job at launch (Kimi MAJOR-4; I re-verified task-runner.sh:62-64 on the pin — CONFIRMED). Fixed to absolute; CLAUDE_MODEL=claude-haiku-4-5 added (Grok m1); per-job TMPDIR posture restored (Grok m3).
9. Review invocation was not sealed while apply was (GLM F6). Fixed: verbatim seal line + ordering rule; full review.conf assignment set incl. zero_streak_threshold=14 (Grok m2); review fail-close retry/skip rule (Grok m5).
10. Confirmatory hierarchy undecided between quote-NG "Primary" and repeat-mistake "headline" (GLM F10 = Grok C1-4). Fixed: single rule — repeat-mistake DiD confirmatory, NG slope co-primary, endpoints secondary, no-harm guard as computable predicate (Grok M5) with T/X sealed.
11. Attribution: promotions=0 ⇒ contrast cannot confirm "review, then promote" (Grok M3 = GLM F3). Fixed: 2×2 attribution rule + per-job source-trailer decomposition report; fold channel stays on (product-as-shipped) and is decomposed, not disabled.
12. Host-writer claims stale (Grok M6 = Kimi MINOR-2), arm-A persistence ambiguity (Kimi MINOR-1 = GLM F17), §8 completeness (GLM F13–F16, F18–F20). All fixed as worded in v0.3.

Escalated to owner (blocks seal; = the C3 second-confirmation packet):

• §1-A claim scope (Grok C3, CRITICAL — the pilots' effect came from class: rule promotions; the sealed autonomous invocation promotes capability-facts only): option A = autonomy-only with reframed claim; option B = pre-registered mechanical rule-approval step in arm C. Recommendation in draft: B.
• §3-R replicates (GLM F11): R1 = ≥2 replicate workspaces per harness arm; R2 = single series + index pairing + limitation note. Recommendation: R2.
• Pin retarget v0.19.0→v0.20.0 re-confirmation (pre-noted; all seats: correctly motivated, Grok m6).

Draft revised to v0.3 (all convergent fixes in; owner decisions marked inline). Next: light full-loop wiring re-check on the pin (scope per §8: cross-week promotion, hook C-on/B-off, source trailers, review ordering, v0.20.0 fact re-confirmation) → owner packet (§1-A, §3-R, pin, budget) → incorporate → one delta round with all three seats on the seal candidate (Grok's NO-GO must flip on evidence; L1-3 loop until blocking=0) → seal fills → owner GO → run.

[shojikumaru]

Light full-loop wiring re-check on the v0.20.0 pin — PASS (2026-08-27, C3 gate item).

Workspace experiments/ev007/ws-wire-020 (outside git), scripts invoked from the pin worktree exactly as sealed in prereg §8 (nightly raw-review.sh --workspace, no --week; APPLY_MAX_AUTO=10 apply-promotions.sh --auto-capability-facts). Corpus: the 6 real raw flush files (W29/W30 originals) re-dated onto 2026-W34/W35 so the shipped trailing-2-week nightly window covers them.

step	result
flush-intake	PASS — fold=30 lines all (source: flush-intake); measured: per-run fold budget defers a large backlog — 6 files needed 6 intake runs to fully archive (new §3 drain rule: repeat until files_scanned=0)
raw-review (nightly, Kimi K3 wrapper)	PASS — window=2026-W34,2026-W35 blocks=22 fabricated=2 rejected=2 candidates=20 error=none; fabrication guard exercised live
apply run 1	PASS — 10 promoted capability-facts, every one k=2 (weeks: W34,W35) = cross-week promotion end-to-end; class: rule themes → awaiting-approval (counted, not written); single-week themes → k-below-2; APPLY_MAX_AUTO=10 cap hit exactly
apply run 2	PASS — fully idempotent (already-applied / awaiting-approval only)
provenance	PASS — promoted lines carry full trailers (source: theme-… ; reviewer: kimi-k3; weeks: …; k=2; approver: auto) ⇒ the §5 fold-vs-promote decomposition report is mechanically computable
hook C-on	PASS — workspace-local .claude/settings.json Stop hook executes under headless claude -p (unconditional trace fired)
hook B-off	Mechanism issue found and resolved in the recipe: the host's personal fable-loop harness registers user-global Stop/PreCompact hooks that fire in any cwd containing STATE.md — both probe workspaces received its checkpoint demands (cross-arm interference from outside the product). The shared pause marker can't isolate B (it kills the task-runner too). Sealed recipe: path-exclusion wrapper on the personal global hook entries for the run's duration (exit 0 under experiments/ev007/, delegate otherwise; applied before job 1, removed after, ledgered). Live wrapper demo = remaining seal-time check
v0.20.0 fact re-confirmation	PASS — three review seats independently re-verified every operational claim at the pin; my spot checks agree

Prereg updated in place (still DRAFT v0.3): §3 intake drain rule; §2 hook recipe refined with the measured facts; §8 wiring line filled with the receipts above.

Seal queue state: all mechanical gate items are green. Remaining before seal: owner decisions §1-A (claim scope: A autonomy-only vs B pre-registered rule-approval — recommendation B) and §3-R (replicates: R1 vs R2 — recommendation R2), pin v0.19.0→v0.20.0 re-confirmation, budget cap → then fills → one delta round with all three seats on the seal candidate (Grok's NO-GO must flip on evidence) → owner GO → seal → run.

[shojikumaru]

C3 second-confirmation round — owner decisions recorded (2026-08-27).

Owner responses to the decision packet (conversation, 2026-08-27):

1. §1-A resolved as B, with a correction to my framing. Owner asked whether rule promotion wasn't already approved to proceed unmanned. Verified against docs/design/apply-step-201.md §4: the shipped contract is approver = Alpha (owner decisions 2026-08-26/27) — rule promotion never required in-run human judgment. Structure B is therefore within the shipped approval flow: the approver's policy is sealed in the prereg ("approve every rule theme passing review hygiene + recounted k≥2"), removing in-run discretion, and the run stays fully unmanned. Prereg §1-A rewritten accordingly (working decision B; final confirmation bundled into the seal GO).
2. §3-R = R2 (single series per arm + cross-arm index pairing + sealed limitation note).
3. Pin: owner asked to re-check tag freshness — correct call. Repo had advanced again: v0.21.0/v0.21.1 released (task-runner: sentinel ledger confluence + task-level task_end (#187) #206 task-runner sentinel-ledger confluence, repo-state: add generation stamp + status.json + caller workflow (family-os#127, pin v0.13.1) #209 repo-state). Owner chose v0.21.1. Executed: caty-agent-harness-ev-pin moved 0babfef → 74b9fbc9cf9f5889eeb2b3e6ea503c662f7d34d6 = git describe → v0.21.1 (exact tag). Measured: git diff v0.20.0..v0.21.1 over scripts/flush-intake.sh, raw-review.sh, apply-promotions.sh, adapters/ = 0 lines — the learning-loop layer is byte-identical, so today's wiring re-check evidence carries to this pin unchanged; the moved layer is task-runner (task-runner: sentinel ledger confluence + task-level task_end (#187) #206), symmetric across both harness arms. Prereg §8 pin line updated.
4. Run ops co-design in progress (owner request): allocate arms to the cswap multi-account rig by measured quota (avoid waste, avoid quota-caused job failures). Account inventory + per-account 5h/7d quota measured via cswap list; allocation + pre-burst probe rules + quota-failure semantics (infra-fail, never counted into metrics) will be sealed in §8 as an ops block.

Remaining before seal: series/calendar table (week-boundary burst design), thresholds/caps, ops block → 3-seat delta on the seal candidate (Grok flip required) → owner seal GO.

[shojikumaru]

Delta rounds complete — all three seats GO on the seal candidate (2026-08-27).

Same panel (Kimi K3 / Grok 4.6 / GLM 5.3; blind delta-1 on v0.4, targeted delta-2 for the two NO-GO seats on v0.4.3; requested = actual for all runs; outputs at the session scratchpad prereg-review/seat-*/delta*.md, verbatim copies ride the sealed record).

seat	delta-1 (v0.4)	delta-2 (v0.4.3)	own v0.2 blockers
Kimi K3	GO — all 8 RESOLVED + NEW-1 (calendar/clobber contradiction, must-fix)	not needed (NEW-1 fix verified by the other two seats' N1 re-check)	8/8 RESOLVED
Grok 4.6	NO-GO — N1 (= NEW-1) + N2 (§8 apply still A-shaped)	GO — N1/N2 RESOLVED with pin evidence; 2 residuals (cap truncation order, stale pin clause) → fixed in-place	C1/C2 RESOLVED, C3 resolved via §1-A=B seal, M1–M6 RESOLVED
GLM 5.3	NO-GO — N1 (same) + N4 (ops block disclosed real account handles into a public-repo doc) + N2/N3 minors	GO — N1/N4/N2/N3 all RESOLVED (mechanical re-check incl. UTC/JST arithmetic and pending-file path at the pin)	F1–F11 RESOLVED

Fixes folded into the candidate along the way (now v0.4.3+, all recorded in the draft changelog): clobber window → j6→j7 single-UTC-day pair; §8 apply = two verbatim commands (auto pass + --approve-file manifest from apply-index.tsv class=rule+awaiting-approval, cap 10/index, ascending theme-id); j6 = trap-instrument feeder for W36 quote citations; j3b extension calendared in all arms; ops block redacted to slot numbers (account handles = offline ledger only — public-repo hygiene, GLM N4); job-hour rule with corrected UTC/JST arithmetic; label re-stamps (§2/§3/§6/#162@v0.21.1); B=0 denominator clause; arm-C row aligned to the plan-F single loop day.

3-seat convergence note: N1 was found independently by all three seats — the convergent-finding-is-adopted principle held through the delta round as well.

Gate state: every mechanical + review gate is green. Remaining: fill the last §8 (seal) slots (hook snippet text, wrapper text, replay-record location, class-map table with the final job list) → owner seal GO (bundles: §1-A=B final confirmation, plan F, 120M cap, pin v0.21.1) → seal (immutable copy) → run starts Fri 2026-08-28 (before-block j1).

[shojikumaru]

🔏 SEALED — EV-007 main-run pre-registration (2026-08-27). Owner GO given; run starts Fri 2026-08-28.

• Owner GO: 2026-08-27, "GO（封印して開始）" — given after the full C3 second-confirmation packet (all-seats-GO review state, wiring PASS, plan F calendar, 120M charge-token cap, §1-A=B rule-approval structure, pin v0.21.1, account allocation).
• Sealed record: experiments/ev007/PREREGISTRATION-MAIN-SEALED-20260827.md (outside git per the declared file set; mode 444; 509 lines)
  sha256 = 4a49dee63d48f23bed2440a0de10d5568312f8490026a50b0de8ecc768f54932
• Pin: caty-agent-harness-ev-pin @ 74b9fbc9cf9f5889eeb2b3e6ea503c662f7d34d6 = v0.21.1 (exact tag, clean).
• Review provenance: r1 3-seat (Kimi/Grok/GLM) → v0.3 structural revision → delta-1 → v0.4.3 fixes → delta-2 all three seats GO (ledger: comment-5438487150). Seat outputs (r1/delta/delta-2, verbatim) archived beside the sealed file under experiments/ev007/tools/review-wiring/prereg-review-archive/ before j1.
• Done-when status after this seal: items 1–5 complete (draft ✓ / 3-seat review with blocking resolved ✓ / pin updated & recorded ✓ [v0.19.0→v0.20.0→v0.21.1, each move measured and owner-confirmed] / wiring re-check ✓ PASS / owner GO → sealed ✓). Remaining: item 6 — execute the run per the sealed prereg (j1 Fri 2026-08-28; pre-run gate = the sealed personal-loop wrapper live demo, result appended here before j1) and publish the report with the EV-008 correction-ledger discipline.

Next action (run operator): install the sealed path-exclusion wrapper → live demo (C fires / B silent / personal loop intact) → post demo result here → j1.