Delegated decisions 2026-09-28: M1/M2 keep, M3 the latest Opus, M4 remove, #444 trading ack, Gate A owner - #462

Scope

• What this PR changes:
  • It records five decisions the user delegated on 2026-09-28 ("you decide with evidance manfiest . research convergence and consensus with peer sessions pls").
  • It retains the returned lane output behind them.
  • It aligns the portable instruction copy with the M3 decision.
• The decisions:
  • M1/M2: keep the top rule verbatim.
  • M3: "The latest Opus at effort max …", in ~/.claude/CLAUDE.md:40 on the host and examples/claude-native/CLAUDE.md:41 here.
  • M4: remove the verification-before-completion trial skill under the skills trial's in-force rule (2026-09-25-skills-trial-and-usage.md:284-286). The removal itself lands in the trial owner's PR together with PR-H: E2E preregistration for the token-save-practice stack (frozen tasks, metrics, gate, procedure) #381 Amendment 3.
  • Prompt audit rounds 2-3: AGENTS.md F1, F4, S1 and X5b; X9 applied to CLAUDE.md, 4 of 4 blind judgments; dashboard manifest bound #444: the live trading-lane session gives the acknowledgement. It was posted 2026-09-28T16:04:57Z.
  • Gate A (PR-H: E2E preregistration for the token-save-practice stack (frozen tasks, metrics, gate, procedure) #381): this session owns it, with a public plan.
• Base commit: 8d8f79cd (rebased after AN-13 M4/M5: two-family lanes; M4 kept, M5 split after blind adjudication; judge-dispatch anti-pattern row #460 merged)
• Lane: lane:foundation. manifests/evidence.json only registers files, in the last commit.
• Owned paths touched:
  • docs/decisions/2026-09-28-delegated-decisions.md
  • evidence/artifacts/delegated-decisions-20260928/
  • examples/claude-native/CLAUDE.md (one line)
  • manifests/evidence.json

SOTA sources

• M3:
  • Anthropic prompt-audit rubric, anthropics/skills@33375500bcea98d610eb30ce10ac4e59b89c390d:skills/claude-api/shared/prompt-audit.md:127-141: "pinned model names silently degrade after the next release", "State the current rule".
  • model-config: what the aliases resolve to, and that this depends on the provider.
• M1/M2:
  • Prompting Claude Opus 5.5: its sample begins "Before taking any action, explore broadly…".
  • Claude Code best practices: review CLAUDE.md "when things go wrong" and test changes by observing behaviour.
• M4: obra/superpowers@8ca22dba, skills/verification-before-completion/SKILL.md (sha256 2befe7fc…). Upstream has not changed that path since, so there is nothing to re-pin to.
• Repository rules:
  • docs/lanes.md:146-160;
  • docs/decisions/2026-09-25-model-fallback-guard.md:68-70;
  • docs/decisions/2026-09-26-harness-rules-cleanup.md:15-16;
  • evidence/artifacts/token-adoption-e2e-20260926/RUNBOOK.md:8-50.

Evidence-class table

Local commands run

$ python3 scripts/validate.py
exit 0 {"components": 69, "hashed_files": 7738, "profiles": 4, "receipts": 163, "status": "passed"}  (at 8d8f79cd + this branch)
$ TMPDIR=<fresh> python3 -m unittest tests.test_install_claude_profile tests.test_grand_dashboard tests.test_adoption_docs_consistency <the three pre-push registry tests>
Ran 115 tests, OK (skipped=4); the three registry tests each "ok" (zizmor on PATH)

Decision record

docs/decisions/2026-09-28-delegated-decisions.md

Review

1. Claude workflow review. Workflow wf_811a77e9-a4e ran five researchers and five refuters. All five recommendations came back upheld: false with amendments, and the record applies them:
   • M1/M2 moved from "narrow" to "keep";
   • M3 moved from bare "Opus" to "the latest Opus";
   • M4 gained the test of the trial rule;
   • Gate A gained the ownership and prerequisite checks.
2. Second family on the instruction items. codex exec -s read-only -m gpt-6-astra -c model_reasoning_effort=max on frozen packets for M4, M1/M2 and M3.
3. Peer objection round. No objections from resolve-prompt-audit-findings, cloudflare-security-audit-sota-review or token-save-practice-e2e-status. native-agent-stack-84 posted the Prompt audit rounds 2-3: AGENTS.md F1, F4, S1 and X5b; X9 applied to CLAUDE.md, 4 of 4 blind judgments; dashboard manifest bound #444 acknowledgement.
4. Cross-family PR review. codex exec -s read-only -m gpt-6-astra -c model_reasoning_effort=max on head 8a1b860b returned "VERDICT: not ready" with 2 major and 2 minor findings. All four are repaired in 8fa26559:
   • the peer positions, the Prompt audit rounds 2-3: AGENTS.md F1, F4, S1 and X5b; X9 applied to CLAUDE.md, 4 of 4 blind judgments; dashboard manifest bound #444 acknowledgement permalink and the Gate A liveness observations are now retained in coordination.md;
   • user-level line 15-16 fragments are redacted from m12/claude-return.md;
   • original_* and published_* digests are split;
   • the fallback-guard citations are fixed (:119-120, :127-129).
     The branch was also rebased onto 8d8f79cd, and the record now cites AN-13 M4/M5: two-family lanes; M4 kept, M5 split after blind adjudication; judge-dispatch anti-pattern row #460's merged overturn clause (2026-09-28-an13-m4-m5.md:240-242).
5. Codex re-check on d58e69fd (the same command). Findings 2-4 were resolved; the recomputed published_* digests and all 13 manifest entries match. Finding 1 was downgraded to minor and unresolved: the ledger claim went beyond what was searched. It was narrowed in dd371284 to "outside directories were not searched". The loop is bounded (one review, one repair, one re-check), so there is no further Codex round, and the narrowing is recorded here as the residual.

Checklist

• New/changed GitHub Actions are pinned to a full commit SHA with a version comment (no floating tags). (none changed)
• New/changed workflows declare top-level permissions: contents: read (or a narrower, explicitly justified addition). (none changed)
• No secrets are printed, logged or committed; no new required secret was added without a documented owner.
• No new paid hosting, subscription or billing surface was introduced.
• Peer-owned untracked files and worktrees were preserved (not deleted, moved or overwritten).

🤖 Generated with Claude Code