#381 pre-run record: Amendment 3 merge checks and recovered inputs; #462 uninstall note

[seathatflowsinourveins]

Summary

This is a dated pre-run record for the #381 token-adoption E2E (Gate A), written by the Gate A owner. It is not an amendment: no task text, information need, check, arm or route changes, and the five sealed files are untouched. No capability probe or organic arm has run.

• Amendment 3 merge. Skills trial: remove verification-before-completion under the conflict rule (M4), #381 Amendment 3 #464 merged at c0966da2 (2026-09-28T18:29:54Z), which is the preregistration_commit checked at launch. At that revision:
  • the frozen table pointer re-serializes to the sealed frozen_input (6899e551…, 6,552 B, 60 records);
  • all three copies of the five role bodies equal their pinned rows;
  • the preregistration tests pass.
    The record also lists the six token-lanes block SHA256s for the launch's carrier check.
• Command identities (reuse-296-00, reuse-343-13). The receipts had truncated the six original observation commands. They are recovered from both source runs' surviving original work files, and each of those files' SHA256 equals its receipt's exact_comparison digest. The six are git log -30, git status, git diff HEAD~5 --stat, grep -rn "def register_file" scripts, ls -la scripts and python3 -m unittest tests.test_host_requests. The record keeps the bare command (the identity) separate from each run's channel (provenance).
• History originals (reuse-296-02, reuse-343-06). git --no-pager log --stat -150 <rev> at each receipt's own catalog_revision reproduces its baseline_sha256 exactly. The match was found by hashing that output for 597 candidate commits. The reports stay private because they carry author identities; the record keeps only the revision, the command and the digest.
• Delegated decisions 2026-09-28: M1/M2 keep, M3 the latest Opus, M4 remove, #444 trading ack, Gate A owner #462 note. The delegated-decisions record gains a dated note on the uninstall command. The unscoped skills remove verification-before-completion -g -y is the complete form, with a precondition that no other agent holds a same-named copy. An -a-scoped removal keeps the canonical folder and lock entry: skills@1.7.0 dist/cli.mjs removeCommand L6834-6838 and L6883-6907, plus getAgentBaseDir L2214. The skills-trial owner found this during the host uninstall.

SOTA sources

• Pre-run record: the protocol's own gates at c0966da2, in evidence/artifacts/token-adoption-e2e-20260926/:
  • preregistration.json pre_run_input_gate of reuse-296-00 and reuse-343-13;
  • README.md:86-90;
  • RUNBOOK.md:30-35 (command identities) and :243 (input_paths);
  • Amendment 3's merge-chronology instruction, README.md:753 ("Record its merge revision and time before execution … re-verify it at the merge revision").
• Recovered values: the source receipts evidence/artifacts/token-e2e-ultracode-20260925/receipt.json (/tools/0, /tools/2) and evidence/artifacts/token-e2e-codex-20260926/receipt.json (/tools/6, /tools/13). Each recovered value is matched to their exact_comparison SHA256.
• History regeneration: upstream git 2.43.0 git log --stat, run at each receipt's catalog_revision.
• Delegated decisions 2026-09-28: M1/M2 keep, M3 the latest Opus, M4 remove, #444 trading ack, Gate A owner #462 note: vercel-labs/skills@v1.7.0, pinned in adoption/skills/manifest.json as npm skills-1.7.0.tgz, sha256 8d1466f792baaee945dae88e05ee403d6f9e78a3ae8dcbf61496035ca274418d. The tarball was re-downloaded and matched. Its package/dist/cli.mjs cites:
  • removeCommand L6834-6838 and L6883-6907;
  • isUniversalAgent L2180;
  • getAgentBaseDir L2214.
• Repository rules:
  • docs/lanes.md:94 (hot-file protocol);
  • docs/acceptance-evidence-policy.md (these are recorded integration checks and recovered historical bytes, not a new run or upstream acceptance).

Verification

• At c0966da2, in a detached worktree, python3 -B -m unittest tests.test_token_e2e_preregistration ran 34 tests: OK.
• At the branch head, in a fresh TMPDIR, python3 -B -m unittest tests.test_token_e2e_preregistration tests.test_token_lanes_subagent_start tests.test_install_claude_profile ran 109 tests: OK (skipped=3).
• Hot-file protocol: the evidence.json edit is in the last commit, via register_file, then component_matrix.py --write and new_host_grand_list.py --write. scripts/validate.py exited 0 (status: passed, 164 receipts).
• Cross-family review of head 208ffb5d: codex exec -s read-only -m gpt-6-astra -c model_reasoning_effort=max, told to refute each added claim from source. It ran 16 ctx_execute recomputations covering the receipts, the retained originals, the --stat -150 regeneration, frozen_input, the block and role-body hashes, cli.mjs and the gate text. It returned NO FINDINGS.
• The diff was scanned for personal paths, session ids and host names (0 matches), and the pre-commit gitleaks scan found no leaks.

Still open before launch

• PR-A's residual measurement scope (#381 PR-A: token-adoption measurement kernel (M3/M4/M5, RTK eligibility, MCP attempt states) #432's Residuals) and its regenerated sanitized baseline, both merge-before-run. This session now owns them.
• The Codex half of the collector join.
• A sealed quiet window W.

🤖 Generated with Claude Code