Receipt: Harbor token-tools E2E (288 trials), figures and hashes of the private originals

[seathatflowsinourveins]

Scope

• Adds one sanitized receipt, evidence/receipts/harbor-e2e-token-tools-20260930.json (kind native_model_e2e), for the Harbor token-tools E2E of 2026-09-29/30, and registers it in manifests/evidence.json (last commit): the hash row in files[] and the entry in the canonical receipts[]. The run's figures and the hashes of its private originals were recorded only in private host state; the foundation record and the new-WSL architecture rows cite this receipt instead.
• Second revision: the Codex review bot left ten threads on the first one (three P1, seven P2). Each is answered on the thread with the field of the receipt that changed, and resolved.
• Base commit: 78881a96 (origin/main when written). The hot-file re-run onto the main of its slot has to carry both the files[] row and the receipts[] entry.
• Lane: lane:foundation
• Owned paths touched: evidence/receipts/harbor-e2e-token-tools-20260930.json, manifests/evidence.json

SOTA sources

• Harbor v0.23.0, the upstream harness the run used with its built-in claude_code agent: https://github.com/harbor-framework/harbor/releases/tag/v0.23.0 (PyPI harbor 0.23.0, uploaded 2026-09-12; read 2026-10-01).
• Task source: SWE-bench Verified, https://huggingface.co/datasets/SWE-bench/SWE-bench_Verified at revision 78f471bf (36 of 36 tasks field-identical and oracle-verified, per the run's provenance check).
• Statistics: scipy 1.18.1, statsmodels 0.15.0 and numpy 2.5.3 (the frozen versions). The cost comparison is a paired geometric-mean ratio with a BCa bootstrap interval and a Holm-adjusted permutation test; the resolve-rate procedure is documented in the receipt (resolve_rate_procedure).
• Docker documentation for the rootless subordinate-id observation: https://docs.docker.com/engine/security/rootless/ and its troubleshooting page (read 2026-10-01).
• The tools under test, each installed as its maintainers document, pinned in run.tools_under_test: rtk-ai/rtk v0.50.0, mksglu/context-mode v1.0.169, oraios/serena (serena-agent 1.7.0), jgravelle/jcodemunch-mcp 1.108.319, chopratejas/headroom (headroom-ai 0.39.1).
• The top rule (AGENTS.md, rule (e), merged in Rule text in every layer: the six standing clauses on AGENTS.md, the portable Claude template and the Codex AGENTS template, anti-pattern rows and the folded finalization record (unit F1) #557) says A/B and E2E use upstream harnesses; this run is the one such E2E already executed on this stack. Nothing here is written from memory: every figure comes from the run's own analysis output, which the preregistered script reproduces byte for byte from the preserved trial outputs (reproduction).
• No upstream tool produces this summary; it is a receipt of our run. The arm wrappers, the analysis script and the scripts that computed the new blocks are local integration code; the last are kept privately with their hashes under originals.

Evidence-class table

Claim	Evidence class	Command / receipt
Per-arm cost ratios, resolve rates and tool use over 288 trials	native_proven (actual execution on this host, 2026-09-30; summary receipt, raw logs private)	results
The cost A/A control (lean-r2 against lean, 1.045 [0.93, 1.20])	native_proven, reproduced by an independent audit	control
The stored analysis is reproduced from the preserved outputs	local_integration (the preregistered script re-run, byte-identical output)	reproduction, trial_output_manifest
308 stream logs = 288 analysed trials + 20 smoke logs; no retries, interruptions or exclusions	local_integration (a scan of the saved run records)	log_accounting
0 of 288 trials started before the freeze at 06:15:44Z; the 17 frozen inputs are unchanged	local_integration (trial start times against the freeze record; inputs re-hashed)	schedule, run.preregistration
Seeded arm order (72 of 72 jobs), blocked by task; usage categories per arm	local_integration (job configs, start times and trial rows)	schedule, token_categories_per_arm
No trial spawned a subagent (308 stream logs, 308 client transcripts, 0 delegation calls, 0 sidechain lines) with positive controls	local_integration (text scans; a live capture and persisted transcripts as controls)	subagent_scan
The token-efficiency layer's own overturn-protocol metric for four of its five arms	local_integration (recomputed by the Gate A owner; not independently reproduced)	catalog_protocol_metric
Resolve-rate procedure, paired tables and a sensitivity run	local_integration (scipy and statsmodels, a fresh generator seed)	resolve_rate_procedure
Rootless Docker needed more than 65,536 subordinate ids for the matplotlib images	native_proven (run records and a read of the current files; Docker documentation read 2026-10-01)	host_prerequisite_observations
The originals are recoverable	local_integration (sha256 and size of 14 private files)	originals

Local commands run

$ python3 -B scripts/validate.py
exit 0; {"components": 69, "hashed_files": 8703, "profiles": 4, "receipts": 180, "status": "passed"}
$ python3 -B scripts/validate.py --scan-file evidence/receipts/harbor-e2e-token-tools-20260930.json
exit 0; {"scanned_files": 1, "status": "passed"}
$ python3 -B scripts/evidence_manifest.py --check
exit 0; {"files": 8703, "status": "passed"}
$ python3 -B scripts/host_receipts.py validate
exit 0; {"receipts": 182, "status": "passed"}
$ git diff --check 78881a96..HEAD
exit 0
$ python3 -B -m unittest tests.test_evidence_manifest tests.test_validate tests.test_ecosystem_manifest (and 14 more modules that read the evidence manifest)
exit 0; Ran 1250 tests in 143.579s, OK (skipped=3)

Decision record

None for this receipt. It is the evidence behind the provisional verdict of the token-efficiency layer in the new-WSL architecture rows: the 14-component profile of #540 stays the install default, savings are "not shown", and the multi-agent adoption E2E decides.

Host evidence

• python3 scripts/host_receipts.py validate passes.
• Independent review: the ten review-bot threads are answered. The cost figures were reproduced by an independent Opus audit on 2026-10-01; the blocks added in this revision (schedule, log accounting, subagent controls, resolve-rate sensitivity, token categories) were computed by the Gate A owner and have had no second reader.
• No platform_status change is made from this receipt.

Checklist

• No GitHub Actions or workflows are changed.
• No secrets are printed, logged or committed: the receipt holds no host path, token, trial or session identifier; validate.py --scan-file and gitleaks found nothing.
• No new paid hosting, subscription or billing surface.
• Peer-owned untracked files and worktrees were preserved.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-01T05:30:58.619210Z	20d6c19	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 20d6c195e9

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".