test: gate Greg and Boris evidence by observer trajectory

[safal207]

Summary

Adds an executable pre-notification evidence gate for the Greg Brockman and Boris Cherny open-work review.

The gate now requires two independent approvals from evidence, not from people:

1. space / transition / time / causality;
2. center of orientation / observer / trajectory continuity.

Coordinate model

N  = read-only evidence observer
Oᴳ = exact Greg source SHA + issue #3 + isolated Linux runtime + observation time
Oᴮ = exact Boris fork/upstream PRs + Claude issues + current GitHub lifecycle time
X  = object space
Y  = state transition
Z  = execution and authority context
T  = time
τ  = ordered evidence trajectory

The observer is outside all third-party repositories and runtimes. It may read public evidence, execute exact public source in isolated CI, and publish an evidence comment only after the gates pass. It may not become an owner, approver, maintainer, closer, merger, or production operator during the trajectory.

Greg evidence trajectory

The workflow checks out exactly:

gdb/tee-output@c41f8ff383200320b746e953e92709ae1b505a71

It runs 20 isolated baseline rounds and 10 counterfactual rounds. Each round writes 64 unique records directly to stdout and stderr file descriptors and calls close() immediately.

The counterfactual changes one variable only: shutdown order.

flush wrappers
→ restore fd 1/fd 2
→ close pipe writers to deliver EOF
→ wait naturally
→ terminate only on timeout

Allowed external classifications:

• baseline loss + all counterfactual rounds pass → CONFIRMED_SHUTDOWN_DATA_LOSS_WITH_PASSING_COUNTERFACTUAL;
• baseline and counterfactual both pass → NOT_REPRODUCED_ON_THIS_RUN_STATIC_RISK_REMAINS;
• incomplete or failed counterfactual → BLOCKED_DO_NOT_NOTIFY.

Boris evidence trajectory

The workflow refreshes through read-only GitHub API calls:

• fork PRs bcherny/openclaw#1, #2, #3;
• upstream PRs openclaw/openclaw#58036, #58037, #58038;
• correction commit b474e098d15d8a0936153118adb6e28255b9071e;
• Claude Code issues #21151, #4937, and #1554.

The fork-to-upstream coordinate-frame change is accepted only through an explicit identity edge: same author, same title, expected fork head, merged upstream PR, and exact merge SHA.

Notification gate

External comments remain blocked unless both results are true:

causal graph verdict starts with READY_TO_NOTIFY
observer trajectory verdict = OBSERVER_TRAJECTORY_ALIGNED_READY_TO_NOTIFY
blocking checks = 0

Even after passing, only evidence comments are permitted. Approval, closing, assignment, merge, labels, and third-party code changes remain prohibited.

Files

• .github/workflows/greg-boris-causal-evidence.yml
• scripts/greg_tee_output_evidence.py
• scripts/boris_open_work_evidence.py
• scripts/validate_greg_boris_causal_graph.py
• scripts/validate_observer_trajectory_center.py
• audits/reviews/greg-boris-causal-gate-config-v1.json
• docs/reviews/GREG_BORIS_CAUSAL_GATE_PROTOCOL_2026-07-20.md
• docs/reviews/GREG_BORIS_CAUSAL_GATE_RUNBOOK.md

This is a stacked draft PR on top of the read-only review PR. It does not modify any third-party repository or issue.

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 4d915581-28d9-433e-897f-29cb2c574cff

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch agent/greg-boris-causal-gate-run-v0-1

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[safal207]

Exact causal-gate result

The stacked evidence workflow completed successfully on exact PR head:

head SHA: 1b1675d48383c607eb2b26f737db93d79ea619fd
execution SHA: 6d5b56e48e54020ae741121693cb10348592a71c
run: 29773637522
attempt: 1
artifact: 8473778255
artifact digest: sha256:f36759b539384b444b427f04213e9f22c453520bf273f2b40824da3e653774bf

Greg result

baseline failures: 0/20
counterfactual failures: 0/10
allowed classification:
NOT_REPRODUCED_ON_THIS_RUN_STATIC_RISK_REMAINS

Each round wrote 64 unique records to each stream through direct fd writes and closed immediately. No records were missing from stdout, stderr, or the combined file. The exact source still has no explicit stdout/stderr flush and closes the writer before sending SIGINT to the reader process, so the original macOS/PTY/startup-race report is not disproved.

Result SHA-256:

6490732f329cc05dc5c06657c92676058b756134b6fc599e4d34c38bfda64736

Boris result

superseded fork PRs: 3/3
upstream correction verified: true
current assigned Claude Code issues: 3/3
verdict: VERIFIED_OPEN_WORK_LIFECYCLE

Result SHA-256:

2d7ac8578a10ed0c150148899d059b0e9b78d537cc9baa65f9a32fd6df733c94

Gates

causal verdict:
READY_TO_NOTIFY_BORIS_GREG_NONREPRO_ONLY

observer trajectory verdict:
OBSERVER_TRAJECTORY_ALIGNED_READY_TO_NOTIFY

blocking checks: 0

Causal result SHA-256:

b2c2857ba08a5125b220f63a17207f575da8f5c7cdae2611ead2c2c71b2bcab2

Observer trajectory result SHA-256:

2cd8ae207a795185b2a7df8d96704462998b587b41eab9f8af450b120a1bd085

The gate authorizes evidence comments only. It grants no authority to approve, close, label, assign, merge, or modify third-party code or issue state.

[safal207]

Evidence delivery

Direct comment creation in gdb/tee-output#3 was attempted after both gates passed, but GitHub returned:

403 Resource not accessible by integration

No attempt was made to bypass third-party repository permissions.

Safe public delivery was completed through mention-addressed evidence issues in this repository:

• Greg: Evidence for @gdb — tee-output #3 bounded reproduction #98 (@gdb)
• Boris: Evidence for @bcherny — openclaw PR lifecycle and Claude Code issue review #99 (@bcherny)

Both issues preserve the exact run, artifact ID/digest, result hashes, coordinate center, allowed classifications, limitations, and links to the original third-party work. No third-party state was modified.