Dated system review →1 October 2026 · 10:13–10:14 KST
Verification standard · 2 October 2026 at 20:45 KST

A claim is only as strong
as its weakest observation.

We improve the system by reproducing a wrong result, repairing its actual code path and testing whether a broken implementation is rejected. Every result keeps its source, population and limits.

Specific local behavior verifiedOperating effect unverified
Evidence chain

Observation → counterexample → repair → independent review.

  1. 01

    Preserve the original failure

  2. 02

    Exercise the actual code path

  3. 03

    Reject bad implementations

  4. 04

    Review the same source bytes

Publication, installed code and natural operating effect require their own observations. Passing this chain for one change does not certify the whole system.

Latest scoped implementation

The test observer must preserve uncertainty.

The latest local repair strengthens failure cleanup and output interpretation in the test probe. Unavailable captured output remains unverified, even when the process reports a successful exit.

Local test population
18 / 18
0 failed · 0 errors · 0 skipped
Bad implementations rejected
6
Controls check that verification can fail.
Review level
Separate internal review
External attestation remains unverified.

Cleanup failure stays unresolved

Before
An unsuccessful Windows tree-termination command could return without attempting the direct-child fallback.
After
A nonzero tree-termination result attempts the direct-child fallback. Cleanup failure remains an unknown outcome.

Controlled dependencies exercise the termination branches. Actual descendant-process reaping is not established.

Timeout cleanup has a finite observation window

Before
An implicit process context exit or an unbounded output read could keep the probe waiting after the execution deadline.
After
The probe retains its execution deadline and limits the post-failure output read to five seconds. Unresolved cleanup returns a timeout outcome.

This bounds the tested cleanup read. It does not prove an absolute wall-clock bound for every operating-system interaction.

Unreadable output cannot become a pass

Before
Locale-dependent text capture could lose UTF-8 output or interrupt attribution; unavailable output could also fail attribution.
After
Binary capture is decoded strictly as UTF-8. Unavailable captured output or a UTF-8 decoding failure produces an explicit unverified outcome even with a zero exit code.

One isolated real child exercises UTF-8 output across a forced legacy text-decoding boundary. The other cleanup counterexamples use controlled process dependencies. This verifies capture availability and decoding, not output completeness; valid empty output retains the process exit-code result.

Dated result and source pins →SHA-256 →Earlier improvements →
Reproduction and access limits

The local result is tied to private source and test bytes. Public JSON and SHA-256 identify this disclosure; they do not independently reproduce or attest to the private execution. The website claim checks can be run from the website repository.

mxroach/neuropublic · kill_tree / run_pytest_once
Commit d6e06a514f5bf07ad7d1d3205975c87a9d611cd9
Source SHA-256 75f6296470174813368f0c95b0bee13b2d15eec4f7c474a8de9fff0bfd15cf91

Earlier disclosures retain their own dates, populations and publication stages. This local test result does not refresh fleet observations or resolve the held full-regression candidate.

Acceptance criteria

What would justify a stronger claim.

Operating effect

Not verified

The installed revision, a natural-run observation from the affected component, declared coverage and a comparison window.

End-to-end decision evidence

Not verified

Research, authorization, execution and accounting records sharing consistent identifiers; missing records retained in the denominators.

Sustained net performance

Not verified

Realized fills reconciled with fees, funding and external flows, matched by account, strategy and time over a declared evaluation window.

External reproducibility

Not verified

A sanitized runnable reproduction bundle and an observer able to check the original evidence independently.

These criteria describe evidence still required. This disclosure supplies no live account results, sustained net-performance claim or external certificate.