Independent testing will determine whether MARGINAL reliably detects unproductive coding-agent loops and intervenes only when warranted without disrupting legitimate progress.

state: expiredheat: lowuncertainty: highconvergesscott: mediumagent-harnesses coding-agents agent-governanceMARGINAL

What is this?

MARGINAL is presented as an open-source runtime governor for AI coding agents, with a contributor call aimed at Claude Code users and repository evidence linking its social card to a Reddit image. Its proposed role is to detect unproductive agent loops and intervene selectively. However, the supplied search snippets discuss general loop controls, independent checks, stopping rules, and semantic drift; they do not identify MARGINAL’s creators or substantiate the claim that independent testing has already confirmed its reliability.

Why it matters to Scott

MARGINAL independently implements Scott’s existing position that long-running coding agents need an external supervisor or closer capable of distinguishing real progress from stalls and intervening through governed harness controls. It is a direct evaluation and dated-receipts opportunity for his runtime-governance and long-running-agent work, but relevance remains medium until independent tests establish intervention precision and an acceptable false-positive rate.

ip:source.handover-notes-for-robots-ebookip:framework.five-surface-loop-anatomyip:concept.runtime-governanceip:concept.evaluation-driven-developmentdev:concept.deterministic-agent-control-planedev:concept.resumable-agent-job-control-planeradar:flow-claude-code-supervisorradar:opencode-guardians-tool-call-verificationradar:concept.agent-governanceradar:concept.agent-reliabilityradar:concept.agent-harnesses

queries asked of Scott's wikis

runtime governance for coding-agent loops
detecting agent stalls versus legitimate progress
independent evaluators and stopping rules for agents
coding-agent harness intervention policies
semantic drift detection in autonomous coding
Claude Code runtime hooks and supervision

Measured heat

no measured readings yet — the hourly heat pass fills this in

How the heat travelled

no chain yet — the hourly chain pass fills this in

Evidence (3) — ⭐ canonical anchor

source	object	author	score	comments
🟠 reddit	Looking for Claude Code contributors 🙏🏽 Open-Source runtime governor for AI coding agents ClaudeAI	Positive-Captain-709	1	9
🟧 echo.github ⭐	The Reddit image is pixel-identical to the repository’s `site/assets/marginal-social-card.png`, introduced by this commit. The commit refram	SignalLayer Labs	—	—
🟠 reddit	Your agent doesn't crash when it goes off the rails. It just keeps billing you[p] MachineLearning	No-Program-5087	0	1

Interpretation history

2026-08-25T09:33:25Z
No independent benchmark, adopter report, or measured intervention result has emerged, and the MARGINAL-specific episode has gone dormant. The broader runtime-governance pattern remains relevant, but this implementation has not advanced its reliability claim.
2026-08-23T08:35:27Z
The latest activity is repetitive amplification without an independent benchmark, adopter report, or measured intervention result. MARGINAL remains an unvalidated implementation of a relevant governance pattern, with no change to the core reliability question.
2026-08-21T07:24:35Z
A separate drift-detector implementation strengthens the broader case for runtime detection of unproductive agent behavior, moving the pattern beyond MARGINAL alone. It does not test MARGINAL or establish detection accuracy, false-positive rates, or safe intervention behavior, so the core reliability hypothesis remains open.
2026-08-21T07:22:43Z
evidence attached: reddit.post.1vu96ci — An independent open-source drift detector supports the same developing episode around detecting and interrupting unproductive agent loops.
2026-08-20T13:27:10Z
The refreshed discussion reinforces the evaluation gap: commenters recognize loop detection as useful but explicitly question the lack of credible benchmarks. No independent test, adopter report, or intervention result changes the creator-asserted status of MARGINAL’s reliability claims.
2026-08-20T10:34:29Z
The reobservation adds no independent testing or implementation results, so MARGINAL remains a promising but creator-asserted runtime-governance experiment. Reliability, intervention precision, and false-positive risk are still unestablished.
2026-08-20T10:32:57Z
grounded: converges/medium — MARGINAL independently implements Scott’s existing position that long-running coding agents need an external supervisor or closer capable of distinguishing real
2026-08-20T10:30:41Z
origin walked (codex/luna, conf 0.99): anchor reddit.post.1vtehus -> echo.github.1675c97079 by SignalLayer Labs
2026-08-20T10:28:31Z
case created — The announced open-source governor is a concrete control-layer experiment spanning Codex, Claude Code, OpenCode, and PrivacyCode.

Decision trace

08-25 19:33expireNo independent benchmark, adopter report, or measured intervention result has emerged, and the MARGINAL-specific episode has gone dormant. The broader runtime-governance pattern remains relevant, but
08-25 19:33alert_silentThe staleness trigger carries no consequential new evidence; Scott can wait unless independent testing establishes detection accuracy, false-positive rates, or safe intervention behavior.
08-25 19:33alert_routeThe staleness trigger carries no consequential new evidence; Scott can wait unless independent testing establishes detection accuracy, false-positive rates, or safe intervention behavior.
08-23 18:35repriceThe latest activity is repetitive amplification without an independent benchmark, adopter report, or measured intervention result. MARGINAL remains an unvalidated implementation of a relevant governan
08-23 18:35alert_silentNo consequential new fact occurred; minor discussion activity does not establish detection accuracy, false-positive rates, or safe intervention behavior and can wait for independent testing.
08-23 18:35alert_routeNo consequential new fact occurred; minor discussion activity does not establish detection accuracy, false-positive rates, or safe intervention behavior and can wait for independent testing.
08-21 17:24repriceA separate drift-detector implementation strengthens the broader case for runtime detection of unproductive agent behavior, moving the pattern beyond MARGINAL alone. It does not test MARGINAL or estab
08-21 17:24alert_silentThe new evidence is an adjacent self-promoted implementation without a verifiable artifact or evaluation results; it can wait until an independent benchmark, repository evidence, or adopter report mea
08-21 17:24alert_routeThe new evidence is an adjacent self-promoted implementation without a verifiable artifact or evaluation results; it can wait until an independent benchmark, repository evidence, or adopter report mea
08-21 17:23alert_silentA lone Reddit self-promotion describes a potentially relevant self-history-based drift detector, but provides no linked repository, release artifact, benchmarks, or evidence of intervention precision.
08-21 17:23surface_candidateA lone Reddit self-promotion describes a potentially relevant self-history-based drift detector, but provides no linked repository, release artifact, benchmarks, or evidence of intervention precision.
08-21 17:23alert_routeA lone Reddit self-promotion describes a potentially relevant self-history-based drift detector, but provides no linked repository, release artifact, benchmarks, or evidence of intervention precision.
08-21 17:22attachAn independent open-source drift detector supports the same developing episode around detecting and interrupting unproductive agent loops.
08-21 17:22propose_attachAn independent open-source drift detector supports the same developing episode around detecting and interrupting unproductive agent loops.
08-20 23:27repriceThe refreshed discussion reinforces the evaluation gap: commenters recognize loop detection as useful but explicitly question the lack of credible benchmarks. No independent test, adopter report, or i
08-20 23:27alert_silentThe new comments add skepticism rather than evidence and do not establish any capability, access, or implementation change. This can wait until an independent benchmark or user report measures detecti
08-20 23:27alert_routeThe new comments add skepticism rather than evidence and do not establish any capability, access, or implementation change. This can wait until an independent benchmark or user report measures detecti
08-20 23:21sensor_dirtycomment_update
08-20 20:34repriceThe reobservation adds no independent testing or implementation results, so MARGINAL remains a promising but creator-asserted runtime-governance experiment. Reliability, intervention precision, and fa
08-20 20:34alert_silentNothing consequential changed: engagement is unchanged and no independent evaluation, adopter report, benchmark, or enforcement result appeared. This can wait for evidence that tests MARGINAL’s core c
08-20 20:34alert_routeNothing consequential changed: engagement is unchanged and no independent evaluation, adopter report, benchmark, or enforcement result appeared. This can wait for evidence that tests MARGINAL’s core c
08-20 20:33alert_silentMARGINAL is a real open-source implementation of an external governor for coding agents and directly matches Scott’s runtime-governance thesis, but the new evidence is primarily creator-authored posit
08-20 20:33surface_candidateMARGINAL is a real open-source implementation of an external governor for coding agents and directly matches Scott’s runtime-governance thesis, but the new evidence is primarily creator-authored posit
08-20 20:33alert_routeMARGINAL is a real open-source implementation of an external governor for coding agents and directly matches Scott’s runtime-governance thesis, but the new evidence is primarily creator-authored posit
08-20 20:32groundMARGINAL independently implements Scott’s existing position that long-running coding agents need an external supervisor or closer capable of distinguishing real progress from stalls and intervening th
08-20 20:30promote_anchororigin walk conf 0.99
08-20 20:28createThe announced open-source governor is a concrete control-layer experiment spanning Codex, Claude Code, OpenCode, and PrivacyCode.