Mian Zhang

Public Release - 2026-06-08

An agent should stop when evidence cannot carry action.

Artifact: Agent Stop-Rule Evidence Check v0. The public question is simple: when public route, reproduction route, hash, DOI/status, freshness, or boundary fields are missing or stale, an AI agent should not receive action permission.

public_route repro sha256 doi_status stop_condition
Registries Evidence Map Challenge Route Boundaries
Technical stop-rule visual showing an AI agent execution pathway blocked by a safety governance gate.

Stop Rule

Fast execution is not permission.

Many agent failures are not caused by a model failing to answer. They happen when a system treats a fluent, plausible, or convenient answer as permission to act. A public stop-rule makes that permission conditional.

Today turns the upload-matrix context around P22 v0.2, W0, P23-P45, F1/P8, and D3 into an inspectable acceptance layer: if the public record cannot carry the claim, the claim is downgraded and action remains blocked or under review.

1

Public route

Every claim that asks for trust needs a public route: DOI, page, registry row, issue route, demo, manifest, or boundary note.

2

Reproduction route

If reproduction is unavailable, pending, or outside the public boundary, the row should say review instead of pass.

3

Hash and DOI/status

Hash mismatch, DOI ambiguity, stale status, or replaced artifacts are stop conditions until repaired in public.

4

Claim boundary

A strong claim needs a negative boundary: what it does not prove, what would downgrade it, and what would withdraw it.

5

Freshness

last_verified and stale_reason stop old evidence from silently supporting a stronger current claim.

6

Action permission

The output is not hype. It is pass, review, stop, or no_credit, with a public reason attached.

Ledger-style evidence check table with link, reproduction, hash, and status fields.

Evidence Check

Link / Repro / Hash is a no-go surface.

The check table is deliberately small. It asks whether the reader can find the public object, whether there is a reproduction or review route, whether the hash or DOI/status still matches, and whether the claim has a stop condition.

Passing this check does not prove deployment, real-robot operation, trading validity, medical validity, or general safety. It only says the public record is structured enough to inspect.

Matrix

Minimum stop-rule fields.

FieldRequired valueStop conditionRepair route
public_routeDOI, registry row, static page, demo, or issue route.Broken or missing route.Add route or mark not public.
reproduction_routeReplay, review, issue, demo, or bounded supplement path.No route, private-only route, or unverifiable route.Downgrade to review.
sha256Public file hash or explicit hash_status.Mismatch, missing hash, or replaced file.Publish corrected status.
doi_statusPublished, concept DOI, pending, stale, replaced, or not_applicable.Ambiguous or stale DOI/status.Add freshness note.
claim_boundarydoes_not_claim, downgrade_trigger, withdraw_condition.Claim has no negative boundary.Split or narrow the claim.
action_permission_statuspass, review, stop, or no_credit.Evidence gap is framed as permission.Block action and rewrite.

Asset Groups

What the check applies to.

The first public pass focuses on P22 v0.2, W0, F1/P8, D3, and the current P23-P45 upload matrix. The check does not add new empirical claims. It adds public fields that make existing claims easier to inspect, attack, downgrade, or repair.

If support material is not public, if a TeX/source bundle is absent from a public publication record, if a link is broken, or if hash/status fields disagree, the correct state is review or stop, not silent trust.

Stop-condition ledger visual showing pass, review, and stop states for public evidence fields.

Stop Conditions

The useful output is a blocked action with a reason.

A stop condition should be specific enough to repair: broken_link, missing_repro, hash_mismatch, doi_status_stale, missing_boundary, stale_evidence, or evidence_gap_used_as_permission.

This is why blocked action is not failure theater. It is accountable engineering work: the public record says why action is blocked and which public field must change before the row can move.

Failure Modes

How a row should fail.

FailureMeaningPermissionPublic repair
broken_linkThe public object cannot be reached.stopFix route or remove claim support.
missing_reproNo public reproduction, review, or counterexample path exists.reviewAdd bounded route or mark unavailable.
hash_mismatchThe artifact no longer matches its recorded digest.stopPublish corrected digest and version note.
stale_evidenceOld evidence is still carrying a stronger current claim.reviewAdd last_verified and stale_reason.
missing_boundaryThe claim lacks a negative boundary or downgrade trigger.stopAdd does_not_claim and withdraw condition.
evidence_gap_as_permissionA missing field is treated as permission to proceed.stopRewrite as blocked or no_credit.

Challenge

Attack one public field.

The best critique is small: report a broken public link, a hash/DOI/status mismatch, stale evidence, missing stop condition, or a sentence that turns an evidence gap into action permission.

Use public materials only. Do not request private runtime logs, customer data, account state, financial execution details, internal prompts, commercial schedulers, private endpoints, or production traces.

Boundary filter visual separating public paper evidence checks from private systems.

Public / Private Boundary

Validate public records, not private systems.

This update is limited to public registries, pages, manifests, demos, schemas, issue routes, boundaries, and counterexample routes. It excludes private runtime, customer data, real accounts, production logs, internal prompts, financial execution details, and any executable action strategy.