Mian Zhang

Public Release - 2026-06-12

When vision is unclear, downgrade the claim.

Artifact: Robot/Vision Shadow Evidence Boundary v0. A robot vision demo should show more than a successful frame. It should show what happens under occlusion, low confidence, bad lighting, and human-review thresholds.

candidate confidence occlusion downgrade review
Papers Evidence Map Counterexamples Registries
Robot vision stop-rule visual showing uncertain recognition moving into downgrade and review.

Artifact

A screenshot is not deployment evidence.

This update clarifies the P41/P44/P45 robotics and vision boundary. Candidate, shadow, and addendum evidence must not be described as real deployment, third-party replication, or unconditional optimality.

The public claim is narrower: when perception is uncertain, the system should downgrade, stop, or ask for review instead of treating a candidate result as an action-ready fact.

1

Observed frame

What the visual material actually shows.

2

Blind spot

What the frame does not show: occlusion, lighting, ambiguity, or alternatives.

3

Confidence state

Whether the system knows the result is uncertain.

4

Downgrade rule

The condition that moves the claim from action-ready to review.

5

Human review

Where review enters before real-world action.

6

Boundary note

No real deployment, no third-party replication, no unconditional optimality.

Vision Boundary

Recognition is only one layer of evidence.

A bounding box around one object can be useful, but it does not prove stable perception across messy scenes. The harder question is how the system behaves when the evidence weakens.

The page separates observed result, missing context, confidence state, downgrade rule, and review route.

Downgrade rule entry for robot vision evidence with confidence and review thresholds.

Downgrade Rule

The stop condition is part of the evidence.

The most important field is not the best frame. It is the rule that prevents a weak frame from becoming an over-broad deployment claim.

If the system cannot say 'I am not sure,' pause, or route the case to review, the demo is easier to overread than to trust.

Challenge Surface

Find any visual claim whose evidence tier is inflated.

A useful challenge points to the exact screenshot, caption, registry row, or wording where candidate evidence is treated as stronger than it is.

The repair is not rhetorical. The claim should narrow, the tier should downgrade, or the missing review route should be added.

Matrix

Robot/vision evidence fields.

FieldRequired valueFailure modeRepair route
visual_artifactThe public image, frame, or demo route.Artifact is treated as complete proof.Add evidence tier.
confidence_stateKnown confidence, uncertainty, or missing threshold.Confidence is implied.Add threshold field.
blind_spotOcclusion, lighting, similar objects, or scene limits.No negative visual boundary.Add does_not_show.
downgrade_ruleWhen the claim moves to review or stop.No downgrade path.Add stop condition.
review_entryWhere human or bounded review enters.Action is implied.Route to review.
does_not_claimNo real deployment or unconditional optimality.Tier inflation.Narrow the claim.

Challenge

Challenge one visual boundary.

Point to any visual claim that upgrades candidate or shadow evidence into deployment evidence.

The strongest challenge names the artifact, the inflated sentence, and the boundary update it requires.