The software sees plenty of messages
AMD GPU · PHYSICAL AI · EVIDENCE ASSURANCE
Before a robot acts,
look twice.
A warehouse robot should not treat six claims copied from one camera frame as six independent observations. Look Twice checks the evidence for the next action—and asks the scout for a new view before the carrier moves.
PREREGISTERED PRIMARY · 30 PAIRED WORLDSACTIVE 29/30 · PASSIVE 0/30
29/30active full-chain direct
0/30passive full-chain direct
60/60missions · 0 unsafe · 0 fallback
Secondary post-hoc feasibility check: all 29 worlds with a feasible corridor went direct; the only dual-blocked world took the safe detour. Offline labels were never available to the controller, and this was not the preregistered endpoint.
AMD FULL-RUN TELEMETRYWAREHOUSE AMR SIMULATIONPYTHON ∧ PURIFY GOLIU LIANG · TRACK 3
WAREHOUSE AMR / ZONE C-04ASSURANCE ONLINE
13-SECOND NATIVE 1080P REPLAY · SIMULATION ONLYRepair → dual admit → direct moveOpen at full resolution ↗ACTIONCROSS REGION
INITIAL GATEDENIED
NEXT ACTIONACQUIRE NEW ROOT
01 / THE PROBLEM
More conclusions do not mean more independent evidence.
Copies add no independent support
Collect what is missing, then decide
02 / THE LOOP
A denial becomes the starting point for the next observation.
Observe
RGB-D Claims arrive with time, scope, calibration, and capture lineage.
Qualify
The system checks physical roots, prediction sets, and the Action Contract.
Repair
A BeliefGap tells the scout which failed clause the next viewpoint should target.
Act
Direct motion opens only when the Python policy and Purify Go both agree.
03 / WHY IT IS DIFFERENT
The model interprets the scene. The Action Contract decides whether the robot may move.
PHYSICAL ROOTS
RGB, depth, and forwarded Claims from one capture count as one physical observation.
ACTION CONTRACT
Evidence must match the robot, payload, corridor, action, time window, calibration, and root requirement.
BELIEF GAP
A denial names what is missing. A fixed, auditable ranking selects a new view that targets the gap instead of repeating the same one.
PURIFY GO
Purify is not another model. It is a separately compiled Go contract gate; direct motion requires Python and Go to agree.
Deny and detour
It stays safe when evidence is weak, but pays for uncertainty with a longer route in every world.
PREREGISTERED CHALLENGE: 0/30 FULL-CHAIN DIRECTMove, repair, then act
The scout acquires an independent side view. The loaded carrier goes direct only after Python and Purify Go both admit the action.
29/29 FEASIBLE WORLDS DIRECT · 1/1 DUAL-BLOCKED DETOURACTIVE-VIEW SELECTOR SUPPLEMENT · 20 PAIRED WORLDSDIRECT OUTCOME HELD AT 20/20
−40.3%mean scout path
−15.0%mean team path
−27.5%mean physical captures
Candidate and baseline both finished direct in 20/20 worlds; the candidate reached the same action outcome with a shorter scout route and fewer physical captures. This is a separate same-generator, non-locked kinematic simulation—not a physical-robot result or a replacement for the primary result above.
04 / COMPLETE DEMO
Watch the complete workflow in place: AMD GPU execution, active evidence repair, dual admission, and the final action.
RECORDED 3:59 AMD GPU WORKFLOW
1920×1080 · 30 FPS · English narration and captions · Simulation only. The video shows the command-line audit, Genesis/ROCm execution, active evidence repair, dual admission, and result boundaries; it is not physical-robot footage.
OPEN VIDEO IN A NEW TAB ▶VIEW THE VIDEO FILE IN THE REPOSITORY ↗DOWNLOAD THE HASH-VERIFIED MP4 ↓WATCH THE 30-SECOND SILENT REPLAY ↗FULL WEBSITE SOURCE AND DATA ARE ON GITHUB