Reasoning in context
Capture the expert’s hypotheses alongside the real fault, while the work is happening.
Beyond task video The debug layer for physical AI
The world breaks.
Experts fix it.
AI should learn how.
We record how experts debug the physical world. The reasoning, the failed attempts, and the verified fix. Every decision that teaches a machine why.
Real faults. Full decision loops. Verified outcomes.
INSIDE AN EPISODE001 / PHONE REPAIR
ILLUSTRATIVE DIAGNOSTIC VIEW
FAULT → DIAGNOSIS → FIX+
01“Could be the battery.”HYPOTHESIS · EXPERT TESTIMONY?
02New battery. Still dead.DEAD END · SIGNAL WORTH KEEPING↳
03Short found. Replaced. Boots.FIX + FUNCTIONAL CHECK✓
A complete decision loop.Not just a video.
Symptom↗Hypotheses↗Test↗Dead end↗Fix↗Verified ✓
01 / THE MISSING LAYER
Doing is visible.
Diagnosing isn’t.
A successful demonstration shows what worked. It leaves out the uncertainty, the wrong turn, and the test that changed the expert’s mind. That’s the data we’re building.
Captured today
What the expert did
- Hands, tools, the finished repair
- Successful runs, filmed cleanly
- Sold by the hour
The missing layer
Why they did it
- The hypothesis, before the outcome was known
- The measurement that changed their mind
- The attempt that ruled something out
The wrong turns, too
A failed attempt rules something out. We keep the dead ends that explain the next decision.
Negative data is valuable dataA fix you can check
The episode ends with a functional check. A narrated claim and an observed outcome stay distinct.
Evidence over assertion02 / THE PRODUCT
The episode
is the product.
One phone repair. From “won’t charge” to a verified fix. Explore the decisions between the two.
Illustrative episode from the pitch deck.REPORTED
A phone that won’t turn on.
The phone is dead and won’t charge. The reported symptom starts the episode; the cause is still unknown.
THE STARTING POINT
“Phone dead. Won’t charge.”
Record the fault before an intervention. A symptom is the starting point, not a diagnosis.
Narration is logged as testimony. Video and the final functional check are the evidence.
03 / FROM THE SHOP FLOOR
Short clips.
Real diagnosis.
Excerpts from accepted episodes: the moment a hypothesis forms, the measurement that contradicts it, and the check that confirms the fix. Recorded with the collector's consent and the shop owner's.
Clip being cut
The expert says what they think is wrong, and why, before opening the device.
Clip being cut
A reading comes back wrong. The expert narrates what it rules out and changes course.
Clip being cut
A functional check against the nameplate. The claim and the observed result stay separate.
Clips are excerpts. Full episodes carry the synchronised video, audio, motion data and the written reasoning trace.
04 / BUILT FOR THE WORKBENCH
Real work.
Purpose-built capture.
Our app and off-the-shelf equipment support capture today. Our wearable v1 is designed to keep the head unit light and every stream on one timeline.
HANDSIGHT / WEARABLE V1Design specification
ONE TIMELINE. THE FULL PICTURE.CONCEPT ILLUSTRATION ↗
01
Multiple camera configurations
Mono, depth, and hand-pose capture for robotics use cases.
02
Hear the reasoning
A microphone array designed for clear narration in a noisy shop.
03
Every stream, in sync
IMU and a sync chip align movement, video, and audio.
04
Less weight on the head
A separate belt unit keeps the wearable focused on capture.
Capture specification
- Resolution
- 1080p selectable to 480p
- Battery
- 7 hours target, full shift
- Upload
- Automatic resumes after dead zones
- Motion
- ~50 Hz IMU on the video timeline
Resolution, upload and motion capture are live in the collector app today. Battery is a v1 hardware target, not a measured figure.
05 / FROM THE SHOP TO YOUR MODEL
Pay for accepted episodes.
Make every one count.
Acceptance criteria come first. Automated checks and expert review assess each episode against your requirements.
A real repair
A natural fault in a working repair shop.
Capture the loop
Video and narration, including the dead ends.
AI first pass
Automated checks prepare the episode for review.
Expert review
Assess the evidence against buyer criteria.
Accepted. Delivered.
You pay per accepted episode.
Rejected episodes: $0.
BUILT FOR
Embodied reasoning & VLM evaluation
Technician copilots
World models
06 / LET’S BUILD YOUR DATASET
Start with one trade.
Learn from every turn.
A paid pilot of 100–200 episodes. Agreed acceptance criteria. Real faults, expert reasoning, and verified outcomes.
100–200EPISODES1TRADE TO STARTFull loopEVERY DECISION
[email protected] Technician or repair shop? We’d love to hear from you, too.