Rook Replay: Rook makes robot failures reproducible
Bring us one failure from your machines' decision layer. We re-run the decisions on a laptop when the recording can support it, and keep the incident as a test your next release has to pass.
01 · A REAL CRASH
Why did the drone do that?
COMMANDED vs ACTUAL ROLL · DECODED FROM THE VEHICLE'S OWN LOG · GRADE: OBSERVATIONAL
The log can describe the crash. It cannot re-run it.
A recording of outputs is not a capture of decisions. That gap is what Rook closes.
02 · A CAPTURED INCIDENT
This one, you can re-run.
OPEN-RMF BLOCKADE MODERATOR · REAL DECISION CODE, SIMULATED ROBOTS · 13 MIN · 1,516 DELIVERIES
ONE SHAREDCHECKPOINT
GRANTED
ROBOT 1 HOLDS2:29.7 · 22.1 S
Same capture, same engine, the same decisions. Byte for byte, on any supported machine.
672 / 672
EXIT 0 · VERIFIEDEXIT 1 · REFUSED
- bag_run_hash
- 99ba8401…d72ec4e5
- bag_output_digest
- 8265ec61…e22aad22
heartbeats_recorded.bin: byte 13227 rewrittenMANIFEST.sha256: repaired to matchdiverged: recorded heartbeat streamfirst divergent tick: 1787185222610908825
RECORDED: EVERY GRANT, EVERY CHECKPOINT ARRIVAL, EVERY TIME. DRAWN: THE MOTION BETWEEN THEM. NO POSE WAS EVER RECORDED.NOTHING REPLAYED.The machine would rather show nothing than show something that is not what happened.
REPLAY INTEGRITY: GUARANTEED · CAPTURE FIDELITY: INSTRUMENTED-COMPLETE
03 · THE RECORDER
This is all Rook is.
RECORD
A recorder sits at the boundary of your decision code and writes down every message it takes and every effect it emits, in order.
GRADE
Every recording is graded before anyone trusts it. The crossing earned instrumented-complete, a recorder at the boundary for the whole run. The crash had only the vehicle's own log, and graded observational, the bottom of the ladder.
RE-RUN
The decision code runs again in a sandbox on a laptop, offline. It reads the recording, makes the same choices, and lands on a hash a stranger can recompute.
KEEP
The incident is kept as a test your next release has to beat before it ships.
2:26.2INROBOT 0 · REACHED CP 3
2:26.2OUTROBOT 0 · GRANTED CP 3 ONLY
2:28.7INROBOT 3 · REACHED CP 3
2:29.2OUTROBOT 3 · GRANTED CP 3–5
2:29.7INROBOT 1 · REACHED CP 2
2:29.7OUTROBOT 1 · GRANTED CP 2–3
2:35.2INROBOT 3 · REACHED CP 4
· · ·ROBOT 1 HELD · 22.1 S
2:51.7OUTROBOT 1 · GRANTED CP 2–4
2:55.8INROBOT 1 · REACHED CP 3
CP 4 IS THE SHARED CHECKPOINT · NINE LINES OF A 13 MIN RECORDING
What runs again is the decision layer, the messages in and the choices out. Perception, physics and radio never run twice. The replay reads them back from the recording, exactly as they happened.
04 · THE STUDY
Send us one failure.
The Repro is $5,000, two weeks from data in hand, money back. A free 48-hour completeness check on the recording comes first, so we never take money for a hopeless case. The first three customers pay $2,500 as named design partners who let us keep the adapter and publish an anonymized writeup.
Then the Library, the customer's repros running before every release so a fixed bug can never quietly return. $30,000 a year for the first release stream and $18,000 for each additional one, an opening hypothesis tested on the first three contracts.
A recording that reaches us is used only to run the Repro. How it arrives, where it sits, and when it is deleted will be written on the Repro terms page.
WHAT TO PUT IN THE EMAIL · ONE LINE EACH
- Which failure would you most want to run again?
- Which service decides what the machines do next?
- What did you keep from that day, and for how long?
Reach out to the founder. He reads every one.
saketh@rookreplay.com