For judges · no upload, no judge code
Judge path. Four stops, in order.
Step 1 re-runs Muse Wonder 1.0’s saved program live against the booklet’s printed conditions; steps 2 to 4 open recorded runs, each labelled with its run id.
Line 001 · The four stops
One live run, three recorded ones.
Open them in order. The SAM 3.1 and Spark 1.3 calls behind step 2 are listed with response ids on /evidence; Wonder’s three program-writing calls are shown with its attempts on /runs.
-
1live
Re-run the program Muse Wonder 1.0 wrote, live
The sandbox re-runs the saved program against the booklet's printed conditions (22/22 at record time). Of the four stops, this is the only one that runs live.
22/22printed booklet conditions reproduced, within 0.1 % · evidence/selftest-booklet.md
-
2recorded
Watch one real hull get measured, then refused
Clip dB7X is third-party test footage with no ground truth. SAM 3.1 tracks the numerals and the water in every frame, Muse Spark 1.3 reads the labels, five code checks confirm them, and code averages the swell to 12.761 m ± 0.086 m (repeatability).
That draft is past the only booklet's table, so the certificate refuses.
REFUSE · OUT_OF_TABLErec-db7x-20261031 -
3recorded
See Wonder's three attempts
Attempt 1 was rejected by the sandbox's code gate; attempt 2 reproduced 14 of 22 conditions; attempt 3, 22 of 22.
Between attempts Wonder saw the gate error or its pass count, the failing ids and its own intermediate values. The printed answers were never in its prompt.
- Attempt 1gate
- Attempt 214/22
- Attempt 322/22
bar length = share of the 22 printed conditions passed · run rec-trieste-20261031 · evidence/selftest-booklet.md
-
4recorded
Open the two certificates
The refusal for the real hull, and the issued survey on the booklet's own ship (condition 4.1, all six drafts entered): 17825.8 t against the printed 17827.7 t, issued with a warning; cargo not computed.

rec-db7x-20261031 · real hull 
rec-trieste-20261031 · condition 4.1 · 6 drafts entered
Line 002 · Optional
Measure your own clip live.
/survey runs SAM 3.1 and Spark 1.3 on your own prepared draft-mark clip. Uploads need the judge code from the submission notes. Our real-hull test clips are third-party footage we do not re-host.
- 002.1An
MP4, cropped to the mark column. - 002.2At most 10.5 s, 64 frames and 3 MB.
- 002.3The judge code from the submission notes; each upload spends Meta API budget.
Line 003 · Limits
What this does not prove
Four of the six honest limits in the README, quoted (each cites its README number); all six are on the landing page.
- 1
“The weight check is on 12 loads of a rendered synthetic barge (code-drawn pixels, truth by physics), not a physical scale: it checks the chain's arithmetic, label checks and wave-averaging, not real water, glare or a real print.”README · Honest limits, 1 (abridged)
- 2
“The real-ship clips are third-party test footage with no ground truth (repeatability and refusal only, no accuracy claim); no real-ship draft accuracy is claimed.”README · Honest limits, 2
- 3
“The test booklet is a Ro-Ro/passenger ship, not a bulk carrier; a draft outside its 0.5-9.3 m table is refused, so the two real clips certify as REFUSE OUT_OF_TABLE.”README · Honest limits, 3 (abridged)
- 4
“This is a prototype, not a surveyor's document.”README · Honest limits, 6 (last sentence)
Then the rest, in one click.
The demo video, the pitch deck and the source.