Meta AI Developer Hackathon · Push the Frontier · Indie
The sea keepsmoving. The readingfinally holds still.
draftmark.edycu.dev · v1.0.0
Before dawn, on a launch
One number, read off a moving sea.
The swell runs the water up and down the painted marks. The surveyor eyeballs a mean, then trusts a per-ship spreadsheet: a judgment, not a record.
Real clip dB7X · redrawn at 1× from recorded geometry · 6 fps · no ground truth
Ten seconds of swell. One line holds.
60/60 · mean 12.761 m · u_rep ±0.086 m · held
band illustrative · label error 0.20 m12.761 m is past the only booklet's 0.5–9.3 m table: no tonnage is guessed.
Measure · SAM + Spark
SAM 3.1 finds the marks in 60 of 60 frames.
“white number”“water”
A label counts only when code agrees
- Monotoneok
- Pitch, 20 cm66.73 px
- Fit residual0.44 px rms
- Two frames agree9 shared
- Metre anchor13M · 14M
Spark 1.3 reads the strings on two frames with a strict JSON schema.
Prove · Muse Wonder 1.0 · the Trieste booklet's 22 printed conditions
Wonder writes the program. The booklet grades it.
| condition | 1 | 2.1 | 2.2 | 3.1 | 3.2 | 4.1 | 4.2 | 4.3 | 4.4 | 4.5 | 4.6 | 5.1 | 5.2 | 6.1 | 6.2 | 7.1 | 7.2 | 8.1 | 8.2 | 9.1 | 9.2 | 10 | score |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Attempt 1 | Static gate refused globals on lines 230, 255 and 260. It never ran. | gate | |||||||||||||||||||||
| Attempt 2 | pass | pass | miss | miss | miss | pass | pass | pass | pass | pass | pass | pass | pass | pass | miss | pass | miss | pass | miss | miss | miss | pass | 14/22 |
| Attempt 3 | pass | pass | pass | pass | pass | pass | pass | pass | pass | pass | pass | pass | pass | pass | pass | pass | pass | pass | pass | pass | pass | pass | 22/22 |
within 0.1 % of the printed displacementoutside it
Attempt 2 missed the second trim correction on 8 conditions.
Attempt 3 saw only the failing ids and its own intermediates, never the printed answers.
Certify · code only
Code prints every number. Then it signs them.
Every station is signed (HMAC-SHA256). Edit one: REFUSE UNSIGNED. Past the table: REFUSE OUT_OF_TABLE.


Architecture · how it fits together

- Meta model call
- Offline, cached
- Code, no model
- HMAC hand-off, verified next
- Settled / PASS
- No database: the browser carries the signed JSON. LIVE_MODE=off falls back to a recorded run.
Checked against a known truth.
A rendered one, for now.
The real clips end in refusal, so the weight is checked on twelve loads of a barge drawn in code: waterline from physics, swell and camera noise added on purpose.
rendered loads within ±2 % of the truth (worst 0.75 %)
median cargo error vs the synthetic truth
clips where SAM found the rendered water (a colour threshold stands in)
No physical scale check yet · no real-ship accuracy claim
Three Meta models · one load-bearing job each · code does the rest
Each model does one job, with a receipt.
Sees
Two video calls per clip: “white number” and “water”.
Reads
Two strict-schema label reads; code must agree.
Writes
The ship's survey program, from booklet pages only.
Prints + signs
Averages the swell, signs each station, issues or refuses.
Condition 4.1: 17825.8 t vs printed 17827.7 t · $0.0397 of Meta calls per dB7X run (estimate)
What is new · measured against our own baselines
Why an agent, not a template.
a generic textbook survey template
the booklet's own procedure, applied verbatim
Wonder, from the booklet alone; our hand-written reference also 22/22
New here
A wave-averaged reading from video, and an agent whose program must reproduce the booklet's printed conditions (at least 20 of 22, within 0.1 %) before it computes.
Honest limits · README
What it does not prove yet.
- §1The weight check is on 12 rendered synthetic loads, not a physical scale.
- §2Real-ship clips: third-party, no ground truth; the ± is repeatability only.
- §3The test booklet is a Ro-Ro/passenger ship, not a bulk carrier.
- §5A consistent whole-metre misread passes the checks; the user confirms the anchor.
- §6A clip reads only the port midship mark; the rest are typed in. A prototype.
Film ten seconds. Keep the number that holds.
draftmark.edycu.dev
/survey · /runs · /evidence
github.com/edycutjong/draftmark
Built by Edy Cu Tjong · MIT · v1.0.0