DraftMark

Meta AI Developer Hackathon · Push the Frontier · Indie

The sea keepsmoving. The readingfinally holds still.

$ python scripts/core_run.py clip dB7X_t45 draft d = 12.761 m u_rep = 0.086 m (k=1) cert REFUSE · OUT_OF_TABLE no guessing past the table

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.

mean = ?
The same 60 heights from clip dB7X, before any averaging.

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 m
12.761 m±0.086 m · k=1 frame 0frame 59 · 10 s
REFUSEOUT_OF_TABLE

12.761 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”

frame 0, 2× crop: SAM boxes, the water, Spark’s reads

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.

Wonder's three attempts graded against the booklet's 22 printed loading conditions
condition12.12.23.13.24.14.24.34.44.54.65.15.26.16.27.17.28.18.29.19.210score
Attempt 1Static gate refused globals on lines 230, 255 and 260. It never ran.gate
Attempt 2passpassmissmissmisspasspasspasspasspasspasspasspasspassmisspassmisspassmissmissmisspass14/22
Attempt 3passpasspasspasspasspasspasspasspasspasspasspasspasspasspasspasspasspasspasspasspasspass22/22

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.

REFUSE, OUT_OF_TABLE, matrix row 4.
rec-db7x-20261031 · real hull
ISSUE WITH WARNING, matrix row 9, positions entered.
rec-trieste-20261031 · 6 drafts entered

Architecture · how it fits together

DraftMark architecture. Per clip, live: phone clip, then SAM 3.1 finds the marks and the water edge, then Muse Spark 1.3 reads the painted marks with five code checks, then code averages the swell to d ± u; each hand-off is HMAC-signed. Per vessel: Muse Wonder 1.0 writes the program offline, a no-network sandbox self-tests it on 22 of 22 printed booklet rows, the proven program is pinned by sha256, and the survey runs it. Code issues a certificate or refuses: OUT_OF_TABLE, UNSIGNED.
  • 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.
SYNTHETIC

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.

12 / 12

rendered loads within ±2 % of the truth (worst 0.75 %)

0.12 %

median cargo error vs the synthetic truth

0 / 12

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.

sam-3.1 · video

Sees

Two video calls per clip: “white number” and “water”.

60/60frames kept, clip dB7X
muse-spark-1.3

Reads

Two strict-schema label reads; code must agree.

5/5label checks pass
muse-wonder-1.0

Writes

The ship's survey program, from booklet pages only.

22/22conditions, 3 calls, est $0.481
code · no model

Prints + signs

Averages the swell, signs each station, issues or refuses.

235testsboth escapes kept as regression tests

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.

0/22

a generic textbook survey template

14/22

the booklet's own procedure, applied verbatim

22/22

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.

Film ten seconds. Keep the number that holds.

draftmark.edycu.dev

QR code: https://draftmark.edycu.dev

Built by Edy Cu Tjong · MIT · v1.0.0