DAVIT

Synthetic data · robotics simulation5/5 checks passed2026-08-05cleared

A depicted bin episode frame: coloured blocks and spheres at rest inside a grey industrial bin

Twelve seeded bin episodes computed by a real solver, five rejected seeds on the record, and a plausibility review that catches generative imitations of its own scenes.

What the checks read

Each line is the buyer's own wording, beside what the check run observed.

every object settled to rest inside the bin, nothing sunk into the floor deeper than two millimetres and nothing tunnelled through a wall
12/12 episodes re-validated from recorded rest poses; 5 rejected seeds on the record
a manifest that hashes every file and records the seeds and solver settings
every named artifact re-hashed to its recorded sha256; two same-seed reruns at 0.0 m delta
Render each episode from its own fixed camera at 832 by 480 and twenty-four frames a second, silent by design, with no black frames and no stuck picture.
12/12 renders passed the deterministic video battery at exactly 832x480 / 24 fps
the independent physical-plausibility review catches every imitation and passes the delivered clips
4/4 delivered samples passed with image citations; 4/4 imitations caught
generate imitations of my scenes from their own first frames
four Cosmos clips anchored on the episodes' own first frames, each sha256-bound to its judge receipt

Project · one contract, four pieces

The simulator computes the physics; the picture only depicts it. The review has to catch the fakes.

The commission

What was asked

The buyer trains bin-picking robots and is done paying for videos that invent the physics. Every trajectory must come from a real simulator, regenerate from its seed, and end with every object at rest inside the bin — and the plausibility review itself must be shown to catch generated imitations before the delivered clips are trusted.

Twelve delivered, five rejected, reruns identical

The clean run

Revision B ran at 48 Hz with a redraw budget of three seeds per delivered episode. Twelve episodes settled inside every sealed bar — rest reached, floor penetration under two millimetres, nothing outside the bin — and five seeds that missed a bar are banked beside the twelve that passed. Two same-seed reruns matched their episodes exactly.

episodes delivered: 12 / 12
rejected seeds on the record: 5
max rerun delta: 0 m
imitations caught: 4 / 4

The simulator computesPhysX GPU at 48 Hz. Every trajectory, contact, and rest state is solver output, seeded and regenerable.

The renderer depictsGPU graphics is impossible at the host-kernel level on the rented VM — banked as a typed receipt, not worked around. The delivered pictures are drawn by a first-party rasterizer that reads exactly the recorded ground-truth poses. It owns no motion, no contact, no timing.

The review catchesgpt-5.6-terra under the judged_physical_plausibility bar — proven against generative imitations before the delivered clips were trusted.

Red first — the validator earns its keep

The red run

Revision A ran the same engine with a deliberately coarse solver step. Nothing settled: every episode was refused on tail drift or missed rest, and the refusals are banked beside the delivery. A validator that has never refused anything proves nothing.

revA · 8 Hz · refused0 / 3

Nothing settled inside the sealed bars.

revB · 48 Hz · accepted12 / 12

Same engine, same bars; only the solver step changed.

obj_1: tail drift 3.12 mm over the final 0.5 s exceeds 0.5 mm
revA delivered=0 rejected=3 → exit 1

revB delivered=12 rejected=5 rerun_delta=0.0 m → exit 0
floor penetration ≤ 2 mm · tail drift ≤ 0.5 mm / 0.5 s

Rest is pose truth: the solver's own velocity readback reported a constant 4.76 mm/s on a static body for nineteen straight seconds, so the validator recomputes stillness from recorded poses and never reads a velocity.

Four imitations, four catches

The imitation test

Cosmos 3 was anchored on each scene's own first frame and asked to imagine the drop — the strongest imitation of the commissioned scenes this checkpoint produces. The independent plausibility review must catch every imitation and pass every delivered clip; a fake it waves through voids the leaf.

sample-1..4.mp4

Four delivered renders — depicted from recorded solver states.

pass

imitation-1..4.mp4

Four Cosmos 3 clips anchored on the same scenes' own first frames, asked to imagine the drop.

caught

The review also refused our own first depiction: at the original camera, settled bodies vanished behind the near bin wall and the pictures honestly read as disappearing objects. The camera sampler now proves the whole inner floor visible before an episode renders, and that refusal is banked beside the delivery.

gpt-5.6-terra · fixture set 2026-08-04.2 · flaw recall 1 · sycophancy 0 · docs/qa/artifacts/20260805-cosmos-isaac/work/davit/acceptance/leaf-4-judged-clean.json

One contract, four priced sub-contracts

The split

episode engine + settled physics£282.50
manifest hashes + determinism reruns£282.50
episode renders + media battery£120.40
judged review + Cosmos generative red-team£157.25

The bytes, and what the checks made of them

The evidence

The buyer's balance was funded through the production Stripe webhook handler, held in escrow against the sealed root, and drawn leaf by leaf only as each banked acceptance receipt verified. Every ledger entry sums to zero.

root: £842.65
worker take: £618.50
ledger: zero-sum

pi_3U0xL1FFRbB25fA70ezkMmhA · Stripe TEST · livemode=false · settled 2026-08-05 · docs/qa/artifacts/20260805-cosmos-isaac/settlement/commission-trace.json

The honest boundary

What this does not cover

This dataset proves solver-computed rigid-body settling under one bin geometry family and depicted appearance, not photorealism and not deformable or fluid contact. The imitation test bounds the review's blindness on these scenes; it does not certify the review against every generator.

docs/qa/artifacts/20260805-cosmos-isaac/

How the money moved

Every step names the banked record it was read from.

asked
Five buyer-quoted criteria sealed across four priced leaves before any episode existed.
1 banked record
- docs/qa/artifacts/20260805-cosmos-isaac/tree/davit-tree.json
checked
The unchanged validator refused revision A outright: at a deliberately coarse solver step, nothing settled inside the sealed bars.
1 banked record
- docs/qa/artifacts/20260805-cosmos-isaac/work/davit/acceptance/leaf-1-revA.json
passed · 3 steps
Twelve episodes settled at 48 Hz with bitwise-identical reruns, and the plausibility review passed every delivered clip while catching all four generative imitations.
3 banked records
- docs/qa/artifacts/20260805-cosmos-isaac/work/davit/acceptance/leaf-1-revB.json
- docs/qa/artifacts/20260805-cosmos-isaac/work/davit/acceptance/leaf-4-judged-clean.json
- docs/qa/artifacts/20260805-cosmos-isaac/work/davit/acceptance/leaf-4-judged-red.json
paid
The production webhook booked the test top-up, escrow held the sealed root, and each leaf drew only as its banked receipts verified — zero-sum.
£842.65
2 banked records
- docs/qa/artifacts/20260805-cosmos-isaac/settlement/stripe-test-receipt.json
- docs/qa/artifacts/20260805-cosmos-isaac/settlement/ledger-receipt.json

Who made it

SeaOtter's Superteam

SeaOtter's own team — first-party work, labeled so. Every number on this page is read from the production receipts; the file that holds each one is named beside it.

£842.65settled

Price yours

A job like yours

Last bankedon a machine£300.00passed 2026-08-02

Your job prices on your own words — a fixed quote before any account.

Price a job like this →