Verification record · Build with Gemini XPRIZE
Polyfork
The traffic-cone geometry claim reproduces exactly from public code; customer and revenue figures remain unverified.
Claim
Polyfork’s submitted claim is that its 3D assets are small programs: changing a knob rebuilds geometry rather than stretching a mesh. Its traffic cone gains a reflective band and increases from 280 to 472 triangles. The submission also reports 54 paying customers, revenue every week since 1 August, and zero related-party revenue.
What we tested
On 16 September 2026, sound.fan fetched the free traffic-cone module without an account and ran it in Node with three.js 0.180. Its executable logic matched the live file. At the default 12 facets, we swept tallness from 0.550 to 0.900 in 0.001 steps, compared a rebuilt cone with a uniformly scaled control, and checked odd facet counts and out-of-range height inputs.
Finding
Reproduced: the cone has 280 triangles at tallness 0.55, 376 at 0.70 and 472 at 0.90. Uniformly scaling the short cone to the tall height leaves it at 280 triangles. The geometry claim reproduces exactly from public code.
The triangle count steps at 0.611 and 0.802. The module’s notes give 0.61 and 0.81; the latter is a 0.008 m rounding difference. Nine facets snap to ten segments, producing 318 triangles; a tallness of 5 clamps to 0.90.
What the evidence establishes
The module’s behavior and anonymous access moved from Claimed to Reproduced. Band placement is deterministic code with a fixed 0.191 m pitch. No model runs when a buyer uses this asset; that is consistent with the claim that agents write models as programs.
Observed, later live surface: Polyfork’s own catalogue file listed 1,519 assets, 712 free, on 16 September. This is participant-controlled evidence, not independent corroboration of the submission’s 583 models, 295 free, or 175 built in the preceding week.
What remains unverified
Not verified: 54 paying customers; weekly revenue; zero related-party revenue; 43 of 100 founding seats sold at submission; 14,524 API/MCP calls from 29 keys; the judge-agent rejection figures; and the claimed Gemini/Vertex AI generation workflow. The later “sold out” message is a participant-published observation, not a reconciliation of August sales. Running one public module does not establish the factory’s production throughput.
What evidence would change that
A dated Stripe export with hashed customer IDs could reconcile distinct paying customers against processor records. API-key call counts, a sample of judge-agent verdict logs, and a Vertex AI billing export for the build period would address the separate usage and AI-role claims.