Stoop proves an edge vision model is fit to fly — precision × runtime × hardware ×
degradation, measured as a shape, not summarized into one clean-data figure. This
board is a snapshot of everything committed so far: two days of real CPU/GPU runs, a
golden set that just found the model's real generalization gap, and a claims ledger
where every public number traces back to a file in this repository.
make reproduce-all is not implemented yet — nothing here re-derives
itself automatically. Read a card, follow its evidence link, check the file.
STOOP Lab v0 — FP32 vs INT8, live in your browser →
peregrine's real detector, quantized the nanostoop way, running side by side via ONNX Runtime Web on a bundled deterministic clip — self-contained, no external requests. Accuracy divergence only; latency is embedded from the real hardware matrix, never measured in-tab (see why inside).
One card per committed evidence run, newest first. A dash
means a number was not measured — see docs/DESIGN.md's truth-badge rule.
Cost figures are computed from live Cloud Billing Catalog rates, not a billing-console
read.
Every public number in this repository, parsed from
docs/CLAIMS.md as currently committed. One row per claim, the evidence
file that produced it, and its current status.
| claim | number | truth | status | evidence |
|---|