YUPAY (Quechua: to count / to reckon / to audit) is our governed multi-model audit harness. We adopt the audit methodology from Kilo Code / André Lindenberg ("We Audited the Same Codebase…"): give every model the same task, then score issues-found · tokens · cost · latency per model. We run it over our own governed open models and emit one DSSE-signed comparison receipt with a Restraint verdict. 0 CDN.
No M3 weights, no M3 derivative. MiniMax M3 is EXCLUDED-BY-DOCTRINE — its open-weight license restricts military/defense use and MiniMax is PRC-based; SZL demos at Defense Unicorns Warhacker. M3 appears below only as a non-participating reference row, never run.
| Model | Label | Issues | Recall | Tokens | Cost (USD) | Cost/issue | Latency |
|---|---|---|---|---|---|---|---|
| Run the audit to see the governed comparison… | |||||||
Run the audit to see the DSSE-signed comparison receipt + Restraint verdict…
locked theorems = 8 {F1,F4,F7,F11,F12,F18,F19,F22} @ kernel c7c0ba17 ·
Λ = Conjecture 1 · Khipu = Conjecture 2 · SLSA L1 honest / L2·L3 roadmap ·
receipts: DSSE ECDSA-P256-SHA256 · 0 CDN · trust ceiling < 1.0 (the recommendation is bounded, never absolute).
Honest labels: no key is wired in this Space, so rows are MODELED (issues from the published Kilo benchmark;
cost = MODELED tokens × published per-token rates, cited). A row is MEASURED only when a real run happens in-process;
M3 is EXCLUDED-BY-DOCTRINE, never run. We never fabricate a benchmark.
Attribution: Kilo Code / André Lindenberg audit methodology (blog.kilo.ai) + MiniMax Sparse Attention paper
(huggingface.co/papers/2606.13392) as INSPIRATION. YUPAY is our own implementation over our own open models —
no M3 weights, no M3 derivative. See NOTICES.md.