Tama checked the loader: score.py needs {id,predicted,scores,latency_samples_ms}
rows plus receipts; run_gate emits {case,rep,route,status,...}. The README had
named it as the scorer without checking.
README: say exactly what verify_export and run_gate check
Aoi Katsuragi
6fff83b
Tama's review: run_gate collects predictions and exits 0 on endpoint errors
(route ERR), so it is not a gate; verify_export's TensorRT check compares
route and token count only, not logits or the tokenizer hash.
Add serving and verification tools
Aoi Katsuragi
a5ee770
code/serve_trt.py serves the TensorRT engine over HTTP (POST /v1/route,
GET /health), defaulting to this model's five labels. code/verify_export.py
checks PyTorch, ONNX and an optional running TensorRT server agree on a dev
set. code/run_gate.py sends a case file to any router endpoint and records
each route and its latency. The README names all three.
Scope tutorial hardware claim to archived receipt
Tama
b2e0569
Archive verified router challenge and tutorial draft
Tama
98a6c22
Validate frozen router run fields in scorer
Tama
ea484dc
Bind router predictions to clean evaluator provenance
Tama
b24e171
Freeze corrected public router challenge and evaluator contract