Local Laya model

--approvals laya serves TypeSafe's POST /systemone route on the proxy itself and points the approval hook there with PADWAN_PROXY_APPROVALS_URL. Verdicts are advisory: the hook always returns ask, whatever the confidence gate.

Prerequisites: the laya extra (uv sync --extra laya, several GB of CUDA torch). The proxy refuses to start without it.

uv run padwan-proxy --backend-url https://api.scaleway.ai/v1/ -m glm-5.2 \
  --claude-config ~/.claude-scaleway --approvals laya

• Weights load once per worker at startup, followed by a warmup call. Startup logs the device, including any CPU fallback.
• First use downloads weights from Hugging Face; set HF_HUB_OFFLINE=1 once cached.
• Inference is serialized under a thread lock. Separate server processes each load their own model.
• Switching back to --approvals jev drops the URL.
• To use a proxy running elsewhere (such as the Docker image), set PADWAN_PROXY_APPROVALS_URL yourself. --approvals laya without --claude-config serves the route and writes no settings.

Checkpoints

Question schemas and option counts (2–20 for choice/score) are checked before inference.

Limitations

• Laya reads about 316 tokens of state (user requests, working directory, tool name and arguments). The tool call sits at the end, so a cut would hide what is being judged: the proxy refuses oversized states with 413 and the hook asks. Expect this on any session with more than a few short user turns.
• Laya's choice confidence is an entropy score, not its winning probability. Lowering the gate does not establish accuracy.
• Local endpoints set through PADWAN_PROXY_APPROVALS_URL always stay advisory; the raw verdict is still logged.

Benchmark

The runner reads benchmarks/data/laya_approvals.jsonl, which is not in the repository: 140 synthetic cases, 70 English/French pairs, split by task family into 70 development and 70 test cases. No transcripts are used. The candidate policy asks three factual questions in one batch and maps the answers in code; it is an experiment, not the deployed policy.

# HF_HOME must point at a cache containing these checkpoints; nothing is downloaded.
HF_HOME=/path/to/huggingface just benchmark-laya --device cuda \
  --models convaiinnovations/laya 'convaiinnovations/laya#multilingual' \
  'convaiinnovations/laya#typed-decisions' --output .scratch/laya-benchmark

The runner records dataset, question, source, config and checkpoint hashes, versions, raw answers, device, memory and latency. Thresholds are chosen on development cases before test inference. Reports include unsafe approvals, false denials, escalation, fit rate, language/tool slices and Wilson intervals (descriptive only: translated pairs are correlated). baseline_production_0_99 is the former automatic policy; always_ask is the current hook. Latency covers the four questions batched together, excluding HTTP and hook overhead.

Results (September 2026, local GPU)

The former 0.99 gate escalated all 70 test cases on every checkpoint. Four approvals are insufficient evidence of reliability, so the candidate was not promoted and no production threshold was lowered.