Jev 1.13

I spent today wiring TypeSafe's Jev — the "System One" decision model that launched on Sept 15 — into a decision gate that routes coding tasks. Most of the time went into finding the endpoint, so here is what's actually true as of 2026-09-20. Everything below is measured, not read off the docs. The endpoint nobody points you at It works on OpenRouter with an ordinary OPENROUTER_API_KEY: POST https://openrouter.ai/api/alpha/decisions Content-Type: application/json Authorization: Bearer $OPENROUTER_API_KEY { "model": "typesafe/jev-1.13", "state": "help! my payouts have been failing for 3 days", "questions": { "tier": { "type": "choice", "instructions": "pick a tier", "criteria": { "billing": "payments", "technical": "bugs" } } } } Response: { "model": "typesafe/jev-1.13-20260917", "answers": { "tier": { "type": "choice", "choice": "billing", "probabilities": { "billing": 0.87, "technical": 0.13 }, "confidence": 0.82 } }, "usage": { "input_tokens": 312, "output_tokens": 35, "cost": 1.31e-05 }, "provider": "TypeSafe" } The traps, each one measured: /chat/completions with typesafe/jev → 400 "is not a valid model ID". Jev does not generate text, so it isn't an autoregressive chat model. Don't try to shim it into standard completions. /api/v1/decisions → 404. The route is currently live under /api/alpha, not /api/v1. Wrapping the body in {"decisionsRequest": {...}} → 400 invalid_type ... path: ["model"]. The body must be flat: model, questions, and state at the root level. GET /api/v1/models (447 entries) lists no typesafe/* ID, and GET /api/v1/models/typesafe/jev-1.13 404s. Yet GET /api/v1/models/typesafe/jev-1.13/endpoints returns 200 with the full provider snapshot, pricing, context, and uptime. Any SDK preflight asserting "does the model exist in the list?" will falsely fail. Make a live call instead. There is no typesafe/jev-latest alias on OpenRouter; that alias throws an error. typesafe/jev-1.13 is the only working ID, and every response names the dated snapshot it served. Measured latency and cost: 239–430 ms per decision, $1.3–2.7e-05 per invocation (roughly $13–27 per million decisions). Output is unmetered. The answer shapes are not uniform choice gives you choice + probabilities + confidence. noul (yes/no) does not: {"type": "noul", "noul": 0.01} For noul, the value is a flat scalar float between 0.0 and 1.0 representing true/yes likelihood (e.g., 0.01 means overwhelmingly false), not a structured dict with probability labels. I had code reading answers.x.probabilities.true, which silently defaulted to 0.0 — so a "does this patch match the intent?" gate answered no to everything, including good patches. If you port Jev snippets around, check that field shape. The failure mode worth designing against My first version of the gate had the classic fallback: except Exception: return choices[0]. With a live key pointed at the wrong route, every call 400'd and the gate answered anyway: triage("refactor auth across 14 modules") → bash_direct risk("auth models, 31 dependents") → low (would have auto-applied) verdict(exit 1, "could not read Username") → complete (would have ended the loop) Confident, cheap, and wrong in the direction that costs you. The fix that stuck: the gate returns unavailable plus a default chosen from the stakes — fail open to the cheapest path for low-stakes routing (restricted strictly to non-mutating checks), fail closed (high, abort) for the risky ones — and the process exit code explicitly indicates whether you got a model verdict or a policy default. Two things I'd have missed without an end-to-end smoke test git diff is blind to files a coding agent creates: they're untracked, so a patch review gate sees an empty diff and rejects a perfectly good change. Use git add -N . && git diff (or git add -A && git diff --cached, but remember to unstage cleanly). I pointed the harness log inside the worktree, so __pycache__ and the run log itself landed in the diff. The gate called that scope creep (noul 0.47 vs 0.94 for the clean version) and it was completely right — plus the messier diff cost 2x because you're billed on input token size. Keep run logs outside your worktree or in .gitignore. Happy to share the routing-gate code if anyone wants it; it's stdlib Python and talks directly to the OpenRouter alpha endpoint. Verify the alpha route before you build on it — I measured this on one box today.