Celeris-1 Decision vs Laya

Celeris-1 Decision

Typed probabilities, image inputs and optional explanations suit routing and classification inside an agent. Both System One and OpenAI Decisions request formats are documented. The service remains early access under its terms, activation can queue, and prepaid credit expires after 30 days. The published accuracy and latency figures are Celeris measurements, not Anchor Terminal tests.

Laya

Apache-2.0 code and weights, installed with pip install laya, with Python 3.10 to 3.13 tested in CI. Base checkpoints score 0.362 and 0.352 on the maintainers' typed-decisions benchmark against a 0.318 random baseline, so it needs fine-tuning.

Celeris-1 Decision

1. Use the decision model path and matching model field. This model has no chat, Responses or models endpoint.
2. Use at most 512 questions per request, or 64 with explanations. Both endpoints reject streaming.
3. Preserve x_celeris in a custom Jev SDK response model when reading explanations. The default response type drops it.
4. Retry 429 with Retry-After and backoff. Stop on 402 until workspace credit is replenished.
5. Read usage.input_tokens for cost, including images and request overhead. Track credit expiry separately from token consumption.

Laya

1. Set LAYA_API_KEY before starting laya-serve. It listens on every interface by default
2. Gate on answer_confidence, not confidence, which measures entropy and doesn't match Jev's field
3. Shortlist choice questions with more than about 20 options using predict_shortlist or the laya_shortlist tool
4. Use semantic or opaque labels such as A and B, not yes and no, in choice questions. The checkpoints can follow the label text
5. Pass model="multilingual" and max_len=8192 for long documents. The English checkpoint stops at 512 tokens