# Typed decisions (Preview)

> Evaluate text or JSON state with open-weight decision models.

Source: https://www.nodus-compute.ai/docs/guides/inference/decisions/
Build revision: 718a07fed53e8ac579fe009e9513f452d8de8e98

**Alpha:** this feature may change.

Sign in to the console, choose **Inference → Models**, and select **Clef** or **Clef Flash** when available. The playground accepts text or JSON and lets you build yes/no, choice and rating questions without writing JSON. The playground places a wide input card beside model results. Select comparison models in the top toolbar, edit questions inline, and run them from the bottom of the input card. Edit instructions, answer types and options directly in each question row. Use **Compare** to evaluate the same input on up to three available decision models. Each column shows answers, probabilities, request latency and receipt-based cost. **Run 10 times** sends up to ten paid requests per selected model and reports agreement with the latest answer. **Stop** prevents subsequent requests; requests already sent may still complete and be billed. The **Code** control shows the first selected model’s request. Available decision models appear automatically in **Models → Decisions** and the playground. No separate customer setup or activation is required. This feature is in Preview.

Use `nodus/clef-flash` for fast decisions or `nodus/clef` for the larger model. Both return structured decisions:

* `choice`: the selected option, its confidence, and a probability for every allowed option.
* `score`: the probability-weighted score across an ordered list, starting at zero, with a legend and probabilities.
* `noul`: the probability that the proposition is true, from zero to one.

These probabilities are model estimates, not guarantees. Set application-specific thresholds and validate them against your own labeled examples before automating consequential actions.

Set `IDEMPOTENCY_KEY` to a unique request identifier and reuse it when retrying the same operation.

Terminal window

```bash
curl https://platform-inference.nodus-compute.ai/v1/systemone \
  -H "Authorization: Bearer $NODUS_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: $IDEMPOTENCY_KEY" \
  -d '{"model":"nodus/clef-flash","state":"Checkout is down for all customers.","questions":{"urgent":{"type":"noul","instructions":"Is this urgent?"},"team":{"type":"choice","criteria":{"technical":"Outages and errors","billing":"Payments and refunds"}}}}'
```

The same operation works at a named endpoint’s `/v1/systemone` URL. Keep an idempotency key stable when retrying the same request. `Nodus-Request-Id` identifies its receipt at `/v1/requests/{id}`.

Send 1–64 questions. Question and choice IDs use letters, digits, underscores, periods or hyphens, up to 100 characters. Choice questions require 2–64 options; score questions require 2–10 ordered levels. Instructions are optional; otherwise the question ID supplies the instruction. State may be a string, object or array. Requests are limited to 1 MiB. This preview does not accept images, video, streaming or chat-completion parameters.

Clef has a 65,536-token context; Clef Flash has 24,576. The full context’s input cost is held before dispatch, then unused credit is released after settlement. Long state may be truncated to fit the model’s available context; keep important context concise and inspect decision quality on representative inputs.

There is **no output-token charge**. Questions and internal schema formatting count toward reported input usage. The live catalog is authoritative for current prices. Missing usage or malformed answers are not charged.

## Model choices and pricing

The preview catalog includes Clef, Clef Flash, Clef Omni, Decider v1.2, Decider Flash v1.2, Joey 1, Elephant 1, Mammoth 1, Span-02, Kev 4B, Laya, Laya Multilingual and Laya Typed Decisions. Availability depends on which routes have completed qualification. Select a model explicitly; cheapest-route selection preserves that model rather than substituting a different model.

The console displays the currently eligible input price. Output is free, including when reported output tokens exceed the number of answers. Some models count the shared state once per question, so sending several questions can consume more input tokens than sending the state once. The temporary credit reservation covers this possibility; the final receipt bills actual usage. Laya accepts at most 32 questions and short contexts (512 or 1,024 tokens depending on the variant).

### Additional open-weight options

The reviewed catalog includes Blink v0.3, d3, Decider v1.1, REAX 1 and REAX 1.2, subject to serving readiness. Blink accepts one question with up to 16 choices. The other four accept up to three questions with up to 64 choices per question. All retain the preview’s ten-level score limit and input-only billing. Model context limits and current USD prices appear in the catalog; compare the exact models needed for your task.

The pinned d3 and Decider v1.1 routes retire on December 1, 2026; the pinned REAX routes retire on December 5. Requests stop when a route retires. Nodus does not silently replace your selected checkpoint with its successor.

## Rate limits and retries

For reranking, put the query and documents in `state` and use one `score` question per candidate. Keep candidate IDs unique across batches and combine answers by ID. Ten candidates is not a fixed API limit: choose batch sizes that fit the selected model’s context, including questions and formatting. Verify every candidate has a score before treating a search as fully reranked. Measure the complete search against its deadline; a successful individual request or settled receipt does not prove the search finished in time.

Nodus does not apply organization, endpoint or local model request-rate, token-rate or concurrency caps to decision requests. Previously saved throughput limits are ignored; model capacity and supplier throttling still apply.

For a `429` response, wait for the `Retry-After` interval before retrying. Pace concurrent requests instead of retrying every candidate immediately. A temporary model-capacity limit can still apply when your organization has no rate cap. Keep the `Nodus-Request-Id` and error message when reporting a failed run.
