Decision API

POST /v1/decisions takes a context and a JEVfire field schema. The schema property is a mapping of field names to definitions; it is not standard JSON Schema. The racing request is a complete example.

Parameter	Default	Contract
context	required	Nonempty string, at most 100,000 characters; rendered prompt must fit backend token limit
schema	required	1–64 independent fields
strategy	auto	auto, batch, prefill_then_batch, aligned_prefill
score_temperature	1.0	0.05–10; rescales label logprobs, not generation temperature
min_probability	unset	Relative-score abstention threshold from 0–1
cache_salt	unset	Nonempty string, at most 256 chars; passed to vLLM

Boolean field: {"type":"boolean","description":"Is boost explicitly charged?"}. Enum field: {"type":"enum","description":"Choose a lane.","choices":["left","hold","right"]}. Enums need 2–255 distinct, nonblank strings; each is at most 1,000 characters. Descriptions are 1–4,000 characters; names are 1–128 characters without controls. Additional properties are rejected.

Response

The response includes:

Property	Meaning
parsed_json	Field names mapped to booleans/strings, or null on abstention
fields	Per-field winner, original candidate scores and normalized probabilities
scores_are_calibrated	Always false
abstained_fields	Fields below the requested relative-score threshold
strategy, requested_strategy	Resolved strategy and caller selection
backend_requests	Number of native vLLM completion calls
elapsed_ms	Service-side elapsed time; client latency includes transport
usage	Logical prompt/output token usage, including cached tokens and padding
max_prompt_tokens, cache_block_tokens	Prompt length and configured cache boundary

Each entry of fields includes value, selected_value, probability, abstained, candidate_probability_mass, and a candidates list containing value, label, probability, and logprob. selected_value remains visible when value is null; do not execute it as though abstention succeeded.

The winning probability is normalized only over that field's candidate labels. It is not calibrated correctness. The model may assign little vocabulary mass to all offered labels. Label mapping and order can influence outcomes.

Strategies

• batch: score all field prompts together.
• prefill_then_batch: finish the first field, then submit the remaining fields.
• aligned_prefill: align the common prefix to the configured cache boundary, finish the first field, then submit the remainder. Requires a positive verified DECISION_CACHE_BLOCK_TOKENS.
• auto: use alignment for at least 16 fields, or at least four fields with context at least one cache block long; otherwise batch. With no configured cache block, always batch.

Each field includes its full label-to-value mapping in every score chunk. For more choices than the score cap, raw logprobs from identical prompt chunks are combined before normalization. The generated token itself is not used.

Failures

HTTP	Meaning
422	Invalid schema, unsupported strategy configuration, or overlong prompt
502	Backend error or missing/nonfinite requested scores
503	/health could not reach a healthy backend
504	Backend scoring timed out

There is no fabricated first-choice answer or external-provider fallback.

Nested objects and arrays

The inference contract is flat. Your application can deterministically compose known nested structure after scoring, as in tool_router.py:

decision = response["parsed_json"]
command = {
    "routing": {"tool": decision["tool"], "priority": decision["priority"]},
    "checks": [{"name": "confirmation", "required": decision["confirmation"]}],
}

This does not infer array length, free-text arguments, sibling dependencies, or arbitrary nested objects. For dependent decisions, use sequential stages or a single enum of valid joint actions. Validate action permissions in your own app.