System One decisions, without decoding

An agent asks small questions all day: is this the right queue, did the test pass, which of these three files matters. Answering each with a generated completion is slow and hard to parse. ignis answers them through the System One API, the typed-decision API the Jev model is served under: the same POST /v1/systemone route and body, so a System One client points its base URL at ignis and keeps working, here on a local model. It reads the answer from the model’s distribution instead of generating one. The evidence is prefilled once, each option gets a one-token label, and the answer is the restricted softmax over those labels at the single position after the prompt. A decision costs one prefill and decodes zero tokens: 36.5 ms at the median in the measured run.

How it works

noul (yes/no), choice and score are pure logit readouts. The server gives each option a single-token label, runs the prompt, and normalizes the probability of the labels at the last position. The top probability comes back as confidence, and the share of the full next-token distribution that the declared options hold is the answer mass, exported as a histogram. A low mass means the model wanted to say something else, the one failure nothing in the answer itself can show.

number, point and box generate one constrained digit per step with the prefix forced ({"x": then three digits), so they cost a prefill plus a round per digit. locate finds a line in a log, an item in a list or a sentence in a text without generating: calibrated attention heads at the end of the prompt narrow the text to a few candidates and a labelled choice decides. A log is first folded into templates plus values, so nothing near its length is prefilled.

POST /v1/decide and POST /v1/systemone are the same handler, and the errors follow System One’s own table; only model must be left out or name the model ignis serves. A decision defaults to the agent lane class: it holds no residency and takes no decode lane, so a fan-out of questions queues behind conversations instead of cutting ahead of them. The request and response are in Decisions.

The numbers

Figure	Measured on	Source
The declared options hold a median 99.8% of the distribution (p5 0.986, min 0.929); the unrestricted winner is a declared letter on 144 of 144	144 authored decisions, 27B, BF16 KV, thinking off	Typed-option logit readout
Prefill p50 36.5 ms, p95 40.7 ms, 28.0 decisions/s serial, no prefix reuse	same, prompts of about 132 tokens	same
Balanced accuracy 0.934, and 0.963 with the evidence moved into the system block (a difference of 4 rows)	same 144 rows	The evidence belongs in the system block
point inside the target on 236/240, 240/240, 120/120 and 60/60 of the four pre-registered sets	served endpoint, attention-head method	The sub-cell reading holds through decide
point errors of 84, 18 and 32 px on three synthetic 4096 px screenshots, every point inside its button	first constrained-digit experiment	Constrained digit readout
locate on logs of 100K, 200K and about 1M tokens: 21 of 23 (91.3%) against a floor of 85% registered in advance; median 2.6 s, p90 6.2 s	one cluster capture	Zero-decode locate, very long logs and prose
locate on prose: 88.0% overall, but 7 of 12 (58.3%) at 1M tokens, which failed its per-group floor	same	same

What these numbers are not

The 36.5 ms is one short prompt, serial, with no prefix reuse. The 144 rows come from one authored fixture whose own provenance says it is model-reviewed, not human-adjudicated, so 0.934 and 0.963 are not a quality claim for your task, and the finding says answer mass and calibration are sturdier than accuracy. The 0.9 abstention threshold needs re-measuring on any workload before it is trusted. point is bounded by the model’s roughly 32 px cell; the 84 px is the worst of three synthetic scenes, and the pre-registered sets are the stronger evidence.

See it, control it

• The route: POST /v1/decide, or POST /v1/systemone; types noul, choice, score, number, point, box and locate, with within as a JSON Pointer for locate.
• In the response: confidence and ranking; for locate also found, segment, value and pointers. usage.output_tokens is 0 for the three logit readouts.
• Metrics: ignis_decisions_total (by primitive), ignis_decision_answer_mass and ignis_locates_total, in the metrics.
• Logs: ignis.decide.locate, ignis.decide.located and ignis.decide.pointing_head.
• Playground: the Decide tab builds requests and shows the full distribution behind each answer; decisions are stored in the browser.

Limits

• Only noul, choice and score are pure logit readouts. number, point and box decode a digit per step, and locate runs several prefills and head passes.
• locate’s 91.3% is for logs. On prose the 1M-token group failed its registered floor.
• A decision has no reasoning: the loop guard does not apply, and the answer is the model’s first reading of the evidence.
• The wire shape is a third party’s, and the Decisions page says so.

Read more

• Decisions: the primitives, locate and the response.
• Images as evidence: decisions over a screenshot.
• Lanes that know who is asking: why a decision defaults to agent.