System One — typed decisions, not text

A System One model takes one state + named typed questions, evaluates every question independently and in parallel against the state, returns typed answers + probabilities. It does not generate text, write code, reason step by step, or call tools. Code owns control flow; the model makes narrow judgments.

Two providers, same wire shape:

	Ollama (local)	TypeSafe (hosted)
Endpoint	POST http://localhost:11434/v1/systemone	POST https://api.typesafe.ai/v1/systemone
Auth	none	Authorization: Bearer $TYPESAFE_API_KEY
Models	nimble, clef, clef-flash, Tev1 (Together AI)	jev-latest → jev-1.13.0, jev-preview
Min version	Ollama v0.35.0 (images v0.35.1)	—
Images	images: [base64…], Clef models only	not supported (text only)
Score levels	2–26	2–10
Choice options	2–26 per model pages (API schema says 255)	up to 255
Questions/request	1–64	not stated
Size limit	64 KiB body (32 MiB with images); must fit loaded context	64k tokens total; 32k for state + longest question
Extra field	keep_alive ("5m", seconds, 0 unload, negative = keep)	—
Cost	free	$0.042 / M input tokens, output free

PHP / Symfony: use Symfony AI 0.15+ (symfony/ai-decision-platform + the Ollama or TypeSafe bridge), not raw HTTP; classes, wiring and tests are in the symfony-ai-platform skill → Decision models, routing between agents in symfony-ai-agent → MultiAgent. This skill stays the source for question design and measured limits.

Details: references/api.md (both HTTP APIs + Python/JS SDKs). Worked examples: references/examples.md. Real nimble runs on invented products (exact request + response, timings, ground truth): examples/ollama/ — 10 object criteria → 400 · 11 structured e-commerce 11/11 · 12 weak spots (dates/math/negation/quantifiers) 7/10 · 13 support ticket routing + guardrail 6/7.

Request

{
  "model": "nimble",
  "state": {"ticket": "I was charged twice. Please refund the extra payment."},
  "questions": {
    "refund":   {"type": "noul",   "instructions": "Is the customer requesting a refund?"},
    "team":     {"type": "choice", "instructions": "Which team should handle this?",
                 "criteria": {"billing": "Payments and refunds", "bug": "Software errors", "account": null}},
    "urgency":  {"type": "score",  "instructions": "How urgently does this need a response?",
                 "criteria": ["Routine: no time pressure", "Soon: a customer is inconvenienced",
                              "Immediate: a critical service is unavailable"]}
  }
}

• state: non-empty string, JSON object, or array. Prefer an object with named parts.
• Question ids are never sent to the model — put the full question in instructions.
• choice.criteria: object {option: description | null}. null = bare label.
• score.criteria: array, index 0 = lowest. Scale is 0 … len-1.
• noul.criteria: optional {"true": …, "false": …}; Ollama defaults Yes/No.
• instructions and every criteria entry may be string, object, or array on TypeSafe. Ollama: only instructions may be structured; criteria must be strings (400 otherwise).
• Answers are not passed between questions. Dependent question → second request.

Response

{
  "model": "nimble",
  "answers": {
    "refund":  {"type": "noul", "noul": 0.9989},
    "team":    {"type": "choice", "choice": "billing",
                "probabilities": {"billing": 0.97, "bug": 0.02, "account": 0.01}, "confidence": 0.95},
    "urgency": {"type": "score", "score": 0.83, "legend": {"0": "Routine…", "1": "Soon…", "2": "Immediate…"},
                "probabilities": {"0": 0.25, "1": 0.67, "2": 0.08}, "confidence": 0.4}
  },
  "usage": {"input_tokens": 174, "output_tokens": 1}
}

• noul = P(yes), a number, not a boolean. No confidence.
• score = Σ level × P(level); can land between levels; not normalized to 0–1.
• probabilities/legend keys for score are level indices as strings (Python SDK: ints).
• TypeSafe's model returns the resolved version (e.g. jev-1.13.0) — log it.
• Ties in choice follow option order.

Confidence (Choice/Score only)

• Ollama spec: 1 − H(p)/ln(N) (entropy). 0 = uniform, ~1 = one candidate dominates.
• TypeSafe docs: Choice (p_max − 1/n)/(1 − 1/n); Score max(0, 1 − Σ p_i|i−m| / MAD_unif) with MAD_unif = (1/n)Σ|i−(n−1)/2|; noul-equivalent |2p−1|.
• The two providers document different formulas — don't port thresholds between them blindly.
• Confidence = peakedness, not correctness. Calibrate thresholds on your own data.
• Alternatives: top probability p_max, or top/second ratio.

Design rules (from the docs, in priority order)

1. Code first. Anything deterministic (dates, arithmetic, regex, status checks) stays in code.
2. Atomic questions. One snap judgment each. "Is this spam?" → six nouls (credentials request, unexpected reward, time pressure, sender/domain mismatch, link mismatch, disguised link) combined in code.
3. Batch everything in one request, including speculative questions only some code paths read. Parallel → near-zero latency cost. Batching 13 questions: ~12× cheaper, ~10× faster, identical answers.
4. Send only relevant state. Irrelevant content lowers accuracy (context rot).
5. Point at state parts with backtick paths: "Does `ticket.messages[0].text` request a refund?".
6. Pick the type by what code does next: Choice → switch, Score → threshold/sort, Noul → if.
7. Route on uncertainty: high → act, middle → confirm/review, low → human or bigger model. Threshold scales with action risk.
8. Composite scores: normalize each Score by len(criteria)-1, weight in code; tune weights, not prompts.

Writing good questions

• Noul: one condition; high value = yes ("contains personal data", not "is free of…"); statement or question both work; add criteria only when the boundary is subtle.
• Noul ≠ degree. "Strong in Python?" at 0.5 means unsure, not medium. Use a Score for degree.
• Score levels describe situations, not degrees. ["0","1","2"] fails. Each level is judged alone — "worse than previous" means nothing. One dimension per Score. Give rare extremes their own top level.
• Choice: give the full option list; add other/none escape hatch; describe options so they separate. Confusable options → objects {what, not_for, examples}.
• Structure (any field): put the question in one key and data in others, reference by backtick name. Field names are free and seen by the model:

  "instructions": {
    "potential_duplicate": {"name": "Jon Smith", "location": "Oakland, CA"},
    "question": "Is the resume for the same person as `potential_duplicate`?"
  }
  

Known failure modes (Jev 1.13; assume similar for local models until tested)

Failure	Do instead
Literal reading	State exact condition; boundary cases in criteria
Counting, math, hex/RGB, interpolating scores	Do it in code; ask one noul per item and sum
Date comparison/arithmetic	Extract parts via Choices, compute in code
Indirection / double negatives	Direct wording; name the state field
Large irrelevant state	Filter first
Adversarial/injected state	Precise criteria; test edge cases
instructions vs criteria contradict	Align them; never map true to "no"
First-option bias in Choice	Reorder options, check consistency
Generation	Use an LLM; let System One pick among candidates

English is strongest; other languages work with lower accuracy.

Ollama errors

400 invalid request / unsupported model or runner / prompt exceeds context / cloud model · 404 model not pulled (ollama pull nimble) · 413 body over 64 KiB (32 MiB with images) · 500 load/render/scoring failed. Body: {"error": "…"}. No streaming, tools, video, or generation options. Input never truncated. Requires GGUF weights; MLX/Safetensors and cloud models unsupported. Not on Ollama Cloud.

Ollama models (from ollama.com/library, read 2026-10-10)

Model	Maker	Size	Input	Effective prompt limit	Notes
nimble (9b)	Bespoke Labs, from Qwen3.5-9B	9.3–9.5 GB	text	8,192 tokens (full state + all questions per question)	<100 ms on M5 Max; contrastive-pair training
clef-flash (9b)	Cloudflare, from Qwen3.5-9B	11–12 GB	text + image	64K	38.8 ms median, lowest latency in Cloudflare's runs
clef (27b)	Cloudflare, from Qwen3.8-27B	18 GB	text + image	64K	~5× slower than clef-flash, "more accurate"
Tev1 (4B, 0.8B)	Together AI	—	text	—	73.3% / 63.5% on Bespoke's suite

Model pages say Choice/Score take 2–26 options (API schema says 255 — trust the model page). Ties follow option order: put the preferred option first.

Vendor benchmarks (no shared head-to-head between nimble and clef-flash; each only vs Jev, different suites):

• Nimble vs Jev 1.13 (Bespoke, 13 public sets, 3,880 decisions): macro 74.8 vs 76.0; Choice 81.6 vs 82.9; Boolean 80.2 vs 84.6; Score 54.6 vs 50.1.
• Clef-flash / Clef / Jev (Cloudflare): BFCL 98.8/98.5/95.8; API-Bank 93.1/91.9/88.2; BANKING77 90.9/94.2/79.7; CLINC150+OOS 66.8/97.4/89.3; Home appliance 97.7/83.0/52.3; ANLI 59.1/69.8/74.8; RouterBench 79.9/79.7/79.9; When2Call 65.6/72.4/81.0; latency median 38.8/209.3/524.1 ms.

Measured (2026-10-10, nimble q8_0, Ollama 0.40.2, CPU-only laptop: size_vram 0)

• 36 KB document as state → 400 {"error":"prompt 0 has 11282 tokens; expected 1–8194 (input is never truncated)"}. Limit is per question prompt (state + questions).
• 17 KB (~5k-token) state + 5 questions → OK, but usage.input_tokens 26,657 (state re-scored per question) and 327 s wall time on CPU.
• Real CPU throughput ≈ 18–20 uncached input tokens/s (604-token section + 1 noul: 32 s warm; 698 tokens: 40 s incl. load). A 5-question run on a 5k-token document fits this too (6k unique tokens / 327 s): usage.input_tokens multiplies the state per question, but the shared state prefix seems to be computed once (prompt cache). Cost ≈ state once + each question, not state × questions.
• ~400-word state + 10 easy factual questions (warm): 10/10, 103 s, usage.input_tokens 13,647 (≈ state × 10). ~8–10 s per extra question on CPU: cheaper than re-sending the state, not free.
• Same state, 10 hard questions: 7/10, 70.7 s (state prefix cached). Failed exactly the documented weak spots: date math ("published more than one year ago?" → 0.41, wrong), double negative ("is it false that X never uses Y?" → 0.10, wrong), arithmetic (monthly price × 12 > threshold? → 0.32, wrong). Passed: counting 3 named items (0.99), multi-hop match, "not stated" trap, close options (0.89, conf 0.67), counting 4 items (> 3? → 0.61, right but unsure). Wrong answers sat at 0.1–0.4, not near 0: middling noul = warning sign. Keep math, dates, and negation in code.
• Harder non-math set (injection, misbinding, tense, NOT-choice, quantifier): 9/10. Miss: "does the state say every project uses language X?" → 0.96 yes (one project's language unstated; overgeneralized). Quantifiers ("every", "all") are a weak spot.
• Ollama rejects object criteria: 400 {"error":"question \"x\": choice criteria must map option keys to descriptions or null"}. Flatten {what, not_for, examples} into one string ("what. Not for: … Examples: a; b"). Structured instructions objects work ({"claim": …, "question": "Is claimsupported byproduct?"}).
• E-commerce product state (nested JSON, ~1k tokens) + 11 structured questions: 11/11, 119.5 s. Passed: care-label misuse vs defect, sarcasm → angry, injection in seller_notes ignored, marketing "pure merino" vs 87% materials, variant pick (size + dark + in stock), health claim, EU→Germany, Kids listing ≠ adult listing.
• Product weak-spot run (12): date and sum-over-variants math wrong (0.28, 0.43); percent-off and double negative right only at ~0.55. Confidence on a wrong Choice can be 0.99 when the question word is ambiguous ("offered" vs "in stock").
• Support-ticket run (13): the one wrong route came back at confidence 0.32; all correct answers ≥ 0.76. Confidence gating works here — gate Choices below ~0.6.
• An identical repeat request returned in 0.48 s (prompt cache) — never benchmark with a repeated request.
• Tiny request: 10.5 s cold, 0.48 s warm. The "<100 ms" claim assumes a GPU (M5 Max). Check GET /api/ps → size_vram before blaming the model.
• Literal reading confirmed: the document said readers "already know the basics", the choice option said "new to the topic" → nimble picked advanced (0.98). Write options that match the text's own wording.

TypeSafe SDK against Ollama (officially documented on ollama.com)

pip install typesafe-sdk
export TYPESAFE_BASE_URL=http://localhost:11434
export TYPESAFE_API_KEY=ollama          # required by SDK, ignored by Ollama
export TYPESAFE_DEFAULT_MODEL=nimble

Use TypeSafeClient(timeout=120) — local first load is slow. Decision models are not in the Ollama CLI or the ollama Python/JS libraries yet (model pages show ollama.systemone(...) snippets but say to use the API or TypeSafe SDK for now).

Unverified

• Whether Jev itself runs on Ollama (not listed).