[iamwavecut]

Closes #64.

What

POST /v1/systemone (TypeSafe's System One API, in the shape llama.cpp, vLLM and Ollama serve) and POST /v1/decisions (OpenAI's Decisions API, text only) answer choice, noul/predicate and score questions about a state with the served chat model, no training and no startup option.

• Each question is one read: a fixed system turn asking for one label, then a user turn with the state, the question and its options under one-token labels (A.. for a choice, yes/no for a noul, 0.. for a score), thinking off, one output token, greedy, no penalties, no graft.
• The answer comes from the first reply token's exact log probabilities of the labels: a new Op readout_logprobs (l[id] - logsumexp(l) over the public vocabulary, one CTA, FP64 oracle test) runs on the head's logits in the Qwen3.5 core's prefill (both the chunked card path and sample_from_hidden), carried through ExecutionOptions::readout_tokens → plan → round buffers → RawTokenLogprob::readout_values → TokenLogprob::readout. Flash-Next and Gemma 4 leave it empty and the serve layer reads the labels from the token's top-20 alternatives (diagnostics.<id>.readout = "top_k", a missing label bounded by the least likely alternative).
• Probabilities are the softmax over the options (a noul's yes+Yes and no+No added up); confidence is the winner's probability over the whole vocabulary; diagnostics.<id> carries label_mass, argmax_is_label and readout. score is the probability-weighted level index with a legend.
• Strict parsing as vLLM: unknown fields, joint, thinking on → 400; images/files/audio/videos → 501; 1–64 questions, 2–26 options, 2–10 levels; a label that is not one tokenizer token → 400.
• The reads run a lane's worth (--max-concurrency) at a time, so other clients get the queue between windows, and every read starts with the state, which the prefix cache prefills once.
• Request log protocols systemone / openai_decisions; GET /v1 index entries; docs in docs/serving.md (Typed decisions), README and the documentation map.

Per-request calibration (temperature, permutations) is deferred by the user's decision.

Verification

Vast RTX 3090 (EPYC 7313, driver 595, CUDA 12.8 build, sm_86), Ternary Bonsai 2 27B, --max-concurrency 4, scripts/pods/decision_checks.sh + decision_bench.py (added):

• ninfer_readout_logprobs_test (full 248320-row and 151936-row vocabularies, 1–32 ids, uniform and extreme logits, untouched slots past count, validation), ninfer_target_logprobs_test, ninfer_logprob_topk_test, ninfer_systemone_test (parsing of both dialects, labels, prompts, reading, responses), ninfer_rerank_test: pass.
• Agreement: a choice's readout probabilities times label_mass equal the chat endpoint's top-20 log probabilities of the same prompt to 1e-5 (billing 0.71710 vs 0.71710; label mass 0.98292 vs 0.98292).
• Both endpoints give the same choice probabilities for the same question; shapes, sums to 1, argmax = choice, expected score, 400/501 refusals checked.
• Latency (wall, median of 3 warm = state cached; cold = state prefilled first):

state	questions	cold	warm	per question
370 tok	1	0.15 s	17 ms	17 ms
370 tok	4	0.60 s	62 ms	15 ms
370 tok	16	0.78 s	235 ms	15 ms
13.5K tok	1	5.3 s	39 ms	39 ms
13.5K tok	4	5.8 s	91 ms	23 ms
13.5K tok	16	6.1 s	325 ms	20 ms

• Concurrency: 16 questions (2K-token cached state) take 0.26 s alone and 0.42 s beside a streaming 256-token chat completion; the chat decodes at 83.5 tok/s instead of 89.7 tok/s meanwhile.
• Prompt tuning (chat top-20 probes on the same model): the user turn alone left the labels 60–90% of the first token's probability ("The", "Based" took the rest, and at 13.5K tokens "The" was the argmax for a choice); the system turn raises it to 98–99% for every type and length, and "in lower case" moves a noul's mass from Yes to yes. Both cases are read anyway.
• Mac: host-only syntax check of every changed C++ file and the serve test built and run with clang; git diff --check.

Not verified / limits

• The top-20 fallback (Flash-Next, Gemma 4) is covered by the unit test only; no GPU run on those models.
• The head's logits are BF16: two options scored within one BF16 step share one probability (seen as A = C = 0.2884 before tuning). Documented.
• Probabilities are a general chat model's under this prompt; not calibrated.
• CI was not waited for; the CUDA build and tests ran on the rented 3090 (GCC 13), the host checks under Apple clang.

Rental: 1.0 h at $0.30/h on Vast (RunPod had no 3090/4090 community instance with a public IP); the instance is destroyed.