AnyJev

• New, opt-in: L0 can read as many option rotations as the decision needs instead of K, with the rotations turning a canonical listing. Opt-in rather than default only because every table in docs/ was measured before it existed; the defaults flip in the release that regenerates them. Decider(adaptive_shifts=True) is now the default and the stopping rule compares the log-odds margin of the running marginal (adaptive_stat="logit") against a threshold with a stated meaning: Decider.calibrate_adaptive(question, states, target=0.01) reads every rotation of a batch of unlabelled states once and returns the cheapest threshold whose disagreement with the full-K answer is under target by a Clopper-Pearson upper bound (anyjev/calibrate/stopping.py). The reference is our own full-strength readout, never a label, so the guarantee costs only calibration compute. Before a question is calibrated the threshold is DEFAULT_LOG_MARGIN (8.5), the smallest value that certified 1% on four (model, task) cells at once. Measured on Qwen2.5-7B and Qwen3-8B over massive_route (K=18) and newsgroups (K=20), 900 states each: the uncalibrated default reads 5.1-9.7 rotations (1.9x-3.9x), per-question calibration reads 4.8-6.1 (2.9x-4.2x) with held-out disagreement 0.000-0.008, and accuracy is within 0.5 points of the full cycle either way. End to end at a certified 1% target on 300 states: 2.22x the decisions per second on a vLLM server and 2.3x-2.7x on the transformers backend (two runs; the adaptive rows are stable and the full-cycle reference varies, so the ratio inherits its variance — docs/rotation_budget.md states both), 7.2 requests per decision instead of 18, agreement 0.987, accuracy 0.703 against 0.697 (bench/results_layout/2026-09-27/, docs/rotation_budget.md, research log entries 23 and 24). Decider(adaptive_wave=n) asks for n rotations per backend call; the default is 2 because a round is a barrier the whole batch waits behind -- on vLLM, waves of 1 and 2 issue the same requests and differ by 1.7x in wall clock, while locally they tie. adaptive_shifts=False reads every rotation.
• Decider(canonical_order=True), also new, is what makes stopping early safe, so turn the two on together. The rotations now turn a listing of the options ordered by their text rather than the caller's listing. Reading every rotation already gave every option every position, but which options sit next to each other still followed the caller's order and a partial budget does not cancel the position bias either, so the decision could depend on how the list was typed. With it the prompts are a function of the option set: any two listings of the same options return identical probabilities, at any rotation budget -- a stronger guarantee than the full cycle used to give. The statistic had to change for the same reason the old margin never bound: a probability gap saturates at 1, which is why docs/results_adaptive.md found margins of 0.05, 0.1 and 0.2 indistinguishable, and why on two of the four cells above no probability-gap threshold could certify a 1% rate at all. adaptive_stat="prob" restores the pre-0.2 rule, adaptive_margin pins a threshold by hand, canonical_order=False restores pre-0.2 prompts. Tests in tests/test_stopping.py, including that a calibration set too small to certify a target returns no threshold and reads every rotation rather than pretending.
• Every result JSON now records the readout it was produced with (env.readout, from bench.run.readout_settings). A benchmark that inherits the library's defaults is fine; one that does not record them is not, because changing a default then moves every published row with nothing in the JSON to show it. This change is exactly that case, which is why the record is centralised in environment() rather than pinned per harness.
• Fix: Decider.load_artifact dispatched on the artifact's

Version 3 of the method: a closed-form head at a fixed depth, routing, label-free adaptation, a packaged demo, and a tree that carries only the shipped code and the result JSON a doc cites (docs/migration_v3.md lists everything removed).

• L2: closed-form per-question heads on the hidden state, at two thirds of the depth. Decider.fit_head(question, states, labels) (or calibrate(..., level="L2")) runs one forward of the labelled states and solves a shrunk-LDA or ridge head on the last-position hidden state; the block it reads, the head kind and the temperature are chosen by out-of-fold NLL. decide(..., level="L2") then sends one prompt per state through a forward that stops at that block (HFBackend.hidden_states_to, bit-exact against the plain forward in fp32, scripts/exit_parity.py) and returns calibrated probabilities with blocks_executed in the diagnostics; require="L2" is enforced like the other levels. Artifacts round-trip through export_artifacts / load_artifacts / load_artifact (refused on another model or question layout). On LocalLLaMA/typed-decisions with 300 labels per question: Qwen3-1.7B 0.730 at 64% depth, Qwen3-4B 0.786 at 67%, Qwen3-8B 0.771 at 67% (L1: 0.499 / 0.567 / 0.648; Jev published 0.727), pooled ECE 0.03-0.05, at 0.68x the batched time of one plain forward on the 8B; Qwen3-30B-A3B 0.799 at 83%, Qwen3-32B 0.798 at 81% (docs/results_exit.md, from bench/results_exit/2026-09-22/). Needs the transformers backend. docs/method_v3.md, docs/jev_mode.md, docs/levels.md, docs/research_log.md.
• L2 follows its question across wordings and listing orders. Decider.route(q) serves a question from its exact head, else from the head of the same kind and option texts under another wording, else (heads fit on random listing orders) from the head of the same option set in another order; Decider(adapt="routed") (default) re-estimates a routed head's feature mean and scale from the unlabelled states the question is asked on (30 by default, the current batch included), which on Qwen3-8B takes a reworded question from 0.65-0.70 (head as is) to 0.74-0.75 with 30 unlabelled states (0.74-0.76 with 300) against 0.77 for a labelled refit (bench/results_paraphrase/2026-09-22/Qwen__Qwen3-8B.paraphrase.b24.json; the 4B goes 0.63-0.68 to 0.74-0.75, Qwen__Qwen3-4B.paraphrase.b24.json; research log entry 13); fit_head(listing="auto") fits on random listing orders up to 8 options (flip under reversal 0.07 instead of 0.18, <model>.order.json) and on the canonical order above (random listings cost the 20-way bench heads 6-18 points in the artifact builds, entry 13); level="auto" picks L2 / L1 / L0 per question. Diagnostics: routed_from, reordered, adapted, adapt_n, listing.
• Closed-form distillation: bench.synth_states writes new cases per workflow from real exemplars (Qwen3-8B, thinking off), bench.distill_heads labels them with a teacher's shipped heads and solves student heads on gold plus teacher-labelled states (hard or soft targets); the 32B's heads lift the Qwen3-1.7B from 0.730 to 0.760 on typed-decisions (95% CI on the gain [+0.015, +0.049]), the 4B and 8B do not move (bench/results_distill/2026-09-22/Qwen__Qwen3-32B.distill_heads.fit.json, entry 14).
• Shipped heads: scripts/build_heads.py fits and validates a model's heads for the 20 typed questions and the three bench tasks through the user path and writes anyjev-heads/<model>.json (with per-question validation, the agreement with the study caches, and env). The head arrays are stored as base64 float32 (anyjev.heads.encode_array, exact, about ten times smaller than number lists): 1.8-4.4 MB per model, 13.6 MB for the five; LinearHead.from_dict and load_artifacts read the list format too, and scripts/compact_heads.py converts in place after a bit-exact round-trip check.
• Demo: python -m demo.jev_mode (demo/jev_mode.py; not in the wheel). --backend fake runs in seconds on CPU with no downloa

• README GIF: scripts/find_flip_example.py finds real items where raw readout flips under option reversal and L0 does not; scripts/make_flip_gif.py renders them (assets/flip.gif, Qwen3-8B on BANKING77). space/ holds the Hugging Face Space entry point.

• Decider.save_artifacts / load_artifacts: L1 artifacts as one JSON file per model, refused on model mismatch.

• ROADMAP.md, CONTRIBUTING.md, CREDITS.md, GitHub Actions CI (ruff + pytest on 3.10 and 3.12).

• Core: Question (choice / score / noul), Decider, Decision with mandatory level.

• Readout: chat-template prompts, single-token label mapping with collision and multi-token errors.

• L0: cyclic-shift permutation marginalization (log-space combine by default) and label-free prior correction: batch calibration by default (low variance in the bench), content-free contextual calibration opt-in (high variance: large wins on one noul task, losses on ordinal scores).

• Readout: noul label tokens stay bound to their option when the phrasing order is swapped (a position-bound readout silently swapped Yes/No in the second phrasing).

• L1: temperature scaling artifacts keyed by (model, question).

• HFBackend accepts revision.

• vLLM backend parity check (scripts/vllm_parity.py): 15/15 argmax agreement with transformers on Qwen2.5-7B-Instruct; restricted log-softmax differs by up to 0.6 nats only on labels below 1e-7 probability.

• Backends: transformers (HFBackend), vLLM OpenAI-compatible server (VLLMBackend, allowed_token_ids + logprobs), synthetic biased model (FakeBackend) for tests.

• Question.score accepts explicit ordered levels; value is the expected level index.

• Bench: bench.providers.nanojev_maze runs AnyJev as the engine inside NanoJev's frozen maze exploration harness; bench.providers.nanojev_native_maze reruns their untuned-Qwen baseline for an apples-to-apples row.

• Bench: bench.run_typed runs LocalLLaMA/typed-decisions (the set Laya and Jev report on) with overall / per-workflow / per-type metrics; bench.providers.laya runs Laya checkpoints on the same decisions.

• Bench: bench.table aggregates result JSONs into one table; runner saves after every task and merges per task across invocations.

• Bench: metrics (accuracy, macro-F1, Brier, NLL, ECE, flip rate, coverage-risk), tasks newsgroups, injection, banking20, runner python -m bench.run.

pip install "anyjev[hf]"