tachyone-multi (System One decision engine)

Status: released (v0.8.0), revision 2026-10-02 — English 1c88ebef (B-13 mixture + P3 evidence confidence), multilingual 3693def1 (card). Trained on a single RTX 3060 12GB (B-5b and B-13 on NVIDIA L4 23GB) and published as LoRA adapters (munod/tachyone-en, munod/tachyone-multi); measured numbers below come from benchmarks/report.md.

This revision adds B-13 for the English checkpoint: the mixture retrain. Same recipe and seed as the B-5 run, one factor changed — the training data: 35,540 records (the 21,000 five-domain records plus 11,000 pinned public records — MultiNLI, BoolQ, Banking77, licences in training/data/jev_sources.lock.json — and 3,540 executable-rule-tree family records), then the identical frozen-trunk choice-bank fit. It sweeps both English splits 0.964 → 1.000 (gate strict 1.000; unseen-text rows 0.9987) — and, measured against a fresh same-recipe control that moves +0.6, it lifts Intelligence on the 231 public JevBench items (evaluation-only, never trained on) from 8.4 to 15.0. The probes below are the honest external check: XNLI 0.341 → 0.566, typed-decisions 0.269 → 0.367.

This revision is B-5b: the multilingual checkpoint now covers five domains. Its trunk is joint-trained on 30,000 multilingual five-domain records (support keeps its 18,000; four new domains 3,000 each) and its choice heads were re-fitted on the frozen trunk by training/fit_choice_bank.py — the same construction as the English artifact, recipe shipped next to the weights as choice_bank_fit.json. On the five-domain split it scores 0.9975 (previous adapter zero-shot: 0.561), gate strict 1.000, all six languages at ECE ≤ 0.004.

Labels (B-11 + B-12, ADR-0014/ADR-0015): every label in all three primitives is derived from the text it accompanies — noul from its phrase bank (request → 1, neutral/empty → 0), score from the tone's level (empty → middle), choice from the option the state names (empty → the catch-all other). All datasets were regenerated and the label audit published with the evaluation reports 0 contradictory rows. These numbers are not comparable with pre-B-11/pre-B-12 measurements: the old labels contradicted 121 of 241 request-toned English rows, left every noul label in es/de/nl at 0, and gave 7.8% of score rows a "near-tie" the text never showed.

Provenance, stated plainly: each adapter carries its recipe next to the weights (finetune_config.json / choice_bank_fit.json) — the English adapter is the B-13 mixture bank over the B-13 joint run's frozen trunk; the multilingual adapter is the B-5b joint run plus the frozen-trunk head fit. Both heads-first constructions exist because training the corrected labels makes noul+score trivial and costs choice (an identical-recipe control landed 13 points lower, L-006) — the fix that worked was re-fitting the head on a trunk that already knew the data.

Model details

• Developed by: The Tachyone Authors.
• Model type: non-autoregressive encoder with three task distributions (noul, choice, score), answering typed questions about a state in one forward pass.
• Trunk: ModernBERT-large (English) and mmBERT-base (100+ languages); see ADR-0007.
• Adapters: munod/tachyone-en, munod/tachyone-multi (LoRA; load base + adapter).
• License: Apache-2.0.
• Repository: https://github.com/munod/tachyone

Uses

Tachyone answers atomic choice / score / noul questions about a state and returns typed values with probabilities and confidence. It speaks the TypeSafe Jev /v1/systemone wire protocol as a drop-in and runs locally/offline with no API key. Compose several atomic answers in code rather than asking one broad question.

Out of scope: free-form text generation, multi-step reasoning, and any decision requiring extended deliberation — decompose those into atomic questions and combine results in code.

Bias, risks, and limitations

• Probabilities are only meaningful after calibration; the shipped temperature must be applied (see docs/training.md).
• Synthetic training data can inherit generator biases; public probes are evaluation-only.
• Confidence is a property of the distribution, not a guarantee of correctness.

Training

Deterministic synthetic JSONL (training/generate_data.py) supervised with an RLCD proper-scoring objective (training/finetune_rlcd.py), then temperature-calibrated on a held-out split (training/fit_calibration.py). Configs and seed live under training/configs/.

Evaluation

Reported by training/evaluate.py and rendered by benchmarks/report.py (accuracy, ECE, p50/p95 latency per primitive and language).

Full-scale run (single RTX 3060 12GB; B-5b on a single L4 23GB): 21,000 five-domain English / 30,000 five-domain multilingual / 18,000 support-only multilingual train / 1,500–7,500 eval deterministic synthetic records (fully localized per language — all five domains now ship the seven training languages, a learnable other option with rich descriptions, per-record RNG, one-in-six distractor clauses), LoRA (r=16 English, r=64 multilingual) plus a dedicated low-rank choice head (near-identity init); 8 epochs for multilingual, batch 16, bf16 + gradient checkpointing. Both published artifacts are bank-over-frozen-trunk: the trunk is trained jointly, then training/fit_choice_bank.py re-fits the choice heads on cached encodings (rank 128, 8 epochs at lr 1e-4), shipping its recipe as choice_bank_fit.json.

Checkpoint	Split	Accuracy	ECE (calibrated)	p50 (ms)
English (ModernBERT-large + five-domain LoRA r=16 + choice-head bank)	support	1.000	0.006	58.0
English, five-domain split	5 domains	1.000	0.009	26.3
Multilingual (mmBERT-base + LoRA r=64 + fitted bank) — routed runtime	support	0.895	0.062	46.9
Multilingual, five-domain split	5 domains	0.9975	0.0014	18.6

Per primitive (English): choice 1.000, noul 1.000, score 1.000; (multilingual, five-domain): choice 0.9924, noul 1.000, score 1.000. Label audit: every noul row is judged against its own text — 0 contradictory in every eval set (positive rates 0.472–0.486), per language in benchmarks/report.md.

Confidence comes from evidence (P3, 2026-10-02). The English adapter ships two extra assets: state_prototypes.json (K=32 centroids of its own 21,000 training states) and confidence_calibration.json (that bank plus a fitted noul map). A noul answer keeps its direction — the cosine still decides yes/no — and takes its magnitude from strength = max_k cos(state, centroid), clamped to [0.5, 1]; choice runs at T=1.5 and score is pinned at 0.1 because its answer is the expected value. Consequences, all measured: accuracy is byte-identical to the previous revision everywhere (argmax and expected value never move), in-domain ECE reads 0.006 / 0.009 instead of an in-sample 0.000 (L-015 — the old number was fitted on the very rows it scored), and off-domain confidence drops to match what the model actually knows: XNLI raw ECE 0.426 → 0.264, typed-decisions 0.578 → 0.416, MASSIVE 0.583 → 0.296, JevBench's Calibration axis 0.0 → 52.6 on the 231 public items (its recorded gate of 60 was not met — .specs/features/jevbench/spec.md P3 results records why).

What the English B-13 row is (in-sample, stated plainly). Both English splits read 1.000 because the held-out synthetic sets share rows with training data (the known 81% choice row collision) — saturation here is a ceiling, not a claim: accuracy on the 791 rows whose text never occurs in training is 0.9987, and the external check is the public one — the 231 JevBench items move Intelligence 8.4 → 15.0 against a fresh control's +0.6 (attribution table in .specs/features/jevbench/tasks.md JB-7), with XNLI 0.566 and typed-decisions 0.367 as the probe deltas. The B-5 trade this row replaces is closed: the previous bank won choice 0.946 → 1.000 at the cost of noul/score (0.992/0.978 → 0.946/0.946) on support; B-13 posts 1.000 across all three primitives on both splits, gate strict 1.000 (0 to shared, 0 wrong domain over 2,500 choice rows). Full tables: docs/benchmarks.md.

The multilingual B-5b gates (fixed before training, all pass). The previous adapter measured 0.561 zero-shot on the new five-domain split (worst new domain agent_tools 0.438, support 0.8413) — the gates were support ≥ 0.8413, worst new domain ≥ 0.70, per-domain ECE ≤ 0.05 (exceptions declared), gate strict published. The published artifact posts support 1.0000, worst new domain 0.993, per-domain ECE 0.0005–0.0038 (zero exceptions) and gate strict 1.000 over 2,500 choice rows; accuracy on the 2,996 rows whose text never occurs in training is 0.996. Per-language ECE on that split is 0.0010–0.0041 — all six languages ≤ 0.05 for the first time on a five-domain set; on the support-only routed split (13–15% of rows fall through to the English checkpoint by language routing, .specs/project/BACKLOG.md L-013) three of six languages remain above the 0.05 target — nl accuracy 0.715 / ECE 0.167, de 0.083, it 0.056 — open under NFR-C06 / BACKLOG B-1. The CUDA-graph fast path (TACHYONE_FAST=1) gives 2.54× p50 on English (17.31 → 6.81 ms) and 3.66× on the multilingual five-domain path (13.97 → 3.82 ms), both with 0 top-label flips — the gate and the keyed heads execute after the encode, which the graphed path never sees.

Robustness (B-4). On a noisy view (one surface edit — typo/accents/casing — applied to 15% of states) English drops only 1.000 → 0.997 and multilingual 0.743 → 0.744, so the released adapters are robust to this noise model.

Full tables and environment are in benchmarks/report.md.

Citation

@misc{tachyone2026,
  title        = {tachyone: a local-first System One decision engine},
  author       = {The tachyone Authors},
  year         = {2026},
  howpublished = {\url{https://github.com/munod/tachyone}}
}