Upload README.md with huggingface_hub

README.md

@@ -77,34 +77,50 @@ split (`training/fit_calibration.py`). Configs and seed live under `training/con
 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):** 9,000 English / 18,000 multilingual train / 1,500 eval
-deterministic synthetic records (fully localized per language, a learnable `other` team 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 (r=32, near-identity init); 4 epochs for English and
-**8 for multilingual**, batch 16, bf16 + gradient checkpointing.
+**Full-scale run (single RTX 3060 12GB):** 21,000 five-domain English / 18,000 multilingual train
+/ 1,500 eval deterministic synthetic records (fully localized per language, 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. The **English artifact** is the B-5 bank
+(ADR-0016): run 5's six-epoch trunk kept **frozen** while the `choice` head was re-fitted as a
+bank — shared + one head per domain, rank 128, 8 epochs at lr 1e-4 — by
+`training/fit_choice_bank.py`, whose recipe ships next to the weights as `choice_bank_fit.json`.
 
 | Checkpoint | Accuracy | ECE (calibrated) | p50 (ms) |
 | --- | --- | --- | --- |
-| English (ModernBERT-large + LoRA r=16 + choice head) | **0.972** | 0.020 | 23.1 |
+| English (ModernBERT-large + five-domain LoRA r=16 + choice-head bank) | **0.964** | 0.023 | 54.4 |
 | Multilingual (mmBERT-base + LoRA r=64 + choice head) | **0.743** | 0.089 | 16.0 |
 
-Per primitive (English): `choice` 0.946, `noul` 0.992, `score` 0.978; (multilingual): `choice`
+Per primitive (English): `choice` **1.000**, `noul` 0.946, `score` 0.946; (multilingual): `choice`
 0.468, `noul` 0.892, `score` 0.870. **Label audit:** every `noul` row is judged against its own
 text — **0 contradictory** in both eval sets (positive rates 0.482 / 0.486), per language in
 `benchmarks/report.md`.
 
+**What this English row trades (published in full, not summarized away).** On the support-only
+split the previous artifact scored 0.972 — `noul` 0.992, `score` 0.978, `choice` 0.946. The
+five-domain bank scores 0.964 there: `choice` becomes **1.000** while `noul`/`score` give up 4.6
+and 3.2 points, and in exchange the adapter covers four domains it could not answer at all
+before — on the five-domain split the previous artifact scores **0.511** overall (worst domain
+0.328) against this one's **0.964** (worst domain 0.963). The gate that routes each `choice`
+question to its domain head scores **strict 1.000** (0 to shared, 0 wrong domain) over 2,500 rows,
+and accuracy on the 473 rows whose text never occurs in training is **0.998**. Full tables:
+[`docs/benchmarks.md`](https://github.com/munod/tachyone/blob/main/docs/benchmarks.md).
+
 `choice` is the weak primitive on the multilingual side (0.468) and it is a *training* trade, not
 a label problem: `choice` labels never changed, and retraining on the corrected labels makes
 `noul` (1.000) and `score` (0.998) trivial — the same tone detector — while the shared trunk
 starves `choice` (an identical-recipe English control landed at 0.841, the five-domain retrain's
-`choice` collapsed to 0.303). That is why the published English adapter keeps its B-11 weights
-(provenance stated in the header). One of six languages meets ECE ≤ 0.05 (`es` 0.045); `pt` 0.062,
+`choice` collapsed to 0.303). The English side did **not** retrain the trunk to escape that: it
+re-fitted the `choice` head on a **frozen** trunk (B-5 / ADR-0016), which lifts `choice` to 1.000
+and leaves `noul`/`score` where the trunk already had them — the cost and the gain in the table
+above are one artifact, not two runs (provenance: `choice_bank_fit.json` next to the weights).
+One of six languages meets ECE ≤ 0.05 (`es` 0.045); `pt` 0.062,
 `fr` 0.101, `de` 0.118, `nl` 0.192 and `it` 0.197 remain above target (NFR-C06), and multilingual
 `choice`/`noul` ECE (0.099 / 0.108) is declared with them. The CUDA-graph fast path
 (`TACHYONE_FAST=1`) gives a 2.65× p50 speedup (8.97 → 3.39 ms) with 0 top-label flips.
 
 **Robustness (B-4).** On a noisy view (one surface edit — typo/accents/casing — applied to 15% of
-states) English drops only 0.972 → 0.969 and multilingual 0.743 → 0.744, so the released
+states) English drops only 0.964 → 0.963 and multilingual 0.743 → 0.744, so the released
 adapters are robust to this noise model.
 
 Full tables and environment are in