model card: composition A (v6 head cat/dec + v4 adapter urgency) = 0.6923 on the ruler

README.md

@@ -49,7 +49,8 @@ per case (joint-schema head).
 
 ## What it is (honestly)
 
-Veredicto was trained as an end-to-end classifier for the three questions
+The legacy v4/v5 pipeline trained Veredicto as an end-to-end classifier
+for the three questions
 (category 17-way, decision 5-way, urgency 4-way). Two independent training
 runs (v4: mixed-quality data, v5: clean deduplicated data) produced the
 same profile: excellent urgency, failed category/decision. On our private
@@ -63,30 +64,73 @@ human-curated ruler (1,499 real cases, seed-42 split, temperature-scaled):
 | overall | 0.3838 | 0.3316 | 0.6233 | **0.7127** |
 
 \* Same ruler, same seed-42 test split, same protocol (n=753 cases).
-v5 = retrained on clean deduplicated data (2 epochs) — category got WORSE,
-confirming the architecture limitation is real, not a data quality issue.
+v5 = retrained on clean deduplicated data (2 epochs) — category got worse.
+These historical scores do not isolate architecture from data or pipeline
+defects. An audit of the observed v4/v5 pipeline copies found targets
+misaligned with reordered `option_ids`, a trained head omitted from
+adapter saves, and hard-label category cases omitted when `probs_alvo`
+was absent. The copies' identity with the remote execution was not
+verified, and the historical head was not recovered; each defect's
+contribution to these scores remains unquantified.
 
 **On urgency, the student beats its own benchmark**: v4 0.745 vs JEV 0.724
 (+2.1 pp) — a 9B model specialized on PT-BR support data outperforms the
-frontier decision API on the question where it was closest. This result is
-**robust**: both v4 and v5 (independent training runs) score 0.73-0.75 on
-urgency, far above any Decisum (ceiling 0.679).
-
-Category and decision did not transfer at any scale or data quality — this
-is a fundamental limitation of the joint-schema prompt format for
-fine-grained PT-BR classification, not a training issue.
+frontier decision API on the question where it was closest. Both
+historical v4 and v5 evaluations score about 0.73-0.75 on urgency; the
+later documented Decisum v10+swa reference reaches 0.6919. These are
+historical point comparisons, without a paired confidence interval.
+
+**Local update — 20 October 2026:** Veredicto v6 A1 trained `head_only`
+with a frozen backbone. Its complete checkpoint `step-2208`, selected
+on dev, scored **66.1355% overall, 68.2603% category, 64.9402% decision
+and 65.2058% urgency** on 753 seed-42 test cases, at T=1 with no fitting
+or calibration on the gold set. This refutes the absolute claim that
+the architecture cannot handle category/decision, without isolating
+each correction's effect. The v6 checkpoint is **not yet published**;
+it is distinct from the legacy v4/v5 LoRA adapter described in this
+card, whose identity and historical scores are retained. Urgency
+regresses against Decisum v10+swa's documented 69.19%. **No production
+promotion or switch was made.** Source:
+[local v6 final report](../veredicto_v6/resultado_final.md).
+
+**Second update — 20 October 2026 (evening):** a zero-training
+composition of the two artifacts above was measured on the same
+official ruler (same 753 seed-42 test cases, T=1, paired per-case
+predictions): the **v6 head answers category and decision**, the **v4
+adapter in this repo answers urgency**. The composition scores **0.6923
+overall** (n=2,259 answers):
+
+| system | category | decision | urgency | overall |
+|---|---|---|---|---|
+| JEV (frontier API) | **0.7251** | **0.6892** | 0.7238 | **0.7127** |
+| **A: v6 head (cat/dec) + v4 adapter (urgency)** | 0.6826 | 0.6494 | **0.7450** | 0.6923 |
+| Veredicto v6 A1 alone | 0.6826 | 0.6494 | 0.6521 | 0.6614 |
+| Decisum v10+swa (champion) | 0.5618 | 0.6481 | 0.6919 | 0.6339 |
+| Veredicto v4 adapter alone (this repo) | 0.2430 | 0.1633 | **0.7450** | 0.3838 |
+
+That is **+5.8 pp over the production champion** and 2.0 pp below the
+frontier teacher — the best local number ever measured on the ruler,
+with each question answered by the artifact that historically masters
+it. Per-subset: human-adjudicated hard core 0.6693, random slice 0.7370.
+A 0.75/0.25 teacher/student blend on urgency reaches 0.6932 overall but
+was selected on the test split — reported as diagnostic only, not as
+the headline. No new training was run and **no production change was
+made**; the serving question (whether the two heads can share one
+backbone instance, given the v6 head was trained on the frozen base and
+the urgency head operates over the LoRA-adapted one) is under separate
+evaluation.
 
 **On urgency, the student beats its own teacher**: 0.745 vs 0.724 (+2.1 pp)
 — a 9B distillation of the decision API, specialized on PT-BR support
 data, outperforms the frontier API on the question where it was closest.
-Category and decision did not transfer at this scale (see below), which is
-exactly why we publish it as a **specialist**:
+The legacy evaluations reported weak category/decision scores, so the
+legacy adapter was published as a **specialist**:
 
 - as a **urgency specialist** (answers one question better than anything
   else we measured, including the teacher);
 - as a **distillation teacher for urgency** into lighter models;
-- as a reproducible data point on what scale does and does not buy on
-  typed-decision tasks.
+- as a historical result whose interpretation must account for the
+  audited pipeline defects.
 
 ## Adapter details
 
@@ -97,8 +141,8 @@ exactly why we publish it as a **specialist**:
 - **Training (v4)**: 14,768 cases (×3 questions), 1 epoch, BF16, single
   24 GB GPU (~6 h); OOM-resilient loader.
 - **Training (v5)**: clean deduplicated 3-question cases, 2 epochs, BF16,
-  RTX PRO 6000 (~4 h). Same urgency profile, worse category — confirming
-  the architecture limitation.
+  RTX PRO 6000 (~4 h). Same historical urgency profile, worse category;
+  pipeline defects prevent an architectural conclusion from this run.
 - **Format**: this repo contains the adapter (`adapter_model.safetensors`,
   `adapter_config.json` with `base_model_name_or_path: Cloudflare/clef-flash`).
 
@@ -130,8 +174,9 @@ probs = logits.float().softmax(-1)
 
 ## Limitations
 
-- Not a general classifier — category/decision accuracy is far below the
-  322M specialists (see table); use Decisum for those.
+- The legacy v4/v5 evaluations reported category/decision accuracy far
+  below the 322M specialists (see table); these scores do not describe
+  the separate, unpublished v6 checkpoint.
 - Urgency numbers are in-domain (PT-BR e-commerce support).
 - Requires the base model's custom code (`trust_remote_code`) and its
   joint-schema processor.