Update README.md

README.md

@@ -125,6 +125,83 @@ a property of the training data. The −0.19 residual is real, and individual ce
 are worse (`double_session_accumulation` at coach register is −2.70), but the
 pattern is inherited.
 
+## Revision history
+
+This is the eighth corpus revision and the eighth corresponding fine-tune. The
+training configuration was **identical across all eight** — LoRA rank 32,
+lr 2e-4, 1 epoch, 20,000 examples, max length 512 — so the only variable in the
+sequence is the corpus.
+
+> **These rows are not directly comparable.** Each revision regenerated the
+> corpus and therefore its own test split; all eight formatted test splits are
+> distinct files. The varying reference ceiling is the tell. Read each row
+> against its own ceiling, not against the row above it. The exception is
+> `sft_only_greedy` / `sft_v2` / `sft_v2_sampled`, which were scored in one
+> evaluation run against one split.
+
+| Model | Ceiling | Reward | Risk | W-in 1 | Severe | Unread. | Overr. | Clin. |
+| --- | --- | --- | --- | --- | --- | --- | --- | --- |
+| *zero-shot floor* | 0.701 | 0.422 | 0.000 | 0.000 | 0 † | 500 | 0.006 | — |
+| sft_only_greedy (v1) | 0.691 | 0.755 | 0.792 | 0.974 | 10 | 3 | 0.758 | 3.00 |
+| sft_only (v1, sampled) | 0.691 | — | — | — | — | — | — | 3.24 |
+| sft_v2 | 0.691 | 0.675 | 0.794 | 0.992 | 4 | 0 | 0.458 | 4.38 |
+| sft_v2_sampled | 0.691 | 0.660 | 0.702 | 0.948 | 13 | 13 | 0.352 | — |
+| sft_v3 | 0.712 | 0.735 | 0.842 | 0.994 | 3 | 0 | 0.518 | 4.09 |
+| sft_v4 | 0.716 | 0.649 | 0.856 | 0.956 | 22 | 0 | 0.482 | 4.23 |
+| sft_v5 | 0.701 | 0.720 | 0.934 | 0.982 | 8 | 1 | 0.552 | 4.23 |
+| sft_v6 | 0.697 | 0.701 | 0.942 | 0.972 | 3 | 11 | 0.512 | 4.11 |
+| sft_v7 | 0.704 | 0.720 | 0.948 | 0.986 | 3 | 4 | 0.554 | 3.89 |
+| **sft_v8 (released)** | 0.701 | **0.668** | **0.960** | **0.994** | **3** | **0** | 0.504 | 4.10 |
+
+† The base model's zero severe errors reflect 500 unparseable completions, not
+accuracy.
+
+### What actually improved
+
+Three trends are large and monotone enough to survive the caveat.
+
+**Risk accuracy rose 0.792 → 0.960.** Most of this tracks repairs to
+declared-target reachability: nine scenario targets produced metric ranges their
+own declarations did not cover at v1, and eight were repaired by v8. Making a
+scenario actually produce its declared range makes it easier to recognize from
+the narrative — which is what this metric measures.
+
+**Unparseable outputs fell to zero**, from 13 at worst. **Severe errors fell to
+3**, from a v4 excursion of 22.
+
+### What did not
+
+**Judged clinical accuracy did not move.** It peaks at v2 (4.38) and sits at 4.10
+in v8, with no trend across the whole repair sequence. Reachability repair makes
+the sampled signals match the declared range; it does nothing to make the
+constant label respond to those signals. Repairing eight of nine unreachable
+targets bought 17 points of risk accuracy and no measurable clinical quality.
+
+**Reward did not improve** — v1 records the highest in the series. Reward and
+accuracy move in opposite directions because they measure different things; see
+the reward decomposition above.
+
+**Overreaching accuracy did not decline, despite appearances.** The drop from
+0.758 to 0.458 between the first two models is a measurement change, not a
+capability loss. From v2 onward the corpus phrases conflicting-signal
+classifications as rationale prose that names no class, so the extraction ceiling
+fell from 1.00 to ~0.55. Normalized against each corpus's own ceiling:
+
+| | v2 | v3 | v4 | v5 | v6 | v7 | v8 |
+| --- | --- | --- | --- | --- | --- | --- | --- |
+| extraction ceiling | 0.550 | 0.550 | 0.551 | 0.552 | 0.553 | 0.552 | 0.551 |
+| **% of ceiling reached** | 83% | 94% | 87% | **100%** | 93% | **100%** | 91% |
+
+The underlying trend is flat to improving — opposite in sign to the reported one.
+
+### If you want an earlier revision
+
+Only v8 is released. Earlier checkpoints are not published: several were not
+retained, they were trained on corpora that are not released, and each would need
+its own contemporaneous test split to evaluate meaningfully. The v1 corpus is
+retained under the `v1` tag on the dataset repository because this card and the
+paper both reference it.
+
 ## Limitations
 
 ### Inherited: self-contradictory outputs