Document per-case score and verdict files

README.md

@@ -112,12 +112,28 @@ Two observations that only the full rubric set exposes:
 
 | File | Content |
 |---|---|
+| [`eval/per_case_scores.csv`](./blob/main/eval/per_case_scores.csv) | **per-case scores, one row per (model, case)**: 800 rows × overall, three family scores, all six Frozen and all Static dimension scores, frozen passed/total, render_ok |
+| [`eval/per_case_verdicts.jsonl`](./blob/main/eval/per_case_verdicts.jsonl) | **every individual rubric verdict with the judge's evidence sentence** — 800 rows, 51,652 verdicts (Frozen + Prompt Fit + Static), so any score can be traced to a specific criterion |
 | [`eval/benchmark_summary_v592.csv`](./blob/main/eval/benchmark_summary_v592.csv) | per-model aggregates: overall, three family scores, Track/Surface splits, Static and Frozen dimensions |
-| [`eval/benchmark_per_case_v592.csv`](./blob/main/eval/benchmark_per_case_v592.csv) | long-form per-case scores (overall + three families) for all 4 models × 200 cases |
+| [`eval/benchmark_per_case_v592.csv`](./blob/main/eval/benchmark_per_case_v592.csv) | per-case scores in wide form: one row per case, four models side by side |
 | [`eval/significance_tests_v592.json`](./blob/main/eval/significance_tests_v592.json) | paired bootstrap (10k resamples) + Wilcoxon signed-rank for all 6 model pairs |
 | [`eval/rubric_stats.json`](./blob/main/eval/rubric_stats.json) | composition of all three rubric families and the aggregation rule |
 | [`eval/reports.html`](./blob/main/eval/reports.html) | self-contained interactive HTML report: model comparison, dimension heatmap, score distributions, per-case tables |
 
+Reading the per-case files:
+
+```python
+import pandas as pd
+
+scores = pd.read_csv("eval/per_case_scores.csv")
+scores.pivot_table(index="surface", columns="model", values="overall_score")
+
+verdicts = pd.read_json("eval/per_case_verdicts.jsonl", lines=True)
+row = verdicts.query("model == '27B_AdamW_step400'").iloc[0]
+[c for c in row.frozen_checks if c["score"] == 0]      # criteria this page failed
+[c for c in row.static_checks if c["score"] == 0]      # integrity/craft failures
+```
+
 ### Checkpoint selection
 
 The `data41287` checkpoints were selected by **running the benchmark**, not by taking the