Upload README.md with huggingface_hub

README.md

@@ -14,7 +14,7 @@ size_categories:
   - n<1K
 ---
 
-# PL-Temporal — Track A
+# PL-Temporal — Track A (Diagnostic Challenge Set v0.2)
 
 A benchmark of temporal legal reasoning over Polish statutes: which version of the law
 was in force on a given date, when did an act enter into force, which Dziennik Ustaw
@@ -42,21 +42,30 @@ https://github.com/bartoszkobylinski/pl-temporal
 | A3 | 24 | whether an act was in force on a given date (TAK/NIE) |
 | A4 | 24 | whether an act was amended in a given period (TAK/NIE) |
 
-## v0.2 results (2026-08-14, N=5 draws where provider quotas allowed, 79 scored)
+## v0.2 results (2026-08-14, 79 scored items; mean over valid draws)
 
-| model | N valid | strict (answered) | abstained/draw |
-|---|---:|---:|---:|
-| claude-opus-5 | 5 | **74.2%** | 11.6 |
-| gpt-5.6-sol | 5 | 69.9% | **0** |
-| gemini-3.1-pro | 3 | 64.6% | 14.0 |
-| gpt-5.6-terra | 5 | 57.2% | **0** |
-| gemini-3.5-flash | 2 | 56.8% | 5.0 |
-| pllum-12b | 5 | 44.3% | **0** |
-| bielik-11b-v3 | 5 | 43.0% | **0** |
+Three metrics — no single leader without a utility function. Coverage: how often
+the model answers. Selective accuracy: how often it is right when it answers.
+End-to-end: usefulness when every question needs an answer.
 
-Gold answers for every analyzer-flagged item were independently verified by a human
+| model | N valid | coverage | selective acc | end-to-end | abstained/draw |
+|---|---:|---:|---:|---:|---:|
+| gpt-5.6-sol | 5 | **100%** | 69.9% | **69.9%** | **0** |
+| claude-opus-5 | 5 | 85.1% | **74.2%** | 63.0% | 11.6 |
+| gpt-5.6-terra | 5 | 100% | 57.2% | 57.2% | **0** |
+| gemini-3.1-pro | 3 | 82.3% | 64.6% | 53.2% | 14.0 |
+| gemini-3.5-flash | 2 | 93.7% | 56.8% | 53.2% | 5.0 |
+| pllum-12b | 5 | 100% | 44.3% | 44.3% | **0** |
+| bielik-11b-v3 | 5 | 100% | 43.0% | 43.0% | **0** |
+
+Majority baselines for the binary families: A3 = 65.2%, A4 = 50.0%. v0.2 is frozen
+as a diagnostic challenge set (no-signal pruning used v0.1 model results); a frozen
+v1.0 with a pre-registered selection procedure is planned. Full findings,
+limitations and incident log: RESULTS-v0.2.md in the canonical repo.
+
+Gold answers for every analyzer-flagged item were verified by the dataset owner
 against ISAP commencement clauses (2026-08-14): zero gold edits, one ambiguous item
-quarantined. Notable finding: zero-abstention is a property of vendor training
+quarantined; an independent second annotation pass is planned for v1.0. Notable finding: zero-abstention is a property of vendor training
 decisions, not of model size — both GPT-5.6 models abstained 0/400, matching the
 Polish models, while legacy gpt-5 was the heaviest abstainer in the roster.