fix(evaluation): harden judge calibration workflow

datasets/regrag-br/calibration/archive/20260726T185553Z/.judge-sealed.json

@@ -0,0 +1,81 @@
+{
+  "run_id": "20260726T185553Z",
+  "judge_scores": {
+    "q199-graphrag-faithfulness": 1.0,
+    "q199-graphrag-answer_relevancy": 0.8347400935056873,
+    "q186-raptor-faithfulness": 1.0,
+    "q186-raptor-answer_relevancy": 0.85198688099497,
+    "q149-contextual-faithfulness": 1.0,
+    "q149-contextual-answer_relevancy": 0.8077504963301628,
+    "q149-raptor-faithfulness": 1.0,
+    "q149-raptor-answer_relevancy": 0.9460890444439526,
+    "q057-graphrag-faithfulness": 0.5,
+    "q057-graphrag-answer_relevancy": 0.0,
+    "q228-parent_child-faithfulness": 1.0,
+    "q228-parent_child-answer_relevancy": 0.6856243446899631,
+    "q188-graphrag-faithfulness": 1.0,
+    "q188-graphrag-answer_relevancy": 0.8336470959717595,
+    "q057-sac_contextual-faithfulness": 1.0,
+    "q057-sac_contextual-answer_relevancy": 0.8387442659907304,
+    "q129-hybrid_rrf-faithfulness": 1.0,
+    "q129-hybrid_rrf-answer_relevancy": 0.9124998734589621,
+    "q017-hybrid_rrf-faithfulness": 1.0,
+    "q017-hybrid_rrf-answer_relevancy": 0.7206135051759966,
+    "q017-hybrid_rrf-abstention": 1.0,
+    "q228-dense-faithfulness": 0.8571428571428571,
+    "q228-dense-answer_relevancy": 0.6870679326468675,
+    "q188-sac-faithfulness": 1.0,
+    "q188-sac-answer_relevancy": 0.8327270764043573,
+    "q162-sac-faithfulness": 1.0,
+    "q162-sac-answer_relevancy": 0.7749749305427135,
+    "q091-raptor-faithfulness": 1.0,
+    "q091-raptor-answer_relevancy": 0.9702814665461529,
+    "q091-raptor-abstention": 1.0,
+    "q147-reranked-faithfulness": 1.0,
+    "q147-reranked-answer_relevancy": 0.9533329415627246,
+    "q029-sac-faithfulness": 1.0,
+    "q029-sac-answer_relevancy": 0.8552758720163123,
+    "q086-contextual-faithfulness": 1.0,
+    "q086-contextual-answer_relevancy": 0.9940291063255439,
+    "q152-dense-faithfulness": 1.0,
+    "q152-dense-answer_relevancy": 0.8740158431078623,
+    "q053-contextual-faithfulness": 1.0,
+    "q053-contextual-answer_relevancy": 0.8949108045847701,
+    "q160-hybrid_rrf-faithfulness": 1.0,
+    "q160-hybrid_rrf-answer_relevancy": 0.7715785879645054,
+    "q011-reranked-faithfulness": 1.0,
+    "q011-reranked-answer_relevancy": 0.8502438689143164,
+    "q057-sac-faithfulness": 1.0,
+    "q057-sac-answer_relevancy": 0.9170167453202009,
+    "q129-contextual-faithfulness": 1.0,
+    "q129-contextual-answer_relevancy": 0.9112467674362174,
+    "q190-parent_child-faithfulness": 0.3333333333333333,
+    "q190-parent_child-answer_relevancy": 0.782537101530835,
+    "q190-parent_child-abstention": 1.0,
+    "q228-sac_contextual-faithfulness": 0.6666666666666666,
+    "q228-sac_contextual-answer_relevancy": 0.7052906182663375,
+    "q011-dense-faithfulness": 1.0,
+    "q011-dense-answer_relevancy": 0.7049204363155358,
+    "q129-raptor-faithfulness": 0.9230769230769231,
+    "q129-raptor-answer_relevancy": 0.9081927270706781,
+    "q017-raptor-faithfulness": 1.0,
+    "q017-raptor-answer_relevancy": 0.7381427478477991,
+    "q017-raptor-abstention": 1.0,
+    "q024-contextual-faithfulness": 1.0,
+    "q024-contextual-answer_relevancy": 0.9541869181118697,
+    "q193-raptor-faithfulness": 0.5,
+    "q193-raptor-answer_relevancy": 0.8704297565744411,
+    "q123-hybrid_rrf-faithfulness": 1.0,
+    "q123-hybrid_rrf-answer_relevancy": 0.8033356630688374,
+    "q205-sac_contextual-faithfulness": 1.0,
+    "q205-sac_contextual-answer_relevancy": 0.7516688618687585,
+    "q093-hybrid_rrf-faithfulness": 1.0,
+    "q093-hybrid_rrf-answer_relevancy": 0.7226008284858803,
+    "q187-graphrag-faithfulness": 1.0,
+    "q187-graphrag-answer_relevancy": 0.5653568713203554,
+    "q128-hybrid_rrf-faithfulness": 0.8571428571428571,
+    "q128-hybrid_rrf-answer_relevancy": 0.8334048457508137,
+    "q107-dense-faithfulness": 1.0,
+    "q107-dense-answer_relevancy": 0.931451117336013
+  }
+}

datasets/regrag-br/calibration/archive/20260726T185553Z/README.md

@@ -0,0 +1,19 @@
+# Legacy calibration workspace
+
+This directory preserves the attempted calibration workspace derived from run
+`20260726T185553Z` for audit history only.
+
+The run predates evidence schema v1 and does not contain the exact ordered
+`judge_contexts` used by the Faithfulness judge. The worksheet reconstructs
+cited source excerpts instead, so neither it nor the sealed scores are eligible
+for judge calibration. Do not fill these labels or use this workspace to report
+agreement.
+
+Original artifact SHA-256 values at archival time:
+
+- `worksheet.md`: `f7ff210d12f7265049cc4cd741d9b0a9f22e520eaefe04fcfb9fc0fbc234ffaa`
+- `labels.json`: `b11eae0437820ac3578c8389a2ef63613b6fa17b45f1eddccb24abd00e5133ca`
+- `.judge-sealed.json`: `af5ed04f740bf3799b0f55ebde36e68a54e3afa2dcb04a1ff134b1d0f621aaf9`
+
+A future eligible run must generate a fresh workspace directly under
+`datasets/regrag-br/calibration/` with `build_calibration_worksheet.py`.