Add scoring_contract: direction, aggregation, query-relevance status

metadata.json

@@ -6,5 +6,22 @@
   "alpha_rank": 1.0,
   "beta_retrieve": 2.0,
   "gamma_stability": 0.05,
-  "heuristic_init": false
+  "heuristic_init": false,
+  "scoring_contract": {
+    "aggregation": "arithmetic mean of all token scores per passage",
+    "score_source": "ImportanceUpdateHead.direct_score(hidden) \u2014 raw logits (no sigmoid)",
+    "direction": "descending (high-first) \u2014 validated 2026-08-01",
+    "passage_encoding": "each passage encoded independently (no query context in scoring call)",
+    "normalization": "none",
+    "query_relevance_status": "context-utility signal; calibrated query-relevance not demonstrated (MRR delta=-0.006, 95% CI [-0.051,+0.041])",
+    "validation_scripts": [
+      "scripts/validate_score_direction.py",
+      "scripts/test_query_specificity.py"
+    ],
+    "validation_results": [
+      "results/score_direction_validation_summary.csv",
+      "results/query_specificity_summary.csv"
+    ],
+    "github": "https://github.com/nitroxido/token-importance-scoring"
+  }
 }