matilda-jev-v1.3: Decision Index 0.3 public suite (index 60.16, kit 62d2f51)

runs/matilda-jev-v1.3/RUN_NOTES.md

@@ -0,0 +1,28 @@
+# Matilda Jev v1.3: Decision Index 0.3 public suite
+
+- **Model:** [Maincode/matilda-jev-v1](https://huggingface.co/Maincode/matilda-jev-v1) @ `c87f575`. The weights used
+  here are byte-identical to that commit: all 11 shards, `readout.safetensors` and the temperature were checked by
+  sha256.
+- **Kit:** apolinario/decision-index `62d2f51` (edition 0.3).
+- **Engine:** `kev.bench:KevEngine`, our in-process engine. Inference settings: bf16, one forward pass per question
+  batch, the stored temperature, no thinking.
+  In a fresh environment, the model repo's own engine (`maincode_jev_serve.engine:MaincodeJevEngine`, from `runtime/`) gave
+  the same choice on all 6,571 questions of 1,500 suite requests, with a maximum probability difference of 0.0000.
+- **Result:** `scores.json` is complete. Public index 60.16 (raw 70.04, breadth 59.21). 140,178 of 140,178 scored
+  requests answered, all ok.
+- **Latency:** median 57 ms, mean 120 ms, p95 383 ms per request (1 × AMD MI355X, in-process).
+
+## How the run was made
+
+The run was made in two parts, as the 0.3 README allows ("a complete 0.2.1 run directory resumed under 0.3 only needs
+the 2,638 rebuilt GSM8K requests"):
+
+1. **The complete 0.2.1 run** (`environment.json`) covered 150,317 requests over 32 shards on MI355X. Its answers to
+   the requests that 0.3 keeps are reused unchanged.
+2. **A 0.3 resume** (`environment-0.3-resume.json`, 1 × MI355X) answered the 2,638 rebuilt GSM8K requests
+   (`--rows gsm8k-rows.jsonl.gz`).
+
+**Cleanup:** an earlier attempt at the resume ran `decision_index run` over the whole suite. Before it was cancelled,
+it answered 48 ToolRet requests that are on the suite's exclusion list and are never scored. Those 48 lines were
+removed from `results.jsonl`; no other line was changed. `status.json` is the resume job's own record, so its
+`completed` count (153,003) still includes them. `results.jsonl.gz` has 152,955 lines.

runs/matilda-jev-v1.3/benchmark-summary.json

@@ -1,12 +1,12 @@
 {
-  "engine": "matilda-jev-v1.3",
+  "engine": "matilda-jev-v1.3-0.3",
   "counts": {
-    "ok": 150317
+    "ok": 152955
   },
   "successful_request_latency_ms": {
-    "median": 57.26611567661166,
-    "p95": 398.21481378749013,
-    "mean": 120.72431615369824
+    "median": 56.82946718297899,
+    "p95": 383.00070504192263,
+    "mean": 119.53657728062879
   },
   "benchmarks": [
     {
@@ -883,38 +883,6 @@
         "custom_metrics": {}
       }
     },
-    {
-      "catalog_id": 30,
-      "dataset": "GSM8K",
-      "requests": 2638,
-      "answered": 2638,
-      "unsupported": 0,
-      "errors": 0,
-      "abstained": 0,
-      "pending": 0,
-      "scored_requests": 2638,
-      "metric": "accuracy",
-      "score": 0.7115238817285823,
-      "reference_same_cases": null,
-      "median_ms": 52.670650416985154,
-      "tracks": {
-        "GSM8K-10choice": {
-          "metric": "accuracy",
-          "score": 0.6338134950720242,
-          "scored_requests": 1319
-        },
-        "GSM8K-4choice": {
-          "metric": "accuracy",
-          "score": 0.7892342683851402,
-          "scored_requests": 1319
-        }
-      },
-      "detail": {
-        "field_accuracy": 0.7115238817285823,
-        "case_exact_accuracy": 0.5815011372251706,
-        "custom_metrics": {}
-      }
-    },
     {
       "catalog_id": 31,
       "dataset": "ChessBench",
@@ -1250,40 +1218,6 @@
         "custom_metrics": {}
       }
     },
-    {
-      "catalog_id": 48,
-      "dataset": "ForecastBench",
-      "requests": 10139,
-      "answered": 10139,
-      "unsupported": 0,
-      "errors": 0,
-      "abstained": 0,
-      "pending": 0,
-      "scored_requests": 10139,
-      "metric": "Brier (lower is better)",
-      "score": 0.17000564947287658,
-      "reference_same_cases": null,
-      "median_ms": 79.54525016248226,
-      "detail": {
-        "field_accuracy": null,
-        "case_exact_accuracy": null,
-        "custom_metrics": {
-          "brier": 0.17000564947287658,
-          "log_loss": 0.49710315964797286,
-          "accuracy": 0.7213729164611895
-        },
-        "subgroups": {
-          "joint": {
-            "n": 3242,
-            "mean": 0.16584322095922385
-          },
-          "scalar": {
-            "n": 6897,
-            "mean": 0.17196223831458488
-          }
-        }
-      }
-    },
     {
       "catalog_id": 50,
       "dataset": "Habermas Machine",
@@ -1306,6 +1240,38 @@
         }
       }
     },
+    {
+      "catalog_id": 30,
+      "dataset": "GSM8K",
+      "requests": 2638,
+      "answered": 2638,
+      "unsupported": 0,
+      "errors": 0,
+      "abstained": 0,
+      "pending": 0,
+      "scored_requests": 2638,
+      "metric": "accuracy",
+      "score": 0.650871872630781,
+      "reference_same_cases": null,
+      "median_ms": 51.691783126443624,
+      "tracks": {
+        "GSM8K-10choice": {
+          "metric": "accuracy",
+          "score": 0.5716451857467779,
+          "scored_requests": 1319
+        },
+        "GSM8K-4choice": {
+          "metric": "accuracy",
+          "score": 0.730098559514784,
+          "scored_requests": 1319
+        }
+      },
+      "detail": {
+        "field_accuracy": 0.650871872630781,
+        "case_exact_accuracy": 0.5193328278999242,
+        "custom_metrics": {}
+      }
+    },
     {
       "catalog_id": 56,
       "dataset": "PhishNChips phishing decisions",
@@ -1441,5 +1407,5 @@
     }
   ],
   "note": "Native benchmark metrics on complete supported case groups. Unsupported cases excluded from accuracy but retained in coverage.",
-  "edition": "0.2.1"
+  "edition": "0.3"
 }

runs/matilda-jev-v1.3/environment-0.3-resume.json

@@ -0,0 +1,24 @@
+{
+  "engine": "kev.bench:KevEngine",
+  "engine_options": {
+    "checkpoint": "/shared/model-cache/matilda-jev-v1.3-trained"
+  },
+  "model_source": {
+    "base_model": "/shared/training/arsh/models/matilda-jev-v1-base",
+    "revision": "matilda-jev-v1-base@f941273735b00913",
+    "checkpoint": "/shared/model-cache/matilda-jev-v1.3-trained",
+    "checkpoint_config_sha256": "19a4e14690068c1d77e827332e1eceb9e65437063dadad47964a80b2440dd4a6",
+    "temperature": 1.2834105376296687,
+    "max_tokens": 131072,
+    "prompt": "kev.model.decision_messages"
+  },
+  "device": "AMD Instinct MI355X",
+  "memory_gib": 288,
+  "torch": "2.14.0+rocm7.2",
+  "hip": "7.2.53211",
+  "loaded_seconds": 17.107464374043047,
+  "frozen_corpus_sha256": null,
+  "rows_path": "/shared/training/arsh/data/decision-index/suite-0.3/gsm8k-rows.jsonl.gz",
+  "runner_sha256": "f4c9f36c4c557f537f1d8184db7bf00c5a84d9f4f0aa67f8c06198cead25818d",
+  "latency": "In-process request wall time: prompt rendering, tokenization and one forward pass per question batch."
+}

runs/matilda-jev-v1.3/index.json

@@ -1,21 +1,21 @@
 {
-  "edition": "0.2.1",
-  "panel_id": "decision-index-0.2.1",
-  "index": 59.59,
-  "raw_index": 69.18,
+  "edition": "0.3",
+  "panel_id": "decision-index-0.3",
+  "index": 60.16,
+  "raw_index": 70.04,
   "scores": {
-    "balanced_skill": 59.59,
-    "balanced_raw": 69.18,
-    "breadth_skill": 58.39
+    "balanced_skill": 60.16,
+    "balanced_raw": 70.04,
+    "breadth_skill": 59.21
   },
   "areas": [
     {
       "id": "knowledge",
       "label": "Knowledge & Reasoning",
-      "raw": 0.5509,
-      "skill": 0.4409,
+      "raw": 0.5796,
+      "skill": 0.4684,
       "coverage": 1.0,
-      "n": 10,
+      "n": 11,
       "benchmarks": [
         25,
         30,
@@ -26,20 +26,20 @@
         44,
         45,
         57,
-        58
+        58,
+        28
       ]
     },
     {
       "id": "language",
       "label": "Language Understanding",
-      "raw": 0.7627,
-      "skill": 0.6627,
+      "raw": 0.7466,
+      "skill": 0.648,
       "coverage": 1.0,
-      "n": 10,
+      "n": 9,
       "benchmarks": [
         11,
         12,
-        28,
         29,
         38,
         39,
@@ -83,16 +83,15 @@
     {
       "id": "arts",
       "label": "Arts & Human Taste",
-      "raw": 0.5888,
-      "skill": 0.4385,
+      "raw": 0.6426,
+      "skill": 0.4622,
       "coverage": 1.0,
-      "n": 7,
+      "n": 6,
       "benchmarks": [
         20,
         21,
         22,
         23,
-        48,
         50,
         64
       ]
@@ -228,8 +227,8 @@
       "in_index": true
     },
     "30": {
-      "raw": 0.7115,
-      "skill": 0.6561,
+      "raw": 0.6509,
+      "skill": 0.5821,
       "coverage": 1.0,
       "random": 0.25,
       "rule": "track",
@@ -237,12 +236,12 @@
       "tracks": [
         {
           "track": "GSM8K-4choice",
-          "score": 0.7892,
+          "score": 0.7301,
           "headline": false
         },
         {
           "track": "GSM8K-10choice",
-          "score": 0.6338,
+          "score": 0.5716,
           "headline": false
         }
       ]
@@ -379,14 +378,6 @@
       "rule": "chance",
       "in_index": true
     },
-    "48": {
-      "raw": 0.32,
-      "skill": 0.32,
-      "coverage": 1.0,
-      "random": 0.25,
-      "rule": "vs baseline",
-      "in_index": true
-    },
     "50": {
       "raw": 0.4541,
       "skill": 0.2077,
@@ -502,5 +493,5 @@
     }
   },
   "coverage": 1.0,
-  "note": "Decision Index 0.2.1 averages 38 benchmarks in five areas. Arts & Human Taste weighs 10%; the other four share 90% in proportion to the square root of their benchmark count (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. ForecastBench enters against its baseline: clip((0.25 - Brier) / 0.25) x coverage, so always predicting 0.5 scores zero. MMLU, ARC-Easy, ARC-Challenge, RouterBench, SGD stay on the board as non-index benchmarks. The six interactive environments are still unrun and stay out. Every entrant on the board has results on all 38 index benchmarks. Point estimates only, no uncertainty intervals yet."
+  "note": "The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers."
 }

runs/matilda-jev-v1.3/results.jsonl.gz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5bf9afe4c0b2bca7fd377fe49fdb2e839d3422767a6c8943fa01b069144d12e3
-size 37891819
+oid sha256:2092be5d02c5e80efcf8ed9b61a5c57941354063311447455ff8a4461523c875
+size 38415177

runs/matilda-jev-v1.3/scores.json

@@ -1,42 +1,43 @@
 {
-  "engine": "matilda-jev-v1.3",
-  "edition": "0.2.1",
-  "generated_utc": "2026-10-01T06:10:43+00:00",
+  "engine": "matilda-jev-v1.3-0.3",
+  "edition": "0.3",
+  "generated_utc": "2026-10-07T02:29:40+00:00",
   "suite": {
-    "edition": "release-v2.1",
-    "requests": 120340,
-    "scoreable": 119898,
+    "edition": "release-v3",
+    "requests": 110201,
+    "scoreable": 109759,
     "excluded": 442,
     "added_requests": 30419,
-    "benchmarks": 44,
+    "benchmarks": 43,
     "rows_sha256": "b2b56d6fb636837ca469e689087bdbf373dda8de7638aa2da6793e6eda0792d5",
-    "added_sha256": "7429f3c9cdddb772c1cfc42bb2a45e8516b0032152b746e6929f1c8b52f4ce89"
+    "added_sha256": "7429f3c9cdddb772c1cfc42bb2a45e8516b0032152b746e6929f1c8b52f4ce89",
+    "gsm8k_sha256": "759334894858a37d6bd7bd07298eac8e55da236e31dad327ccd06a9979f4738a"
   },
-  "completed": 150317,
+  "completed": 140178,
   "complete": true,
   "counts": {
-    "ok": 150317
+    "ok": 140178
   },
   "latency_ms": {
-    "median": 57.3,
-    "p95": 398.2,
-    "mean": 120.7
+    "median": 56.8,
+    "p95": 383.0,
+    "mean": 119.5
   },
-  "decision_index": 59.59,
-  "raw_index": 69.18,
+  "decision_index": 60.16,
+  "raw_index": 70.04,
   "scores": {
-    "balanced_skill": 59.59,
-    "balanced_raw": 69.18,
-    "breadth_skill": 58.39
+    "balanced_skill": 60.16,
+    "balanced_raw": 70.04,
+    "breadth_skill": 59.21
   },
   "areas": [
     {
       "id": "knowledge",
       "label": "Knowledge & Reasoning",
-      "raw": 0.5509,
-      "skill": 0.4409,
+      "raw": 0.5796,
+      "skill": 0.4684,
       "coverage": 1.0,
-      "n": 10,
+      "n": 11,
       "benchmarks": [
         25,
         30,
@@ -47,20 +48,20 @@
         44,
         45,
         57,
-        58
+        58,
+        28
       ]
     },
     {
       "id": "language",
       "label": "Language Understanding",
-      "raw": 0.7627,
-      "skill": 0.6627,
+      "raw": 0.7466,
+      "skill": 0.648,
       "coverage": 1.0,
-      "n": 10,
+      "n": 9,
       "benchmarks": [
         11,
         12,
-        28,
         29,
         38,
         39,
@@ -104,16 +105,15 @@
     {
       "id": "arts",
       "label": "Arts & Human Taste",
-      "raw": 0.5888,
-      "skill": 0.4385,
+      "raw": 0.6426,
+      "skill": 0.4622,
       "coverage": 1.0,
-      "n": 7,
+      "n": 6,
       "benchmarks": [
         20,
         21,
         22,
         23,
-        48,
         50,
         64
       ]
@@ -249,8 +249,8 @@
       "in_index": true
     },
     "30": {
-      "raw": 0.7115,
-      "skill": 0.6561,
+      "raw": 0.6509,
+      "skill": 0.5821,
       "coverage": 1.0,
       "random": 0.25,
       "rule": "track",
@@ -258,12 +258,12 @@
       "tracks": [
         {
           "track": "GSM8K-4choice",
-          "score": 0.7892,
+          "score": 0.7301,
           "headline": false
         },
         {
           "track": "GSM8K-10choice",
-          "score": 0.6338,
+          "score": 0.5716,
           "headline": false
         }
       ]
@@ -400,14 +400,6 @@
       "rule": "chance",
       "in_index": true
     },
-    "48": {
-      "raw": 0.32,
-      "skill": 0.32,
-      "coverage": 1.0,
-      "random": 0.25,
-      "rule": "vs baseline",
-      "in_index": true
-    },
     "50": {
       "raw": 0.4541,
       "skill": 0.2077,
@@ -935,38 +927,6 @@
       "chance": 0.25,
       "in_index": true
     },
-    "30": {
-      "catalog_id": 30,
-      "dataset": "GSM8K",
-      "requests": 2638,
-      "answered": 2638,
-      "unsupported": 0,
-      "errors": 0,
-      "abstained": 0,
-      "pending": 0,
-      "scored_requests": 2638,
-      "metric": "accuracy",
-      "score": 0.7115,
-      "reference_same_cases": null,
-      "median_ms": 52.7,
-      "tracks": [
-        {
-          "track": "GSM8K-4choice",
-          "score": 0.7892,
-          "headline": false
-        },
-        {
-          "track": "GSM8K-10choice",
-          "score": 0.6338,
-          "headline": false
-        }
-      ],
-      "index_raw": 0.7115,
-      "index_skill": 0.6561,
-      "coverage": 1.0,
-      "chance": 0.25,
-      "in_index": true
-    },
     "31": {
       "catalog_id": 31,
       "dataset": "ChessBench",
@@ -1269,26 +1229,6 @@
       "chance": 0.1641,
       "in_index": true
     },
-    "48": {
-      "catalog_id": 48,
-      "dataset": "ForecastBench",
-      "requests": 10139,
-      "answered": 10139,
-      "unsupported": 0,
-      "errors": 0,
-      "abstained": 0,
-      "pending": 0,
-      "scored_requests": 10139,
-      "metric": "Brier (lower is better)",
-      "score": 0.17,
-      "reference_same_cases": null,
-      "median_ms": 79.5,
-      "index_raw": 0.32,
-      "index_skill": 0.32,
-      "coverage": 1.0,
-      "chance": 0.25,
-      "in_index": true
-    },
     "50": {
       "catalog_id": 50,
       "dataset": "Habermas Machine",
@@ -1309,6 +1249,38 @@
       "chance": 0.311,
       "in_index": true
     },
+    "30": {
+      "catalog_id": 30,
+      "dataset": "GSM8K",
+      "requests": 2638,
+      "answered": 2638,
+      "unsupported": 0,
+      "errors": 0,
+      "abstained": 0,
+      "pending": 0,
+      "scored_requests": 2638,
+      "metric": "accuracy",
+      "score": 0.6509,
+      "reference_same_cases": null,
+      "median_ms": 51.7,
+      "tracks": [
+        {
+          "track": "GSM8K-4choice",
+          "score": 0.7301,
+          "headline": false
+        },
+        {
+          "track": "GSM8K-10choice",
+          "score": 0.5716,
+          "headline": false
+        }
+      ],
+      "index_raw": 0.6509,
+      "index_skill": 0.5821,
+      "coverage": 1.0,
+      "chance": 0.25,
+      "in_index": true
+    },
     "56": {
       "catalog_id": 56,
       "dataset": "PhishNChips phishing decisions",
@@ -1443,6 +1415,6 @@
       "in_index": true
     }
   },
-  "panel_id": "decision-index-0.2.1",
-  "note": "Decision Index 0.2.1 averages 38 benchmarks in five areas. Arts & Human Taste weighs 10%; the other four share 90% in proportion to the square root of their benchmark count (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. ForecastBench enters against its baseline: clip((0.25 - Brier) / 0.25) x coverage, so always predicting 0.5 scores zero. MMLU, ARC-Easy, ARC-Challenge, RouterBench, SGD stay on the board as non-index benchmarks. The six interactive environments are still unrun and stay out. Every entrant on the board has results on all 38 index benchmarks. Point estimates only, no uncertainty intervals yet."
+  "panel_id": "decision-index-0.3",
+  "note": "The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers."
 }

runs/matilda-jev-v1.3/status.json

@@ -1 +1 @@
-{"time": "2026-10-01T06:10:20.753450+00:00", "event": "complete", "completed": 150317, "counts": {"ok": 150317}, "shards": 32}
+{"time": "2026-10-07T02:28:35.862316+00:00", "event": "complete", "engine": "kev.bench:KevEngine", "completed": 153003, "counts": {"ok": 2638}, "elapsed_seconds": 137.7}