Unify audited E1 and E5 result sources

INTACT-unified/PAPER_E5_GOAL_MANIFEST.json

@@ -22,17 +22,19 @@
   },
   "evaluation": {
     "interface": "Direct",
+    "result_revision": "2026-09-14",
+    "evaluation_protocol": "official_expert_continuation_v1",
     "candidate_sequences": 0,
     "terminal_cost_calls": 0,
     "evaluation_seeds": [0, 1, 42],
     "episodes_per_evaluation_seed": 100,
-    "evaluator_fingerprint_sha256": "1e475a069338bdad6c9e6a38b8ea4e2d5557ecc4cff88844e5efee79121a9fc7",
+    "history_boundary": "strictly pre-start actions only; append each policy-executed action exactly once",
     "aggregate_official_sr_percent": {
-      "pusht": {"mean": 80.2222222222, "sample_std": 1.2619796324},
-      "cube": {"mean": 99.5555555556, "sample_std": 0.7698003589},
-      "reacher": {"mean": 95.6666666667, "sample_std": 1.7638342074},
-      "tworoom": {"mean": 82.1111111111, "sample_std": 4.1141130181},
-      "macro": {"mean": 89.3888888889, "sample_std": 0.7742977703}
+      "pusht": {"mean": 86.1111111111, "sample_std": 0.96},
+      "cube": {"mean": 100.0, "sample_std": 0.0},
+      "reacher": {"mean": 97.2222222222, "sample_std": 0.84},
+      "tworoom": {"mean": 81.5555555556, "sample_std": 3.67},
+      "macro": {"mean": 91.2222222222, "sample_std": 0.51}
     }
   },
   "training_seeds": [
@@ -40,7 +42,7 @@
       "seed": 0,
       "asset": "intact-goal-e5-seed0.tar.gz",
       "shared_state_sha256": "05c6b2f030c11698038134789e02ad00bc6d01e50aebb56910de05dfaf2d5178",
-      "official_direct_sr_percent": {"pusht": 79.3333333333, "cube": 100.0, "reacher": 93.6666666667, "tworoom": 86.6666666667, "macro": 89.9166666667},
+      "official_direct_sr_percent": {"pusht": 85.0, "cube": 100.0, "reacher": 96.3333333333, "tworoom": 85.3333333333, "macro": 91.6666666667},
       "shards": [
         {"task": "pusht", "path": "checkpoints/recovery_delta_full_pusht_s0/weights_epoch_5.pt", "bytes": 84734848, "sha256": "80312771c15e50ba4b3eb3eb9a910fe6dc0a3bd140f7df1396eea306bf89d012"},
         {"task": "cube", "path": "checkpoints/recovery_delta_full_cube_s0/weights_epoch_5.pt", "bytes": 84858432, "sha256": "96a4d525e996c2b78fa384873496396024364850ad4af0003274b8c1426ea270"},
@@ -52,7 +54,7 @@
       "seed": 42,
       "asset": "intact-goal-e5-seed42.tar.gz",
       "shared_state_sha256": "e8adcbe1c27a564b3be8193c1ce0737a50247686e647e544f9985856d635661d",
-      "official_direct_sr_percent": {"pusht": 81.6666666667, "cube": 100.0, "reacher": 96.3333333333, "tworoom": 81.0, "macro": 89.75},
+      "official_direct_sr_percent": {"pusht": 86.6666666667, "cube": 100.0, "reacher": 97.3333333333, "tworoom": 81.3333333333, "macro": 91.3333333333},
       "shards": [
         {"task": "pusht", "path": "checkpoints/recovery_delta_full_pusht_s42/weights_epoch_5.pt", "bytes": 84734848, "sha256": "fbcf4dd23eb1ffefdde3ec133609ff00eba4774bd8a697ef3e1586b0fdca917d"},
         {"task": "cube", "path": "checkpoints/recovery_delta_full_cube_s42/weights_epoch_5.pt", "bytes": 84858432, "sha256": "d155d9d0c71cc48614d822b14456f341748a25524c7a899b959a54312755e94a"},
@@ -64,7 +66,7 @@
       "seed": 3072,
       "asset": "intact-goal-e5-seed3072.tar.gz",
       "shared_state_sha256": "7cf4b0ca65011c7f54bcabdfe6664e9c82fe650e13656fbb4fe0351c582e65b8",
-      "official_direct_sr_percent": {"pusht": 79.6666666667, "cube": 98.6666666667, "reacher": 97.0, "tworoom": 78.6666666667, "macro": 88.5},
+      "official_direct_sr_percent": {"pusht": 86.6666666667, "cube": 100.0, "reacher": 98.0, "tworoom": 78.0, "macro": 90.6666666667},
       "shards": [
         {"task": "pusht", "path": "checkpoints/recovery_delta_full_pusht_s3072/weights_epoch_5.pt", "bytes": 84734848, "sha256": "24a598acac5faa1e6e548e84c4d17bd79d9e9e48aae38eee384a92ce3269b590"},
         {"task": "cube", "path": "checkpoints/recovery_delta_full_cube_s3072/weights_epoch_5.pt", "bytes": 84858432, "sha256": "d2abfd5d907491cf770cc9d3f5bd003f53d0d16f01ac4732111b9dc459895f07"},

INTACT-unified/PAPER_E5_MATRIX_MANIFEST.json

@@ -7,6 +7,8 @@
   },
   "protocol": {
     "name": "Controlled Math-SDPA shared-encoder E5 matrix",
+    "result_revision": "2026-09-14",
+    "evaluation_protocol": "official_expert_continuation_v1",
     "epoch": 5,
     "training_seeds": [
       0,
@@ -25,7 +27,7 @@
       "reacher",
       "tworoom"
     ],
-    "evaluator_fingerprint_sha256": "1e475a069338bdad6c9e6a38b8ea4e2d5557ecc4cff88844e5efee79121a9fc7"
+    "history_boundary": "strictly pre-start actions only; append each policy-executed action exactly once"
   },
   "cells": [
     {
@@ -168,24 +170,24 @@
       "native_interface": "Direct",
       "aggregate_official_sr_percent": {
         "pusht": {
-          "mean": 35.0,
-          "sample_std": 1.45
+          "mean": 36.11,
+          "sample_std": 0.19
         },
         "cube": {
-          "mean": 61.89,
-          "sample_std": 1.26
+          "mean": 67.56,
+          "sample_std": 4.03
         },
         "reacher": {
-          "mean": 61.0,
-          "sample_std": 0.58
+          "mean": 90.56,
+          "sample_std": 1.35
         },
         "tworoom": {
-          "mean": 68.78,
-          "sample_std": 3.83
+          "mean": 78.56,
+          "sample_std": 6.68
         },
         "macro": {
-          "mean": 56.67,
-          "sample_std": 0.22
+          "mean": 68.19,
+          "sample_std": 0.67
         }
       },
       "training_seeds": [
@@ -436,24 +438,24 @@
       "native_interface": "Direct",
       "aggregate_official_sr_percent": {
         "pusht": {
-          "mean": 70.67,
-          "sample_std": 1.76
+          "mean": 81.78,
+          "sample_std": 0.96
         },
         "cube": {
           "mean": 100.0,
           "sample_std": 0.0
         },
         "reacher": {
-          "mean": 82.33,
-          "sample_std": 3.18
+          "mean": 88.67,
+          "sample_std": 0.33
         },
         "tworoom": {
-          "mean": 69.44,
-          "sample_std": 6.75
+          "mean": 69.22,
+          "sample_std": 7.04
         },
         "macro": {
-          "mean": 80.61,
-          "sample_std": 2.11
+          "mean": 84.92,
+          "sample_std": 1.86
         }
       },
       "training_seeds": [
@@ -704,24 +706,24 @@
       "native_interface": "Direct",
       "aggregate_official_sr_percent": {
         "pusht": {
-          "mean": 80.22,
-          "sample_std": 1.26
+          "mean": 86.11,
+          "sample_std": 0.96
         },
         "cube": {
-          "mean": 99.56,
-          "sample_std": 0.77
+          "mean": 100.0,
+          "sample_std": 0.0
         },
         "reacher": {
-          "mean": 95.67,
-          "sample_std": 1.76
+          "mean": 97.22,
+          "sample_std": 0.84
         },
         "tworoom": {
-          "mean": 82.11,
-          "sample_std": 4.11
+          "mean": 81.56,
+          "sample_std": 3.67
         },
         "macro": {
-          "mean": 89.39,
-          "sample_std": 0.77
+          "mean": 91.22,
+          "sample_std": 0.51
         }
       },
       "training_seeds": [

INTACT/AUDITED_RESULTS_20260914.json

@@ -0,0 +1,38 @@
+{
+  "schema_version": 1,
+  "result_revision": "2026-09-14",
+  "authoritative_for": "project_headline_success_rates",
+  "protocol": {
+    "benchmark": "Official LeWM",
+    "training_scope": "task-specific E1",
+    "training_seeds": [3072, 3073, 3074],
+    "evaluation_seeds": [0, 1, 42],
+    "episodes_per_evaluation_seed": 100,
+    "history_boundary": "strictly pre-start actions only; append each policy-executed action exactly once"
+  },
+  "results": {
+    "direct": {
+      "interface": "Direct",
+      "candidate_sequences": 0,
+      "official_sr_percent": {
+        "pusht": {"mean": 87.44, "sample_std": 1.26},
+        "cube": {"mean": 100.00, "sample_std": 0.00},
+        "reacher": {"mean": 97.33, "sample_std": 0.00},
+        "tworoom": {"mean": 97.67, "sample_std": 1.20},
+        "macro": {"mean": 95.61, "sample_std": 0.59}
+      }
+    },
+    "guarded_a": {
+      "interface": "Guarded A 128x3",
+      "candidate_sequences": 384,
+      "official_sr_percent": {
+        "pusht": {"mean": 91.56, "sample_std": 0.69},
+        "cube": {"mean": 99.67, "sample_std": 0.33},
+        "reacher": {"mean": 97.56, "sample_std": 0.51},
+        "tworoom": {"mean": 97.56, "sample_std": 1.26},
+        "macro": {"mean": 96.58, "sample_std": 0.44}
+      }
+    }
+  },
+  "manifest_note": "MANIFEST.json preserves the identity, hashes, and package-specific validation of the released reproduction archives; this file is authoritative for project headline success rates."
+}

INTACT/MANIFEST.json

@@ -7,6 +7,11 @@
     "path": "INTACT",
     "source_revision": "intact-goal-e1-single-p77-v1"
   },
+  "result_scope": {
+    "role": "archive-specific reproduction validation",
+    "project_headline_results": "AUDITED_RESULTS_20260914.json",
+    "use_for_project_headline": false
+  },
   "model": {
     "name": "Goal-displacement INTACT single-task reproduction",
     "epoch": 1,

INTACT/README.md

@@ -15,23 +15,24 @@ checkpoints: four tasks for each training seed (`0`, `42`, and `3072`).
 - No current or future expert action is visible
 - Official Direct evaluation, `H=5`, `RH=5`, zero search
 
-## Official Direct results
+## Audited project results
 
-| Task | Mean SR (%) | Train-seed sample std (pp) |
-|---|---:|---:|
-| PushT | 86.44 | 1.07 |
-| Cube | 100.00 | 0.00 |
-| Reacher | 97.78 | 0.51 |
-| TwoRoom | 98.78 | 0.69 |
-| **Macro** | **95.75** | **0.52** |
+| Inference | PushT | Cube | Reacher | TwoRoom | Macro SR |
+|---|---:|---:|---:|---:|---:|
+| **Direct, zero search** | 87.44 +/- 1.26 | 100.00 +/- 0.00 | 97.33 +/- 0.00 | 97.67 +/- 1.20 | **95.61 +/- 0.59** |
+| **Guarded A 128x3** | 91.56 +/- 0.69 | 99.67 +/- 0.33 | 97.56 +/- 0.51 | 97.56 +/- 1.26 | **96.58 +/- 0.44** |
 
-The matrix contains 36 evaluations: training seeds `0/42/3072`, evaluation
-seeds `0/1/42`, four tasks, and 100 episodes per cell.
+[`AUDITED_RESULTS_20260914.json`](AUDITED_RESULTS_20260914.json) is the
+machine-readable source of truth for project headline SR. `MANIFEST.json` is
+retained for the identity, hashes, and package-specific validation of the
+released `0/42/3072` reproduction archives; its validation aggregate is not
+the project headline.
 
 ## Files
 
 - `intact-goal-e1-single-seed<seed>.tar.gz`: four task shards for one seed
-- `MANIFEST.json`: release contract, aggregate results, and shard hashes
+- `AUDITED_RESULTS_20260914.json`: authoritative project headline SR
+- `MANIFEST.json`: archive identity, package validation, and shard hashes
 - `RESULTS.csv`: complete 36-cell continuation evaluation
 - `SHA256SUMS`: archive checksums
 

README.md

@@ -33,6 +33,12 @@ point and appends every policy-executed action exactly once. The history-free
 ablation removes the previous-action channel from both training and evaluation.
 Guarded A is an optional local verifier around the Direct plan; it evaluates
 384 sampled candidates rather than the 9,000 used by CEM 300x30.
+Machine-readable headline values are frozen in
+[`INTACT/AUDITED_RESULTS_20260914.json`](INTACT/AUDITED_RESULTS_20260914.json)
+and
+[`INTACT-unified/AUDITED_RESULTS_20260914.json`](INTACT-unified/AUDITED_RESULTS_20260914.json).
+Model manifests retain archive identities and hashes; use the audited result
+files for project-level reporting.
 
 ## Download