Update logbook: Repro - LOGDIFF Exact Boolean Guidance

logbook.json

@@ -11,7 +11,7 @@
     "icml2026-repro",
     "paper-OAM1jJsMGp"
   ],
-  "updated_at": "2026-07-19T18:56:33+00:00",
+  "updated_at": "2026-07-19T18:56:37+00:00",
   "root": {
     "slug": "index",
     "title": "Repro - LOGDIFF Exact Boolean Guidance",
@@ -85,6 +85,6 @@
       }
     ]
   },
-  "agent_view_tokens": 3150,
-  "revision": "1784487393151426000"
+  "agent_view_tokens": 3626,
+  "revision": "1784487397121561000"
 }

pages/claim-5-completeness-certificate/page.md

@@ -1 +1,15 @@
 # Claim 5 - completeness certificate
+
+
+---
+<!-- trackio-cell
+{"type": "markdown", "id": "cell_350f94a88f97", "created_at": "2026-07-19T18:56:34+00:00", "title": "Exhaustive Proposition C.2 certificate"}
+-->
+Two constructive cases from Proposition C.2 are evaluated against an independent full-joint oracle across 25 posterior settings.
+
+| Case | Exhausted semantic events | Max probability error | Max score error |
+| --- | ---: | ---: | ---: |
+| Independent categorical groups | 20,650 | 2.22e-16 | 3.33e-16 |
+| Nested or mutually exclusive taxonomy | 6,350 | 2.22e-16 | 2.22e-16 |
+
+All checks pass. The taxonomy has 15 predicates over 8 leaves, and the fail-closed control rejects an overlapping pair that is neither nested nor mutually exclusive.

pages/source-scale-cmnist-classifier/page.md

@@ -1 +1,22 @@
 # Source-scale CMNIST classifier
+
+
+---
+<!-- trackio-cell
+{"type": "markdown", "id": "cell_26911a0d4863", "created_at": "2026-07-19T18:56:36+00:00", "title": "Exact upstream classifier trained on the full source split"}
+-->
+The missing noise-aware CMNIST composition classifier was rebuilt using the pinned author trainer, its 50-epoch configuration, all 60,000 training examples, and the expected epoch=49-step=2950 checkpoint. SHA-256: a77cafc37a7be88eaec5aead5ffe024a995fbf2421d736ff009b3c9dd96d36ba.
+
+A deterministic audit evaluates all 10,000 test examples at fixed diffusion times.
+
+| t | Digit accuracy | Colour accuracy |
+| ---: | ---: | ---: |
+| 0 | 97.56 percent | 100.00 percent |
+| 50 | 97.20 percent | 100.00 percent |
+| 100 | 96.38 percent | 99.98 percent |
+| 250 | 83.67 percent | 97.15 percent |
+| 500 | 33.89 percent | 48.82 percent |
+| 750 | 13.09 percent | 15.28 percent |
+| 999 | 10.64 percent | 10.79 percent |
+
+The decay toward chance at maximum noise is expected and is now measured rather than hidden inside one random-timestep average.