Update logbook: Repro: Representational Curvature Modulates Behavioral Uncertainty in Large Language Models

logbook.json

@@ -8,7 +8,7 @@
     "icml2026-repro",
     "paper-fufl3hBXMq"
   ],
-  "updated_at": "2026-07-17T21:48:50+00:00",
+  "updated_at": "2026-07-17T21:50:58+00:00",
   "root": {
     "slug": "index",
     "title": "Repro: Representational Curvature Modulates Behavioral Uncertainty in Large Language Models",
@@ -64,6 +64,6 @@
       }
     ]
   },
-  "agent_view_tokens": 9464,
-  "revision": "1784324930791081900"
+  "agent_view_tokens": 9714,
+  "revision": "1784325058383601200"
 }

pages/methods-and-provenance/page.md

@@ -13,3 +13,10 @@
 {"type": "markdown", "id": "cell_31e5e0e39e8d", "created_at": "2026-07-16T19:03:32+00:00", "title": "Spot check PASSED (HF Jobs t4-small, job 6a592a82, ~2 min, ~$0.01). GPT-2 small…"}
 -->
 **Spot check PASSED (HF Jobs t4-small, job 6a592a82, ~2 min, ~$0.01).** GPT-2 small, 10 LAMBADA passages, methodology validated end-to-end: (1) manifest hash verified in-job; (2) block-output hooks cross-checked against output_hidden_states — caught TWO real traps: hidden_states[-1] is post-ln_f (raw hook differs, ln_f(hook)==hidden_states[-1] verified exactly), and GPT2Block returns tensor not tuple in current transformers (hook now shape-asserted); (3) mean contextual curvature 113-118 deg across layers — matches paper Fig 2's 112-120 deg range; (4) raw corr(C,H) positive at all layers (+0.06 to +0.10 on 10 passages, GPT-2 small) — correct sign and magnitude ballpark pre-CV; (5) manual token table semantically sane (mid-BPE-word continuations show ~0 bits entropy, open contexts ~10 bits). Entropy/curvature ranges asserted. UD manifest also frozen: en_ewt train+test with Sec-2.2 filters = 5927 sentences vs paper's 5815 (+1.9%), closest reconstruction; documented as approximate.
+
+
+---
+<!-- trackio-cell
+{"type": "markdown", "id": "cell_d395d7699623", "created_at": "2026-07-17T21:50:55+00:00", "title": "Reproduction workspace + data bundle (artifacts). All scripts, adversarial revi…"}
+-->
+**Reproduction workspace + data bundle (artifacts).** All scripts, adversarial reviews, and progress log: https://huggingface.co/datasets/solarkyle/curvature-repro-results under workspace/. Per-claim data (per-token geometry/entropy parquets, perturbation results, checkpoint sweeps, training metrics+checkpoints, Gemma extension) under runs/, training/, training_eval/, extension/; figures under figures/; frozen dataset manifests under configs/. Dual-review reports (Codex GPT-5.6 + Opus 4.8) for claims 1, 2, and final verdicts under workspace/reviews/. Total GPU spend ~$35 on HF Jobs (T4/L4/A10G/A100). Extension code also merged into github.com/solarkyle/jspace under curvature/.