Update logbook: Reproduction: Improved Dimension Dependence for Bandit Convex Optimization with Gradient Variation

logbook.json

@@ -11,7 +11,7 @@
     "icml2026-repro",
     "paper-X8evkEdMxb"
   ],
-  "updated_at": "2026-07-23T15:57:10+00:00",
+  "updated_at": "2026-07-23T15:57:19+00:00",
   "root": {
     "slug": "index",
     "title": "Reproduction: Improved Dimension Dependence for Bandit Convex Optimization with Gradient Variation",
@@ -67,6 +67,6 @@
       }
     ]
   },
-  "agent_view_tokens": 6420,
-  "revision": "1784822230436092000"
+  "agent_view_tokens": 6696,
+  "revision": "1784822239074189000"
 }

pages/claim-4-small-loss-f-t/page.md

@@ -8,3 +8,12 @@
 **VERIFIED.** Theorem 4 gives O(sqrt(d F_T)+d) for convex and linear functions, with F_T defined as the best fixed cumulative loss on X minus the sum of per-round minima on the enlarged domain X+. Appendix C.7 plugs the lemma E[Vbar_T] <= 16 d L E[F_T^w] into the regret decomposition.
 
 The audit recomputed the displayed envelope over d=16...256 and F_T=1,16,256,1024, recording both the theorem scale sqrt(d F_T)+d and the visible 6R sqrt(d^2+16 d L F_T) proof envelope. Artifacts: `outputs/audit/claim4_ft_formula_audit.csv`; source: https://arxiv.org/abs/2602.04761.
+
+
+---
+<!-- trackio-cell
+{"type": "markdown", "id": "cell_8e0958c35f41", "created_at": "2026-07-23T15:57:16+00:00", "title": "Repair audit verdict"}
+-->
+**VERIFIED.** The repair audit computes F_T directly as best fixed cumulative loss on X minus per-round minima on the enlarged domain X+. For linear losses the ball minima are closed form; for convex diagonal quadratics the same trust-region closed form is used. The 20 rows are in `outputs/anchored/wt_ft_direct_diagnostics.csv`, and the summary reports max positive regret/envelope 0.0 and max absolute regret/envelope 0.31463 for the paper-visible envelope 6 sqrt(d^2 + 16 d F_T).
+
+Controls confirm the enlarged-domain quantity is nonnegative across the executed rows, and the source rows include the X+ and 16dL appearances used in the proof audit. Artifacts: `outputs/anchored/anchored_controls.csv`, `outputs/anchored/anchored_summary.json`; Hub Space: https://huggingface.co/spaces/Srishti280992/repro-bco-gradient-variation .