Remove stale FP-Stalker backfill claims (block is fully populated)

cano_paper_v2.md

@@ -195,9 +195,6 @@ model-agnostic perturbations. Gaussian provides little additional transfer
 protection (1.04x). C&W's transfer reduction is negative
 (anti-protective on small synthetics) -- its ratio is reported as n/a.
 
-> **Note.** Transfer numbers are from the utility-metric subset; the FP-Stalker
-> block does not yet have transfer values computed -- backfill planned.
-
 ## 3.3 Noise Utility Metrics
 
 **Table 3.** Per-strategy noise-quality metrics (n = 5,924 in-scope rows).
@@ -253,10 +250,13 @@ concentration on importance-ranked rather than variance-ranked features.
 | `synth_50u_20s` | 50 | 0.003 | 0.514 | 0.356 | n/a | n/a |
 | `synth_5u_30s` | 5 | -0.006 | 0.208 | 0.072 | n/a | n/a |
 
-> **Note.** "n/a" means a strategy was not evaluated on that dataset
-> (backfill planned). FP-Stalker (Vastel et al. [10]) is the closest dataset
-> to deployment conditions; on it, CANO closes most of the gap to Gaussian
-> (0.276 vs 0.340).
+> **Note.** "n/a" cells in Table 5 mark configurations where a strategy was
+> not run on a particular dataset; this primarily affects Laplace and PGD on
+> the smaller synthetic variants and on `keystroke_cmu_51users`, which were
+> added to the strategy roster after those datasets had already been
+> evaluated. FP-Stalker (Vastel et al. [10]) is the closest dataset to
+> deployment conditions and is fully populated across all six strategies; on
+> it, CANO closes most of the gap to Gaussian (0.276 vs 0.340).
 
 ## 3.6 Statistical Significance
 
@@ -311,14 +311,19 @@ adversaries.
   (HTillmann; BrFAST extended) are the next integration.
 - 9 synthetic features with artificial importance concentration.
 - `cybersec_intrusion` (2 users) excluded from aggregates.
-- Utility metrics on 5,924-row subset; backfill re-run planned.
+- Utility metrics (sparsity, KL, deviation, sensitivity) cover the 5,924-row
+  subset of runs from 2026-04-05 onward; older synthetic-only runs predate the
+  instrumentation and contribute only adaptive accuracy_reduction.
 - RL at 50 users; architectural changes needed for scaling.
-- Transfer numbers not yet computed for the FP-Stalker block.
+- Laplace and PGD were not run on the older small-synthetic variants
+  (`overlap_*`, `synth_NuMs` family) -- visible as `n/a` cells in Table 5.
 
 ## 4.3 Future Work
 
-1. Backfill transfer-attack and noise-utility metrics for the FP-Stalker block.
-2. Extend utility-metric coverage across all historical evaluation runs.
+1. Extend utility-metric coverage across the older synthetic runs (the
+   ~50,000 rows that predate the noise-quality instrumentation).
+2. Fill the remaining `n/a` cells in Table 5 by running Laplace and PGD on
+   the small-synthetic variants and on `keystroke_cmu_51users`.
 3. Formal DP guarantees for CANO's allocation mechanism.
 4. Larger RL training (1,000+ users); online policy updates in deployment.
 5. Theoretical analysis of conditions under which feature-weighted noise
@@ -385,5 +390,5 @@ generalizes better under realistic conditions.
 
 ---
 
-*Generated: 2026-04-26 03:41:25*
+*Generated: 2026-04-26 03:47:41*
 *Data source: 19 merged `eval_*.jsonl` files (68,885 raw configs, 54,281 in-scope after excluding `cybersec_intrusion`).*