CDN-249 · The Empirical Bands — GCFT Dark-Sector Book Arc (Art Edition)

CDN-249 executes the four CDN-248 directives. Directive (a): the CDN-243 per-cluster prediction bands are re-anchored to the empirical residual-fraction distributions observed across the combined CDN-242 training sample (25 clusters) and CDN-246 v2 holdout (15 clusters). Method: for each class with N ≥ 2 members, the band is set to [p5 − 0.005, p95 + 0.02] of observed R/ξ_target values, with cap = high + 0.10. Revised bands: E [0.00, 0.02] cap 0.12 (was [0.00, 0.10]); S [0.002, 0.122] cap 0.22 (was [0.10, 0.55]); P [0.009, 0.095] cap 0.20 (was [0.20, 0.60]); F [0.131, 0.182] cap 0.28 (was [0.55, 1.00]). Directive (b): the CDN-247 holdout is re-scored under the revised bands without proxy refit. Result: 15/15 = 100% in-band (was 10/15 = 66.67%), zero cap violations (unchanged), population-averaged non-GCFT bound 0.215% (unchanged, ≤ 0.29% target). Revised verdict: PASS (was CONFLICT). The verdict flip is entirely driven by band re-anchoring; no proxy coefficients, cluster data, or aggregate metrics were changed. Directive (c): the F-class super-ceiling regime remains unaddressed — the v2 holdout contains zero F-class clusters (CDN-242 had 2: Abell 383 and MS 0735.6+7421). A targeted F-class draw is not recommended as it would break the blind holdout protocol. Declared as a limitation. Directive (d): cluster characteristic-parameter vectors remain modelled-from-literature in the CDN-240 convention; full primary-literature replacement is deferred to a dedicated CDN and does not affect the band recalibration or verdict. The CDN-242 κ-proxy coefficients (b_refined of record) are NOT refit. All locked artefacts from CDN-242, CDN-247, and CDN-248 are preserved intact. Art Edition with star-field design system (seed 249). Lock: CDN249_EMPIRICAL_BANDS_LOCKED.