Edge guards, a one-page report template, and acceptance checks.
E7) Edge cases & guardrails
# Mismatched candidate counts Aggregate at the decision level (per item). If per-request pairing is required, first pool within each request in u-space, then pool across requests. # Missing telemetry for gate Set g := 1 and flag the row; still log RSI and band. # Vendor-specific truncation or calibration Do not normalize RSI post-hoc. Comparability comes from (-1,+1) bounds. # Shard merges Only merge (U,W). Never average RSI directly. # Fallback on breach If acceptance gates fail (parity, order invariance, clamp bounds, gate purity), revert analysis to classical m-based baselines for that slice and flag the bucket. # Numeric hygiene Always clamp before atanh. Carry (U,W) as float64. Guard denominators with max(W, eps_w).
E8) Report template (per bucket)
bucket RSI_pool_env 95% CI (u-space) A++/A+/A0/A-/A-- tokens lat_ms Vendor A 0.510478 [+0.08, +0.21] 0/1/2/0/0 0.92x 310 Vendor B 0.393545 [ 0.00, 0.00] 0/1/2/0/0 1.00x 345 # Verdict line (bounded index) ΔRSI ≈ +0.116933 in favor of A, significant (p ≈ 0.03) computed in u-space (mean(atanh(RSI_env_A)) - mean(atanh(RSI_env_B)) with bootstrap).
E9) Acceptance checklist (pass/fail)
Determinism Same manifest + inputs ⇒ identical RSI_pool and RSI_pool_env Order/shard Permutations and shard merges leave pools unchanged (U/W fuse) Boundedness All RSI, RSI_env in (-1,+1); pooled results too Stamp completeness Each bucket summary includes knobs_hash No mutation of m Verified by collapse parity: phi((m,a)) = m
One-line takeaway. Publish a stamped, one-page bake-off per surface: bounded chooser, u-space pooling, CI in rapidity space — order-invariant, shard-safe, reproducible, and classical values remain untouched via phi((m,a)) = m.
Navigation
Previous: SSM-AI – Appendix E — Vendor Bake-off Protocol (E4–E6)
Next: SSM-AI – Appendix F — SSM-Audit CFO Pack (3–5 KPI lanes, weekly roll-ups, ROI)
Directory of Pages
SSM-AI — Table of Contents