Weekly bounded KPI lanes, clear roll-ups, ROI — no change to your numbers.
F1) The CFO view (what they get)
- One bounded portfolio index (weekly).
RSI_port := tanh( U_port / max(W_port, eps_w) ). - Band shares.
%A++, %A+, %A0, %A-, %A--per service and portfolio. - Before vs SSM-AI worksheet. Token/cost/latency savings with audit-ready math.
- Stamped ledger. Replay and roll-up are calculator-fast and deterministic.
- Invariant. Classical numbers remain untouched:
phi((m,a)) = m.
F2) KPI lens library (declare 3–5 lanes; dimensionless)
All inputs normalized to [0,1]; publish transformations in the manifest.
Lane 1 — Cost efficiency (tokens & retries)
tokens_drop := clamp( (tokens_before - tokens_after) / max(tokens_before, eps), 0, 1 )
retry_drop := clamp( (retry_before - retry_after ) / max(retry_before, eps), 0, 1 )
e_cost := (alpha_tok * tokens_drop + alpha_ret * retry_drop) / Unit
Lane 2 — Latency health (tail aware)
on_time_rate := clamp( 1 - miss_rate, 0, 1 )
tail_p95_penalty := clamp( (p95_ms - SLO_ms) / max(SLO_ms, eps), 0, 1 )
tail_p99_penalty := clamp( (p99_ms - SLO_ms) / max(SLO_ms, eps), 0, 1 )
e_lat := (alpha_on * on_time_rate - beta_p95 * tail_p95_penalty - beta_p99 * tail_p99_penalty) / Unit
Lane 3 — Quality stability (retries/contradictions)
e_qual := (alpha_succ * success_rate - beta_retry * retry_rate - beta_contra * contradiction_rate) / Unit
Lane 4 — Incidents & escalations (optional)
mttr_drop := clamp( (mttr_before - mttr_after) / max(mttr_before, eps), 0, 1 )
mtbf_gain := clamp( (mtbf_after - mtbf_before) / max(mtbf_before, eps), 0, 1 )
e_inc := (alpha_mttr * mttr_drop + alpha_mtbf * mtbf_gain - beta_page * page_rate) / Unit
Mapping (symmetric, bounded)
a_out := tanh( +c * e_out )
a_in := tanh( -c * e_in )
# If a lane emits a single signed contrast e, use:
a := tanh( c * e )
Weights for roll-ups
# Use portfolio weights per row (declare once):
w := revenue_share # or
w := 1 # uniform
One-line takeaway. Publish 3–5 KPI lanes with a tanh-mapped, bounded chooser and roll up weekly via u-space fusion—clear to finance, trivial to replay, and observation-only with phi((m,a)) = m.
Navigation
Previous: SSM-AI – Appendix E — Vendor Bake-off Protocol (E7–E9)
Next: SSM-AI – Appendix F — SSM-Audit CFO Pack (F3–F5)
Directory of Pages
SSM-AI — Table of Contents
SHUNYAYA 