Vendor delta analytics and weekly roll-up reducer for CFO and ops teams.
N4) Vendor delta worksheet (copy-paste; per service)
All *_pct columns are fractions in [0,1]; positive = savings vs baseline.RSI_pool_env_delta := RSI_pool_env_A - RSI_pool_env_B.
svc comparison tokens_saved_pct retry_drop_pct latency_saved_pct unit_cost_saved_pct RSI_pool_env_delta weekly_savings_usd
search Vendor A vs B 0.0876 0.3 0.1014 0.1 0.151 182400
tools Vendor A vs B 0.0957 0.3333 0.1026 0.1 0.157 96300
Formulas (illustrative):
RSI_pool_env_delta := RSI_pool_env_A - RSI_pool_env_B
weekly_savings_usd := savings_usd_A - savings_usd_B
Reference quantities (declare once per worksheet):
T_week := tokens_per_week_in_1k_units
P_ref := dollars_per_1k_tokens_baseline_or_blended
R_week := requests_or_tool_calls_per_week
C_tool := dollars_per_tool_or_API_call
C_ms := dollars_per_ms_latency # optional, set 0 if not monetized
Spend_ref_week := baseline_weekly_AI_spend_for_this_service
Per-vendor decomposition:
savings_usd_i :=
T_week * P_ref * tokens_saved_pct_i
+ R_week * C_tool * retry_drop_pct_i
+ R_week * C_ms * lat_saved_pct_i
+ Spend_ref_week * unit_cost_saved_pct_i
Guardrail (band parity).
Count savings_i only if RSI_pool_env_i >= band_min (e.g., A0).
N4.1) Adoption premise (zero extra infra)
SSM-AI runs directly on your existing stack.
It requires only symbolic math and a small manifest — no retraining, no new services, and no PII.
Classical numbers remain unchanged via phi((m,a)) = m.
Selection and routing use bounded alignment (a ∈ (-1,+1)) and a chooserRSI := tanh( (V_out - U_in) / max(W_in, eps_w) ).
Order/shard invariance follows fromU += w*atanh(a) ; W += w ; a_out := tanh( U / max(W, eps_w) ).
N4.2) Savings premise (plug into worksheets)
Use Annual_Savings ≈ S_base * r_save with r_save ∈ [0.10, 0.20].
Typical ranges:
- Large orgs: $500000–$2000000
- Mid orgs: $30000–$160000
- NGO/public: $5000–$40000
Decompose into tokens + tool/API + vendor arbitrage + capacity deferral, all stamped for replay.
N4.3) Acceptance checks (pass/fail)
- Band parity. Compare vendors only where both meet
band_min(e.g., A0). - Order/shard invariance.
RSI_pool_envcomputed via(U,W); batch == stream == shard. - Apples-to-apples. Same manifest (
knobs_hash), prompts, datasets. - Unit discipline. All
*_pctcolumns are fractions in[0,1]; avoid mixed % formats. - Replayability. Recomputing from
(U,W)and manifest must reproduceRSI_pool_envandweekly_savings_usdwithin dtype tolerance.
N5) Roll-up reducer (reference pseudocode)
def weekly_rollup(rows, eps_w=1e-12):
# rows: per-decision/session logs with U,W,tokens,lat_ms,retries,cost
by_key = {} # (week_start, svc, vendor, knobs_hash) -> accum
for r in rows:
k = (r.week_start, r.svc, r.vendor, r.knobs_hash)
acc = by_key.setdefault(k, {
"U":0.0,"W":0.0,"tokens":0,"succ":0,
"lat_ms":[],"retries":0,"cost_tokens":0.0,"g":[]
})
acc["U"] += r.U_kpi
acc["W"] += r.W_kpi
acc["tokens"] += r.tokens
acc["succ"] += r.success
acc["lat_ms"].append(r.p95_ms)
acc["retries"] += r.retries
acc["cost_tokens"] += r.cost_tokens
acc["g"].append(r.g)
out = []
for k,acc in by_key.items():
U, W = acc["U"], acc["W"]
RSI_pool = tanh(U / max(W, eps_w)) if W > 0 else 0.0
g_week = max(0.0, min(1.0, sum(acc["g"])/max(len(acc["g"]),1))) if acc["g"] else 1.0
RSI_env = g_week * RSI_pool
band = to_band(RSI_env)
tps = acc["tokens"]/max(acc["succ"],1)
out.append({
"key": k,
"RSI_pool_env": RSI_env,
"band": band,
"tokens_total": acc["tokens"],
"tokens_per_success": tps,
"retry_rate": acc["retries"]/max(acc["succ"],1),
"cost_per_1k": acc["cost_tokens"]
})
return out
Quantiles. Maintain an external quantile sketch (e.g., t-digest) for p95_ms.
SSM-AI itself does not alter latency; it only provides order-invariant symbolic alignment.
Navigation
Previous: SSM-AI – Appendix N — Economics & Rollout Worksheets (N1–N3)
Next: SSM-AI – Appendix N — Economics & Rollout Worksheets (N6–N9)
Directory of Pages
SSM-AI — Table of Contents
SHUNYAYA 