SSM-AI – Two-Minute Gate Calibration (4.5)

From raw telemetry to a stamped gate in five steps.

4.5 Two-minute calibration plan (reproducible)

Goal. Pick weights and smoothing so the gate turns turbulence into a transparent, bounded brake on decisions—without ever touching m (collapse parity phi((m,a)) = m).

Protocol (paste-ready).

Sample (no PII).
Collect 1–2k recent steps with lanes F,D,L,E,V,Q and baseline RSI.
Normalize lanes to [0,1].
Use simple, declared transforms (or your existing ones):

F := clamp(frac_violations, 0, 1)
D := clamp((1 - cos_sim)/max(1 - cos_ref, 1e-12), 0, 1)
L := clamp((latency - p50)/max(p95 - p50, 1e-12), 0, 1)
E := clamp(error_count/max(calls,1), 0, 1)
V := clamp(edit_distance/max(length,1), 0, 1)
Q := clamp(queue_depth/max(q95,1e-12), 0, 1)

Grid a tiny search over weights & smoothing.
Try w ∈ {0.5, 1.0, 1.5} for each lane, rho ∈ {0.10, 0.20, 0.30}, mode ∈ {"mul","u_scale"}. For each combo:

W := wF + wD + wL + wE + wV + wQ
mix := (wF*F + wD*D + wL*L + wE*E + wV*V + wQ*Q)/max(W,1e-12)
g_inst := clamp(1 - mix, 0, 1)
g_t := (1 - rho)*g_prev + rho*g_inst
RSI_env_mul    := g_t * RSI
RSI_env_uscale := tanh(g_t * atanh(RSI))

Score against day-one KPIs (observation-only).
Prefer the setting that improves: retries↓, time-to-first-correct↓, bad escalations↓—while keeping parity:

# Purity check (must hold)
phi((m,a)) = m

Freeze & stamp the manifest.
Publish chosen weights, rho, g_min, mode, bands policy, and acceptance notes. Compute a knobs hash and include in your ASCII stamp.

Quick worksheet (copy/paste).

# Inputs: columns F,D,L,E,V,Q,RSI, time
W      = wF + wD + wL + wE + wV + wQ
mix    = (wF*F + wD*D + wL*L + wE*E + wV*V + wQ*Q)/max(W,1e-12)
g_inst = max(0, min(1, 1 - mix))
g_t    = (1 - rho)*g_prev + rho*g_inst
RSI_env = (mode == "u_scale") ? tanh(g_t * atanh(RSI)) : g_t * RSI
band    = to_band(RSI_env)

Worked mini (calculator-fast).

RSI = 0.70
F=0.20, D=0.10, L=0.30, E=0.15, V=0.20, Q=0.10
wF=1.0, wD=1.0, wL=1.0, wE=1.0, wV=0.5, wQ=0.5  -> W=5.0
mix = (0.20+0.10+0.30+0.15+0.10+0.05)/5.0 = 0.18
g_inst = 0.82 ; rho = 0.20 ; g_prev = 1.00 -> g_t = 0.964
mode="mul":     RSI_env = 0.964 * 0.70  = 0.6748  -> band A+
mode="u_scale": RSI_env = tanh(0.964*atanh(0.70)) ≈ 0.688 -> band A+

Acceptance checklist (pass/fail).

# Gate purity: alignment-only
assert phi((m,a)) == m

# Monotonicity: worse lanes -> g_t must not increase
assert worsen(F|D|L|E|V|Q) ⇒ g_t_new <= g_t_old + tol

# Bounds
assert 0 <= g_t <= 1
assert |RSI_env| < 1

# Determinism
same inputs + manifest ⇒ identical g_t, RSI_env (within dtype tolerance)

Minimal manifest diff (freeze).

"gate": {
  "weights": {"F":1.0,"D":1.0,"L":1.0,"E":1.0,"V":0.5,"Q":0.5},
  "rho": 0.20,
  "g_min": 0.00,
  "mode": "mul",              # or "u_scale"
  "eps_g": 1e-12,
  "band_policy": {
    "promote_if": "RSI_env >= A+",
    "pause_if":   "RSI_env in A0",
    "block_if":   "RSI_env <= A-"
  }
}

One-line takeaway. In two minutes: normalize lanes → grid tiny weights/smoothing → pick the setting that improves retries/time without touching m → freeze the manifest → stamp. RSI_env := gate(RSI); phi((m,a)) = m remains inviolable.

Navigation
Previous: SSM-AI – Ready-Made Gate Presets (4.4)
Next: SSM-AI – Gate Stamp Fields (4.6)

Directory of Pages
SSM-AI — Table of Contents