Tiny sweeps that explain the lift
Purpose. Run small, orthogonal sweeps to show which knobs drive quality/efficiency gains while keeping classical outputs intact via phi((m,a)) = m.
Knobs to sweep (independent).
• Weights policy (gamma) in w := |m|^gamma → {0, 0.5, 1, 2} (use w := 1 for pure comparability).
• Lens gain (c) in a := tanh(c*e) → {0.7, 1.0, 1.3}.
• Gate mode → {"mul", "u_scale"} where RSI_env := g*RSI (mul) or RSI_env := tanh(g*atanh(RSI)) (u_scale).
• Prior strength (beta) in u-space → {0, 0.1, 0.2} with bounded b ∈ [-1,+1].
Fixed invariants.
All runs must satisfy boundedness and parity: |a|<1, |RSI|<1, |RSI_env|<1, and phi((m,a)) = m.
Sweep scaffolding (copy-paste).
# Grid (example)
Gamma := {0, 0.5, 1.0, 2.0}
C := {0.7, 1.0, 1.3}
Gate := {"mul","u_scale"}
Beta := {0.0, 0.1, 0.2}
for g in Gamma:
for c in C:
for mode in Gate:
for beta in Beta:
run_id := stamp(g,c,mode,beta)
# compute RSI, RSI_env with declared manifest + (g,c,mode,beta)
# log metrics and bands for Δ vs baseline
Record these for each cell (Δ = SSM-AI − Baseline).
first_pass_correct_delta
retries_delta
t_first_correct_delta
exposed_risk_delta # incorrect in A-/A--
tokens_per_solved_delta
calls_per_solved_delta
stability_eps # expect ~0 across batch/stream/shuffled
pearson_r, spearman_r
sat_rate := mean(|a| > 0.9) # want < 0.10
dead_zone := mean(|a| < 0.1) # want < 0.70
Interpretation rules (quick).
• If sat_rate > 0.10, reduce c or increase Unit.
• If dead_zone > 0.70, increase c or lower Unit; try gamma := 0 (uniform).
• If high-confidence damping is too harsh, switch gate to "u_scale".
• If vendor comparability is primary, prefer gamma := 0 and publish it.
• If signal strength should matter, use gamma := 1 and report sensitivity.
Mini ablation table (example layout).
Gamma C Gate Beta ΔFirst% ΔRetries ΔTokens sat_rate dead_zone Notes
0.0 1.0 mul 0.0 +3.1% -18% -9% 0.06 0.58 good default
1.0 1.0 u_scale 0.1 +4.2% -22% -12% 0.08 0.55 gentler at A+/A++
2.0 1.3 mul 0.2 +4.4% -25% -10% 0.14* 0.41 *saturated → reduce c
One-line takeaway. Small, orthogonal sweeps over gamma, c, gate mode, and beta reveal a stable operating point with saturation < 10% and dead-zone < 70%, while collapse parity phi((m,a)) = m guarantees classical values never change.
Navigation
Previous: SSM-AI – Empirical Validation & Mini Benchmarks — Metrics (calculator-fast; no retraining) (6.2)
Next: SSM-AI – Empirical Validation & Mini Benchmarks — Results (tiny tables; reproduce from stamps) (6.4)
Directory of Pages
SSM-AI — Table of Contents
SHUNYAYA 