Quick checks to guarantee bounded behavior; fast fixes for common symptoms.
7.6 Robustness & Acceptance (quick checks)
• Collapse parity. phi((m,a)) = m before/after any lane/gate step.
• Order/shard invariance. Batch vs stream vs shuffled ⇒ same RSI (within tolerance).
• Boundedness. All a, RSI, RSI_env, RSI_path satisfy |x| < 1.
• Edge clamps. Inputs at or beyond ±1 clamp to ±(1 - eps_a).
• Division policy. Under "strict", near-zero divisors ⇒ fallback to classical actuation; lane still stamps context.
• Determinism. Fixed manifest ⇒ identical outputs (within dtype), across machines.
# Core invariants to assert in CI
assert phi((m,a)) == m
# Order/shard invariance (sketch)
RSI_batch := tanh( SUM w*atanh(a) / max(SUM w, eps_w) )
RSI_stream := tanh( U/W ) # carried U,W with same items
RSI_sharded := tanh( (SUM_shards U_k) / max(SUM_shards W_k, eps_w) )
assert close(RSI_batch, RSI_stream) and close(RSI_batch, RSI_sharded)
# Boundedness + clamps
a_c := clamp(a, -1+eps_a, +1-eps_a)
assert abs(a_c) < 1 and abs(RSI) < 1 and abs(RSI_env) < 1
7.7 Troubleshooting (symptoms → fixes)
• Symptom: frequent a near ±1 (saturation).
Fix: reduce lens gain c or increase Unit; verify calibration.
• Symptom: RSI hovers near 0 (dead-zone).
Fix: increase c (or adjust Unit); consider uniform weights w := 1.
• Symptom: slight drift between batch and stream.
Fix: clamp before atanh; use pairwise/Kahan sum for U; check eps_w.
• Symptom: unstable ratios.
Fix: ensure division_policy = "strict"; verify bounds for denominators; lane math stays M2.
• Symptom: gating too aggressive at high confidence.
Fix: switch gate to “u_scale” mode.
# Strict division policy (control, not value)
division_policy := "strict"
if |m_den| <= eps_div:
use_lane_for_actuation := False # fallback to classical selection
stamp("DIV_GUARD")
# Gate alternatives
# multiply mode (default):
RSI_env := g_t * RSI
# curvature-preserving mode:
RSI_env := tanh( g_t * atanh(RSI) )
One-line takeaway. Accept only what’s bounded, invariant, and reproducible: assert collapse parity, clamps, and shard/order equality; when symptoms show up, tweak gain/Unit, strengthen clamps or sums, and fall back to classical actuation under "strict" policy — never touching m.
Navigation
Previous: SSM-AI – Scalability: HW Parity & Performance (7.4, 7.5)
Next: SSM-AI – Integration Quickstarts: LLM Decoding Hooks (8.1)
Directory of Pages
SSM-AI — Table of Contents
SHUNYAYA 