SSM-AI – Appendix K — SSM-Search (K7–K9)

Guardrails, two-channel ranking code, and acceptance checks.

K7) Guardrails & Policies

  • Parity. Classical retrieval stays intact: phi((m,a)) = m.
  • Clamp. Enforce |a| < 1 with eps_a = 1e-6 before any atanh.
  • Visibility. Log m_retrieval, RSI or a_search, band, g_t.
  • Search-safe defaults. Missing features ⇒ treat as 0, set RSI := 0 (neutral), rely on m_retrieval.
  • Click/open policy (by band). A++/A+ open; A0 preview; A-/A– require extra confirmation.
  • Shard/stream parity. Always merge via (U,W) only; never average a or RSI directly.
  • Determinism. Same manifest ⇒ identical ranking and bands.

K8) Pseudocode (two-channel lens + ranker)

def lens_search(feat, cfg):
    p = cfg["alpha"]*feat.get("quality", 0.0) \
      + cfg["beta"] *feat.get("freshness", 0.0) \
      + cfg["gamma"]*feat.get("authority", 0.0)
    n = cfg["delta"]*feat.get("risk_penalty", 0.0) \
      + cfg["eta"]  *feat.get("coherence_penalty", 0.0)

    a_out = tanh(cfg["c"] * p / max(cfg["Unit_out"], 1e-12))
    a_in  = tanh(-cfg["c"] * n / max(cfg["Unit_in"], 1e-12))
    return a_in, a_out

def rank_results(items, cfg, g_t=1.0, eps_w=1e-12, eps_a=1e-6):
    ranked = []
    for it in items:
        a_in, a_out = lens_search(it["feat"], cfg)
        a_in  = max(-1+eps_a, min(1-eps_a, a_in))
        a_out = max(-1+eps_a, min(1-eps_a, a_out))
        U_in  = atanh(a_in); V_out = atanh(a_out); W = 1.0
        RSI   = tanh((V_out - U_in) / max(W, eps_w))
        RSI_e = g_t * RSI
        ranked.append({"doc_id": it["doc_id"], "m": it["m"], "RSI": RSI, "RSI_env": RSI_e})
    return sorted(ranked, key=lambda r: (r["RSI_env"], r["m"]), reverse=True)

Notes

  • Primary ranker: RSI_env (or RSI if g_t=1).
  • Tie-break: m_retrieval to preserve legacy behavior.
  • Bands: Use A++/A+/A0/A-/A– for UI/policy, not for averaging.

K9) Acceptance Checklist (must pass)

  • Order/shard invariance. Shuffle and per-shard merges reproduce identical lists (via (U,W)).
  • Boundedness. a_search, RSI, RSI_env strictly in (-1, +1); bands per manifest thresholds.
  • Parity. m_retrieval exactly matches the baseline system.
  • Drift sanity. Hour/day roll-ups from (U,W) yield the same a_pool across pipelines.
  • Determinism. Identical manifest ⇒ identical ranking, bands, and stamps.

Navigation
Previous: SSM-AI – Appendix K — SSM-Search (K4–K6)
Next: SSM-AI – Appendix L — Governance Quick Reference (L1–L3)

Directory of Pages
SSM-AI — Table of Contents