SSM-AI – Integration Quickstarts: LLM Decoding Hooks (8.1)

Add a bounded chooser without touching logits or probs.

Where to hook. After you have candidate tokens/logits and any side signals for the lens.

Minimal flow.

  1. Compute contrasts e_in, e_out.
  2. Map to alignments: a_in := tanh(-c*e_in), a_out := tanh(+c*e_out).
  3. Chooser (rapidity mean):
    RSI := tanh((SUM w*atanh(a_out) - SUM w*atanh(a_in)) / max(SUM w, eps_w)).
  4. Gate: RSI_env := g_t * RSI (or curvature-preserving mode "u_scale" → RSI_env := tanh(g_t * atanh(RSI))).
  5. Pick by RSI_env; keep classical m intact via phi((m,a)) = m.
    Clamp rule (always). Before any atanh, clamp: a_c := clamp(a, -1+eps_a, +1-eps_a).
    Weights policy. Default w := |m|^gamma with gamma = 1; w := 1 if declared.
# LLM decoding hook (beam/greedy, callback-style)
def on_candidates(cands, lens, g_t=1.0, eps_w=1e-12, eps_a=1e-6, gate_mode="mul"):
    scored = []
    for cand in cands:  # cand carries m (logprob/prob) and lens_items: [(e_in, e_out, w), ...]
        U_in = 0.0; V_out = 0.0; W = 0.0
        for (e_in, e_out, w) in cand.lens_items:
            a_in  = tanh(-lens.c * e_in)
            a_out = tanh(+lens.c * e_out)
            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 += w * atanh(a_in)
            V_out += w * atanh(a_out)
            W    += w

        if W <= 0:
            RSI  = 0.0
            band = "A0"   # insufficient_evidence
        else:
            RSI  = tanh((V_out - U_in) / max(W, eps_w))
            band = to_band(RSI)  # A++/A+/A0/A-/A--

        RSI_env = (tanh(g_t * atanh(RSI)) if gate_mode == "u_scale" else g_t * RSI)
        RSI_env = max(-1+eps_a, min(1-eps_a, RSI_env))
        scored.append((cand, RSI_env, RSI, band, U_in, V_out, W, g_t))

    best = max(scored, key=lambda kv: kv[1])[0]
    return best  # selection by RSI_env; classical m remains intact (phi((m,a)) = m)

Stamp fields (suggested).
token_id, m, RSI, RSI_env, band, U_in, V_out, W, g_t, lens_id, Unit, c, eps_a, eps_w, weights_policy, combine_policy, gate_mode, dtype, knobs_hash, stamp

One-line takeaway. Compute RSI in rapidity space, apply a calm gate, select by RSI_env, and stamp — your logits/probs stay exactly as they are (phi((m,a)) = m).

Navigation
Previous: SSM-AI – Robustness & Troubleshooting (7.6, 7.7)
Next: SSM-AI – Integration Quickstarts: RAG Pipeline Hook (8.2)

Directory of Pages
SSM-AI — Table of Contents