SSM-AI – Core Mechanics at a Glance (0B)

Kernel, streaming fuse, bands, and selection

Datum and collapse
Carry two lanes for every quantity you already compute.
x := (m, a) with a in (-1,+1) and collapse parity phi((m,a)) = m.

Clamp
Bound alignment before any map or composition.

a_c := clamp(a, -1+eps_a, +1-eps_a)   # default eps_a = 1e-6

Rapidity map
Compose safely in u-space, then map back to a bounded lane.

u  := atanh(a_c)
a' := tanh(u')     # after composition in u

Lane for mul/div (M2 policy, default)
Multiply or divide m as usual. Compose the lane with signed rapidities.

a' := tanh( atanh(a1) +/- atanh(a2) )
# "+" for product, "−" for ratio, declared in manifest

Optional M1 policy (multiply-only, rare, opt-in)
Only when declared and both operands share one lens with small |a|.

a' := clamp(a1 * a2, -1+eps_a, +1-eps_a)

Streaming fuse, order-independent
Batch equals stream equals shuffled within fixed epsilon.

U += w*atanh(a)
W += w
a_out := tanh( U / max(W, eps_w) )    # default eps_w = 1e-12

Weights
Pick once and pin in the manifest.

w := |m|^gamma    # default gamma = 1
# or
w := 1            # uniform

Lens → align → RSI (single bounded chooser)
Turn a declared contrast into a bounded chooser used for selection.

# given contrasts e_in, e_out and scale c>0
a_in  := tanh(-c*e_in)
a_out := tanh(+c*e_out)

U_in  := SUM w*atanh(a_in)
V_out := SUM w*atanh(a_out)
W_in  := SUM w

RSI := tanh( (V_out - U_in) / max(W_in, eps_w) )   # RSI in (-1,+1)

Calm gate (alignment-only, read-only actuation)
Scale selection pressure without touching values.

RSI_env := g_t * RSI      # g_t in [0,1] from telemetry

Bands (global defaults, normative)
Publish once and use everywhere for legibility and policy.

A++: a >= +0.90
A+ : +0.60 <= a < +0.90
A0 : -0.60 <  a < +0.60
A- : -0.90 <  a <= -0.60
A--: a <= -0.90

Division policy, normative
Keep division safe and explicit.

division_policy := "strict"   # default
# alternatives "meadow" or "soft" require explicit declaration + numeric guards
# combine policy for alignment remains M2 unless declared otherwise

Non-negotiable invariants, normative
(i) Collapse parity. phi((m,a)) = m
(ii) Order invariance. Batch == stream == shuffled within epsilon
(iii) Determinism. Same manifest implies same outputs
(iv) Lane purity. Policies, bands, and gates act on a only. m is never mutated
(v) Division near zero. Default strict, alternatives declared
(vi) Cross-vendor parity. Identical manifests yield comparable a and RSI

Dtype and epsilon note
Prefer float64 for u-space accumulation. If constrained to float32, set eps_w >= 1e-8 and keep |a| < 1 - eps_a with eps_a >= 1e-6 to avoid atanh(±1) and division by zero in a_out := tanh(U / max(W, eps_w)).

Manifest crib, minimum fields
Pin the knobs that govern reproducibility and parity.

eps_a, eps_w, gamma, weights_policy,
combine_policy=M2, division_policy=strict,
bands{...}, lens_id, lens_params, g_t_source,
dtype, knobs_hash


Navigation
Previous: SSM-AI – Under the Hood & Integration (0A continued)
Next: SSM-AI – Surfaces → Kernel → Acceleration (0C)


Directory of Pages
SSM-AI — Table of Contents


Explore Further
https://github.com/OMPSHUNYAYA/Symbolic-Mathematical-AI