digital-drosophila/models/male-cns-lif/deterministic-runtime-v2/candidate/MaleCNS-NumericalContract-v1.md

MaleCNS-NumericalContract-v1

Status: immutable candidate contract for MaleCNS-LIF-runtime-v2-deterministic-candidate.

Frozen tensors and arithmetic

• State (V, g, ring-buffer arrivals, effective edge weights): IEEE-754 binary32.
• CSR indptr and edge ids: int64; target indices: int32; refractory/counts: int32.
• Effective edge weights are the frozen Phase 4C signed float32 weights; raw synapse counts remain unchanged.
• Canonical contribution key: ['target_idx', 'source_idx', 'edge_id']. Since global CSR edge ids are source-row ordered, stable (target_idx, edge_id) sorting is equivalent.
• Each target is owned by one thread and accumulated sequentially with round-to-nearest-even __fadd_rn. Scientific fan-out contains no atomic add.

Per-timestep total order

1. Read and clear the current exact ring-buffer slot.
2. g = float32(g * g_decay), then g = float32(g + recurrent_arrival).
3. Decrement refractory by one; while refractory remains nonzero, hold V = V_reset.
4. For available neurons, forward Euler: V = float32(V + alpha * (g - (V - V_rest))), with every elementary operation rounded to float32.
5. Detect spikes using the frozen comparison V >= V_threshold; reset and set exact refractory steps.
6. Apply same-step external forced-spike events in ascending target index; a neuron already spiking or refractory is not double-counted.
7. Enumerate spiking source indices in ascending neuron index. Expand each CSR row in edge-id order.
8. Stable-sort contributions by target only, preserving source/edge order, then perform fixed sequential segmented sums.
9. Add each segment once to the ring-buffer slot exactly 18 steps ahead.
10. Emit events in (simulation_step, neuron_idx) order. Simultaneous events never receive a hidden priority beyond neuron index.

No threshold epsilon, voltage rounding/snap, float64 model substitution, approximate delay, topology change, or parameter change is permitted.