Neural Tissue as a Third Domain of the Computational-Primes Compiler

A compiler splitting a computation between physics and digital logic needs a cost model: a numeric feasibility condition per elementary operation and a price per domain hand-over. A published framework supplies both for silicon, over twelve operations (computational primes). We state them for cortical tissue, which carries every prime in some form but none unconditionally. Tissue is a third compiler domain, not a noisy analogue chip: the cost model inverts on the time axis. Fast electrical signalling is the physics side. The receptor-molecule count storing a synaptic weight is the slow digital side, discrete but not binary. Writing a weight is 2.8–4.6 orders of magnitude slower than the forward step, spike to calcium to kinase, and more precise (4.7 against 1–2 bits). Yet it costs only 4.2 times more energy per event (triggering spike excluded). The hand-over is priced in latency, not energy; the assignment rule reverses the chip's: compute electrically, write to the molecular store as rarely as the rate budget allows. Only erasure is capped by energy, near 0.09 erasures per second and synapse. Three more constraints apply. No read-only mode: recall-type retrieval writes the retrieved item (1.3–2.4 restudy presentations, est.); holding writes far less, with little or inconsistent priority selectivity. Redundancy fails under common-mode error: no group size suffices once weights frozen inside a group correlate above 0.13–0.22. Selection costs neurons per item (6–12 at a proxy value); its numbers rest on an unmeasured rate contrast between items. The mechanisms are known; the contribution is a catalogue of conditions.