Stimulating only AVB, AVA, SMDD, SMDV and RIV, an evolved policy reaches an odour source 2.5 mm away within 40 s in 91 % of trials (random policy: 4 %). The minimal circuit needs all three of reversal, deep bend and steering.
With only sensory channels (ASE, AWC, AWA, ASK, ASH, PLM, ALM, AVM) the same learner stays at chance (0.125). In this model the sensory-to-command step of the connectome is silent — a negative result confirmed both with static stimulation and with learning.
Given reversal and deep-bend channels only, evolution rediscovers the textbook rule: concentration falling → reverse + omega turn, rising → run (Pierce-Shimomura et al. 1999). Reversal onset rate falls monotonically with dC/dt in the full policy too.
In a corridor four body-widths wide, head steering alone cannot take a corner (0/32). Only a deep bend pressed against the wall does — and its dorsal/ventral direction is the turn direction. With ventral-only bends the worm enters the left arm of a T-maze on 128/128 trials and never takes a right corner.
Letting SMDD/SMDV set the bend direction (as reported for real worms, Nature Neurosci. 2026) makes the right-turn corridor learnable (reach 1.00) — but the learned policy becomes right-handed and loses the left corner.
With a left/right concentration difference at the nose (an idealised head-sweep sample), the policy chooses the bend direction toward the source in 73 % of pulses (62 % without it) and navigates as fast as an oracle that knows the direction.
In narrow corridors a wall stall reads as "concentration stopped rising", which triggers the open-field reflex reverse + ventral bend. That reflex is the only way to take a corner — and the reason a correctly aimed right entry is aborted. Learning settles on "go left first, reverse if worse".
In narrow T-mazes, arm choice should be decided by the dorsal/ventral direction of the bend executed at the junction; conditions that suppress dorsal turns should bias animals to the ventral arm and delay reaching the other goal. Wall stalls should trigger pirouettes.
Model: OpenWorm c302 (C2) neurons re-implemented in JAX with fitted dynamics ("Vfit"); Boyle, Berri & Cohen 2012 motor layer gated by AVB/AVA; 2D viscoelastic body with wall contact. Learning: evolution strategies over a small MLP. Full methods, decision records (ADR-001…017), and every negative result are in the repository: github.com/kairess/worm-whisperer. Korean documentation lives in docs/.