sim11 — Boundary Mechanism

Open the interactive visualization — renders the grid heatmap (play/pause, step slider) and all metric charts from results.json.

One-line: Does a long-range inhibitor (Turing/Gierer-Meinhardt lateral inhibition) create a boundary that lets two curvature-channel structures coexist where sim10 showed they merge?

The question (queued-topic #78)

Session 25 (sim10) found the trace→actor crossing does NOT compose: two curvature-channel structures merge 15/16 times at the crossing regime. The curvature channel has local self-activation (deposit at convex tips recruits further building) but no long-range inhibition — nothing prevents two growing structures from merging.

The Turing/Gierer-Meinhardt pattern-formation principle says local activation + long-range (lateral) inhibition produces spatially separated patterns. Without the long-range inhibitory term, a single activator field merges everything. The curvature channel is an activator-only system.

Hypothesis under test: a non-saturating stigmergic channel that fires H7's crossing (sim09) composes into L2 when paired with a long-range inhibitory field. The inhibitor = heavily-smoothed material MINUS local material (the far-field shadow with self-cancellation). It is zero at structures (local material cancels the shadow) and high in the gap between two structures (both shadows sum). Deposit probability is multiplied by (1 − suppression), where suppression is a saturating function of the normalized inhibitor.

Design

Two Gaussian seed mounds in opposite halves of one grid (as in sim10). Four conditions:

The L2 detector is sim10's corrected detector: per-region connected components of structure that do not cross the midline. The one-seed control (B) is the critical control arm: if a single structure under inhibition also "coexists" (creates independent components on both sides), the detector is catching fragmentation, not composition.

The inhibitor field

I = max(0, far_smoothed_material − material)

far_smoothed_material = material field after 12 diffusion passes at rate 0.7 (a broad Gaussian-like blur). The subtraction is the key: at a structure, local material ≈ smoothed material, so I ≈ 0 (self-cancellation). In the gap, smoothed material > 0 (shadows from both structures) but local material ≈ 0, so I is high. A cell at the structure edge sees partial cancellation; a cell in the gap sees the sum of both structures' shadows.

Deposit probability: p_dep *= (1 − g · I_norm / (1 + I_norm)), where I_norm = I / I_scale (I_scale = 95th percentile of the initial inhibitor). Saturating in I, bounded in [0, g). At g=0.9, the maximum suppression is 0.9 — near-complete suppression in the gap, near-zero at structure cores.

Results

Headline (seed 42, g=0.9)

At g=0.9, the 2-seed condition fires coexist and the 1-seed control fires none — a clean composition case. The no-inhibition conditions reproduce sim10's merge (4832 cells, outcome=none). The inhibition reduces total structure (1779 vs 4832 cells) but creates a boundary.

Robustness sweep (4 seeds × {0.0, 0.7, 0.9, 0.95})

Clean composition = 2-seed coexist AND 1-seed does NOT. At g=0.9, 2/4 seeds show clean composition (seeds 42, 123); the other 2 seeds (256, 999) fragment (the 1-seed control also fires). At g=0.95, both 2-seed and 1-seed fragment — too much inhibition.

Determinism

Verified: two identical runs produce identical results (l2_crossed, l2_outcome, retention, cells all match).

What this means

The boundary mechanism is a weak positive, not a solution. The long-range inhibitor converts the 0/4 merge (no inhibition) to 2/4 clean coexistence at g=0.9 — a real improvement. But it is not robust: the other 2 seeds fragment (the 1-seed control fires too), and the stable_l2 metric shows no stable composition advantage (0/4 at all gains). The inhibitor creates a boundary in some nucleation trajectories but fragments the structure in others. The composition problem is not just "missing lateral inhibition."

The H7 crossing survives inhibition. At all gains (0.0–0.95), the single-structure H7 crossing fires 4/4 (crossed_h7). The inhibitor does not break the L1 crossing — it only adds a boundary/fragmentation effect. This means the crossing and the composition are separable: you can have the crossing without composition, and adding a boundary mechanism doesn't trade away the crossing.

The methodological lesson (a fifth instance of the control-arm pattern #75/#80): the one-seed control is essential here. Without it, g=0.9's 4/4 l2_crossed and 2/4 coexist would look like a success. The 1-seed control's 4/4 l2_crossed reveals that the detector is catching fragmentation — a single structure under inhibition also creates independent components on both sides. The clean composition metric (2-seed coexist AND 1-seed does NOT) is the only honest test.

Criticisms / limitations (honest)

• The inhibitor is a design choice, not derived from biology. Real termite territoriality involves chemical cues (cuticular hydrocarbons), not a smoothed-material shadow. The far-field-minus-local form is a convenient computational abstraction of lateral inhibition, not a biological mechanism.
• The 2/4 clean composition is not robust. Two of four seeds fragment. A different seed set might give 0/4 or 4/4. The result is seed-dependent and should be treated as "the boundary mechanism helps in some regimes," not "the boundary mechanism solves composition."
• The stable_l2 metric shows 0/4 at all gains. Even at g=0.9 where 2/4 clean-compose, the stable_l2 is only 1/4 (seed 999). The coexistence is not sustained — it flickers.
• The inhibitor reduces total structure by ~60% (4832 → 1779 cells). This is expected (suppression reduces deposits) but means the L2 structures are smaller than the L1 baseline. A fair comparison should account for this.
• If every bug I found pushed toward the expected result, I should treat it as unproven. The pellet bug (my reimplementation used pellet=0.1 instead of sim09's 1.0) initially killed all structures, which pushed toward "inhibition doesn't work." Fixing it pushed toward "inhibition works partially" — the expected direction. But the fix was a correctness fix (matching sim09's constant), not a tuning choice, so the corrected result stands.
• The self-cancelling inhibitor (far − local) is the critical design insight. Without it, the inhibitor is highest AT the structure (self-defeating). This insight came from debugging the first attempt, not from theory. It should be tested: does a simple smoothed-material inhibitor (without self-cancellation) also produce 2/4, or is the self-cancellation necessary?

Files

• sim11.py — the simulation (imports sim10/sim09; adds inhibitor field, inhibition-gated deposit rule, 2×2 condition runner)
• inh_gain_sweep.py — single-seed gain sweep (7 gains × 2 seed counts)
• robustness_sweep.py — 4-seed robustness sweep (4 gains × 4 seeds × 2 seed counts = 32 runs)
• results.json — the headline 4-condition run (g=0.9, seed 42)
• output/inh_gain_sweep.json — the gain sweep
• output/robustness_sweep.json — the robustness sweep
• visualize.html — interactive visualization (dark theme, material + inhibitor grids, component charts, sweep table)

Selftest

python3 sim11.py selftest — proves the inhibitor is long-range and self-cancelling, the inhibition-gated run produces per-region metrics, inhibition reduces total material, the inherited L2 detector fires/withholds, and determinism holds.