sim06 — Termite Mound

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

One-line: A minimal Grassé stigmergy termite-mound model testing whether an accumulated environmental trace can cross from a passive coordination signal to a self-maintaining actor (hypothesis H7).

What it tests

sim06 is the concrete substrate for the H7 Trace→Actor Crossing Hypothesis: that accumulated stigmergic traces can cross from being passive coordination signals to being a self-maintaining actor — a structure that (1) has dynamics not reducible to individual traces, (2) constrains the agents that build it, and (3) resists perturbation by repairing itself.

Two conditions are contrasted:

• baseline — termites deposit soil pellets laced with cement pheromone; the pheromone decays and diffuses; the structure is inert (only maintained by termites happening to wander back).
• self_maintenance — the accumulated structure itself re-emits cement pheromone proportional to its material (the trace→actor feedback loop). The structure becomes a source of the very signal that recruits builders.

A crossing detector flags the timestep at which the trace field becomes an actor per H7's three operational criteria; a perturbation experiment damages the structure mid-run and measures self-repair.

Hypotheses tested

• H7 (primary) — Trace→Actor Crossing. Three operational criteria: (1) persistence despite erosion (structure_stability ≥ 0.90), (2) non-reducible dynamics (pheromone stays elevated ≥0.5 over structure while the structure's mass has saturated, |material_growth_rate| < 0.01 — the field stays energized without continued net accumulation), (3) constraint on agents (≥60% of deposits land on existing structure). A crossing is declared when all three hold for ≥4 consecutive samples.

  Criterion 2 was corrected on 2026-07-27. It previously required the deposit rate to fall below its early-run average — unsatisfiable in this model by construction, since Grassé positive feedback makes deposits accelerate as structure forms. That clause held only during warm-up (samples 0–5), before any structure existed, so the detector could never fire and the original null was a property of the detector rather than evidence about H7. The results below are from the corrected detector, which selftest Part 5 verifies both fires on an ideal crossing history and withholds on each single-criterion negative. See ../REVIEW.md §1.

• H1 / H10 (context) — multi-scale composition. sims 03–05 showed composition fails without an explicit mechanism; sim06 tests whether stigmergic self-maintenance is that mechanism.

• H4 (context) — dynamic environment. The field has its own dynamics (decay, diffusion, erosion, self-emission), not just a passive recording medium.

Design

• Grid: 100×100 toroidal. 200 termites, 4000 steps, sampled every 25 steps.
• Termites (vectorized): Moore-neighborhood random walk; loaded termites follow the pheromone gradient with probability PHERO_FOLLOW. Loading via off-grid reload (RELOAD_PROB) + rare cell pickup (turnover). Deposit probability p = DEPOSIT_BASE + DEPOSIT_GAIN · local/(1+local) (Grassé positive feedback, saturates ~0.95). Tuned: deposit_base=0.02 (low nucleation), phero_follow=0.6.
• Field: cement pheromone decays (PHEROMONE_DECAY=0.02), diffuses (3×3 toroidal blur), material erodes (material_decay=0.01, tuned). Self-maintenance condition adds pheromone += MAINTAIN_GAIN · material over structure cells.
• Metrics per sample: total_material, n_structure_cells, mean/max pheromone, n_pillars (8-connectivity components), compactness (cells/bbox), stability (fraction of structure surviving window-to-window), deposit_on_structure_fraction.
• Crossing detector: post-pass over history; requires 4 consecutive samples satisfying all three H7 criteria.
• Perturbation: at 60% of steps, zero a central 25%-area square of material + pheromone; measure recovery = current / pre-damage total_material.

Results (from results.json — tuned params: material_decay=0.01, deposit_base=0.02, maintain_gain=0.3)

Honest result: a partial/null finding. Self-maintenance produces a real quantitative difference — it builds 66% more structure (1876 vs 1131 cells) with higher retention (0.98 vs 0.96). But the crossing detector does not fire for either condition, and perturbation recovery is ≈1.0 for both (the damaged patch fully regrows).

Why no crossing? Per-criterion pass rates over the 160 samples:

Criterion 1 is the binding constraint. Structure stability runs 0.849–0.893 (baseline) and 0.746–0.802 (self-maintenance) against a 0.90 threshold — close, but never reached. Cells sitting near STRUCTURE_THRESHOLD flip above and below it every window, so the structure churns rather than persisting.

For self-maintenance, criterion 3 also fails outright (0/160), and in the direction opposite to the hypothesis: the self-maintenance condition is more fragmented than baseline (219–297 components vs 66–109) and less selective (deposits on structure 0.43–0.53 vs 0.70–0.79). The cause is saturation: maintain_gain=0.3 drives pheromone to ~15.6 over structure, but the deposit response p = base + gain·φ/(1+φ) is flat above φ≈1, so combined with diffusion the deposit probability goes uniform at ~0.87 across the whole grid. The self-emission loop destroys the spatial contrast stigmergy depends on instead of reinforcing it.

Root cause: the model's stigmergy is too spatially diffuse. Deposits spread across the grid faster than they consolidate, so the structure never coalesces into few large, stable pillars. The self-maintenance feedback amplifies building (66% more structure) but actively works against consolidation.

This is a legitimate result: stigmergic self-emission alone is insufficient to produce the trace→actor crossing in this minimal model, and a saturating deposit response makes matters worse. Spatial consolidation is the missing ingredient — pointing to what sim07 must add.

Key findings (project arc)

sims 03–05 showed multi-scale composition fails without an explicit mechanism (H10: unbounded space is necessary but not sufficient). sim05's closing recommendation was stigmergy as the "glue." sim06 tested the simplest stigmergic self-maintenance and found: the feedback loop amplifies building but does not produce a self-maintaining actor per H7's operational criteria. The mechanism is present in weak form (structure-size boost) but the crossing — the moment a trace becomes an autonomous actor — is not achieved. H7 is not supported by this model as implemented; the result is suggestive but inconclusive, and points to the need for a consolidation mechanism (deposit inhibition, directional bias, or explicit pillar-merging rules) in a follow-up sim.

Note on the strength of this null (2026-07-27): between Session 9 and this revision the crossing detector was found to be incapable of firing at all, so the original null carried no evidential weight. With criterion 2 corrected the detector demonstrably fires on an ideal history and the null is now a real measurement — criterion 1 misses by roughly 0.01–0.05 of stability. That is a near miss, not a categorical failure, and it should not be read as strong evidence against H7.

Limitations

• Toy 2D model; no real termite biology, airflow, or thermoregulation.
• Single mechanism (cement pheromone self-emission); no competition between structures.
• The crossing criteria thresholds (0.90 / 0.5 / 0.01 / 0.6) are reasonable but arbitrary; a different operationalization might classify the same dynamics differently. Criterion 1 misses by ≤0.05, so it is threshold-sensitive — the thresholds were deliberately not retuned after the criterion-2 fix, to avoid selecting a detector that produces the desired answer.
• structure_stability is not invariant to sample_every: it measures survival of structure cells across one sampling window, so at material_decay=0.01 a 25-step window erodes ~22% of mass and cells near STRUCTURE_THRESHOLD drop out. A shorter sampling interval would raise the same structure's measured stability. Cross-run comparisons are only valid at equal sample_every.
• The pre-2026-07-27 parameter search (material_decay 0.005–0.4, deposit_base 0.005–0.05, phero_follow 0.6–0.95, maintain_gain 0.1–0.5, reload_prob 0.15–0.3) reported no regime where the crossing fires, but it ran against the broken detector and so establishes nothing. It has not been repeated.
• Perturbation patch is a clean square; real damage is irregular. Recovery ≈1.0 means the damage is invisible at these params — a stronger test would need higher erosion or larger damage.
• 2D toroidal grid removes boundary effects that real mounds use (edges, substrate).

What it teaches / next steps

The weak result is itself informative: stigmergic self-emission is not the missing composition mechanism by itself. The structure amplifies but doesn't consolidate. A sim07 could test whether adding spatial consolidation — e.g. deposit inhibition (termites avoid cells above a density cap), directional building bias along existing walls, or explicit pillar-merging — lets the trace→actor crossing occur. If sim07 produces the crossing, the project can then ask sim05's open question with the stigmergic glue in place: do two self-maintaining structures compose into a higher-order (L2) organization?

How to run

cd ~/brain/artificial-life/simulations
uv run python3 sim06_termite_mound/sim06.py run            # full experiment -> results.json
uv run python3 sim06_termite_mound/sim06.py sweep_plot     # parameter sweeps -> output/*.png
uv run python3 sim06_termite_mound/sim06.py sweep_crossing # broad crossing search -> output/sweep_crossing_results.json
uv run python3 sim06_termite_mound/sim06.py selftest       # internal sanity checks (incl. Part 5d deposit_response guard)
uv run python3 sim06_termite_mound/cue_response_sweep.py  # Session 22: cue-based non-saturating control -> output/cue_response_sweep.json

Session 22 — the deposit_response parameter (cue-based non-saturating control)

sim06's as-built deposit rule is the saturating cue p = base + gain·φ/(1+φ) (flat above φ≈1). A deposit_response parameter ("saturating" | "linear") switches the cue-response curve. The "linear" form is the non-saturating cue p = base + gain·φ (clamped to 1.0) — the remaining cell of the 2×2 (queued-topic #67), completing the cue × {linear, saturating} factorial that mirrors sim09's action × {linear, saturating} factorial.

Result — the non-saturating cue crosses LESS, not more (opposite of the action family). Seed-42 factorial (64 conditions): saturating cue 32/32 stable (hold 1.000); linear cue 16/32 stable (hold 0.527). Without self-maintenance: saturating 16/16 stable; linear 0/16 stable. Self-maintenance rescues the linear cue (16/16 stable). The linear cue clamps to p=1.0 at φ≈1.15, flattening the gradient (mean pheromone 0.467 < 0.5 threshold vs saturating's 0.749). The non-saturating property reverses sign across families: it helps in the action family, hurts in the cue family. See output/cue_response_sweep.json, hypotheses/logs/H7.md and hypotheses/logs/H11.md (Session 22 refinements).

Visualization: cd sim06_termite_mound && python3 -m http.server 8080 → open http://localhost:8080/visualize.html. Also deployed at alife.vancedubberly.com/sim06_termite_mound/visualize.html.