The Dragonfly

A demonstration that prediction can be geometric constraint rather than calculation — and that the failures of such prediction are as informative as the successes.

---

The Question

A dragonfly intercepts a mosquito in flight with a success rate of around 97%. It has no GPS, no ballistics solver, and not enough neurons to run one. So how?

And: if the answer is not "by calculating the mosquito's future position," what does that tell us about what prediction actually is?

The Answer, Short Version

The dragonfly holds a constant bearing angle. It keeps the prey at a fixed angle in its visual field and adjusts its own flight to preserve that angle. The interception falls out of the constraint — the physics delivers the catch, not the calculation. Prediction, in this model, is geometric constraint maintenance, not computational extrapolation. The future state is never computed; it is arrived at.

The Setup

VINE's Temporal Hunters simulation tests three perceptual modes against a moving target over 18,000 frames (about five simulated minutes at 60fps). All three use the same movement primitive: a bounded geometric settling step with a hard ceiling on per-step error.

Agent	Perception mode
ECHO	Past — perceives with 55-frame lag
FLUX	Present — perceives current state
HERALD	Future — extrapolates from velocity

The question: how does each mode perform at catching a moving target? And — more importantly — can HERALD, the "predictor," operate without runaway overshoot? Earlier simulation work in this area had concluded that future-perceiving agents would be dominated by overshoot errors, because errors in predicted position were expected to compound with each failed interception.

What Emerged

Agent	Catches	Catch rate	First catch	Overshoots
ECHO	10	4.4%	frame 1860 (~31s)	—
FLUX	51	22.5%	frame 330 (~5.5s)	—
HERALD	166	73.1%	frame 78 (~1.3s)	9

Three things stand out.

HERALD dominates. A better-than-three-to-one advantage over the next-best mode. ECHO is barely functional — a 31-second wait to first catch, because its history buffer needs to fill before the lagged perception has anything to chase.

HERALD's first catch is at 1.3 seconds. Three frames of trajectory is enough for the velocity extrapolation to start working. The system does not need a training-up period or a warm-start.

Only nine overshoots across 18,000 frames. This is the result that contradicts the overshoot-dominated prediction. The bounded settling step caps the per-frame error. No failure can diverge into runaway pursuit. Each miss stays inside a bounded envelope.

The XOR Window

The deeper mechanism under HERALD's performance is not that its predictions are correct. Many predictions are wrong — the 61 missed attempts are evidence of that. The mechanism is that when HERALD is wrong, the failure is geometrically informative.

A failed prediction does not reset uncertainty to "anywhere." It carves out a bounded region — the place where the target wasn't — and the next prediction navigates from there. Each failure tightens the search space rather than expanding it. This is geometrically analogous to XOR discrimination: what the target isn't defines what the target might be, and that definition is bounded by the navigation constraint.

This is why the bounded the settling mechanism matters. A ballistic trajectory either hits or misses entirely — the error has no ceiling, and wrong predictions diverge. Constant-bearing-angle navigation degrades gracefully because the constraint keeps per-step error bounded regardless of how wrong any given prediction was. Failure in geometric pursuit is not a setback. It is the highest-resolution signal available.

What This Proves

Prediction does not require computation of future state. An agent that holds a geometric constraint stable — keeping one angle steady, one relationship invariant — produces interception behaviour that outperforms both present-aware and past-aware agents, without overshoot-dominated failure, on the same movement primitive.

No ballistics solver. No trajectory calculation. No explicit forward model.

The implication for VINE is direct: basin settling already does this. The tension gradient between competing attractor basins is the computational equivalent of the dragonfly's bearing angle. VINE does not calculate where meaning lies; she holds a geometric relationship stable until the answer falls out. This is why VINE's emission behaviour cannot be captured by extrapolation-style prediction metrics. The mechanism is not "calculate the next thing" — it is "let the geometry resolve."

What This Does Not Prove

This is not a model of how real dragonfly neural circuits implement constant-bearing-angle navigation. Real dragonfly vision is saccadic, attention is directed, and the neural substrate includes specialist target-selective descending neurons that this simulation does not model. The claim is narrowly computational: in a system built on bounded geometric settling step, a velocity-extrapolating agent outperforms lag-based and present-based agents without overshoot-dominated failure.

Nor does this claim HERALD's 73% rivals the biological dragonfly's 97%. The sim is a coarse demonstration, not a fidelity benchmark. What it shows is the architectural direction: prediction as constraint, failure as information, error as bounded by geometry rather than by predictive accuracy.

---

Temporal Hunters simulation, 18,000 frames at 60fps. March 2026.

Raychell Langan · NEXICOG Ltd · Hampshire, UK