The Bats Had Momentum: What Fundamental’s Field Is Missing

In Batman Returns (1992), the bats that pour through Gotham and the regiment of penguins that march on the city were not hand-animated. They were boids — a modified version of Craig Reynolds’ 1987 distributed behavioral model, one of its first big-screen outings. Each bat saw only its neighbors and followed a few small rules, and the swarm fell out of the interaction. It was one of the earliest times a film audience watched emergence do an animator’s job, and it worked because the creatures didn’t just arrange themselves — they moved like things with weight. They carried momentum. A bat that banked hard kept going; the flock had inertia you could feel.

Fundamental is a descendant of that 1992 swarm. It’s a relational field runtime — it lets the elements of a web page become bodies in an invisible physics field — and, as it happens, it flocks. Its forces for spacing, aligning, and grouping particles are even named, in the source, after Reynolds’ three rules. Drop an autonomous agent into the field and a murmuration emerges with no code that says “flock.”

But watch it closely and something the bats had is missing. Fundamental flocks like a boid, but it doesn’t move like one. And the reason is a single word: momentum.

Where the momentum leaks

Two facts about how the engine works today.

First, bodies don’t recoil. When a body in the field pushes a particle — attracts it, repels it, swirls it — the particle moves, but the body doesn’t move back. The body is an immovable source. In the real world, and in Reynolds’ bats, every push has an equal and opposite push; that’s how a flock’s total momentum stays put while individuals trade it back and forth. In Fundamental, the push goes one way and the bookkeeping never balances.

Second, mass is nominal. By default every particle behaves as if it weighs exactly one unit — the engine advances velocity by simply adding the force (v += F), not by dividing force by mass (a = F/m). So “heavier elements swing wider, lighter ones dart” is, right now, a thing the field can’t express by default. There’s an opt-in for real mass, but the recipes are all tuned around unit weight, so almost nothing uses it.

Add a dab of friction every frame — which the engine does deliberately, to keep interfaces calm — and the result is a field that is driven and damped rather than one that conserves. Momentum doesn’t circulate through it the way it circulates through a flock of bats. It leaks.

This isn’t a bug. It’s a design that optimized for a calm, legible interface and got one. But it leaves a real capability on the table, and — interestingly — the capability is the one most aligned with Fundamental’s own headline idea.

The idea: give bodies reaction, give matter mass

The proposal is two linked changes, both opt-in.

1. First-class mass. Make a = F/m available as a real default for a field, with a body’s mass deriving from something physical — its rendered area, say. A heading becomes heavy; a tag becomes light. The machinery already exists in the engine; what’s missing is making it the basis of a field’s motion rather than a rarely-used flag.

2. Newtonian reaction. When a force acts on matter, apply the equal-and-opposite impulse back to the source body. Not a new force — a property that existing forces opt into: when attract pulls a particle in, the body feels a small tug toward the particle in return. Bodies stop being immovable sources and become participants that can be moved by what they move.

You can’t have the second without the first — you can’t recoil an infinite-mass body, so reaction needs bodies to have finite, real mass. The two changes are one change. And the engine team has already half-seen the hole: the wall primitive “does not yet recoil,” and collision recoil onto the DOM is already filed as proposed. This would generalize that instinct across the whole force set.

What we’d expect from it

This is a proposal, not a shipped feature, so the right register is expected results — hypotheses, not findings. Here is what adding momentum should do.

Reciprocity becomes physical, not just bookkeeping. Fundamental’s whole thesis is reciprocal: bodies bend the field, and the field bends them back. But today that “back” is a feedback loop — the field writes a CSS variable onto the element and a style rule reacts. With reaction, the loop also closes through motion: a body that throws its weight around gets pushed by the matter it disturbs. The headline becomes literally true at the level of physics, not only of styling. This is the single most thesis-aligned upgrade available.

Weight becomes legible. A massive element should settle slowly and overshoot; a light one should snap. Importance could read as inertia, not just size — a heavier body resists being shoved aside by a crowd of lighter ones. That’s a new expressive channel that costs no new vocabulary.

Momentum can be conserved on purpose. With reaction in place and friction turned off, a field becomes a closed system: total momentum is preserved, the way it is in a real flock. That’s an opt-in “physics-truth” mode for the cases that want it (a simulation, a teaching demo) sitting alongside the calm, damped default for the cases that don’t.

The clump might ease — but we should test, not assume. Today, opposing forces can cancel and friction freezes the leftover into a stalled clump. Real mass changes those dynamics: matter carries through a balance point instead of stopping dead at it. We’d expect momentum to soften the clump, but the honest move is to measure it, because it could equally introduce new oscillation. Reaction adds energy paths that didn’t exist before.

And there are costs worth naming. Momentum makes things bouncier, and bounciness is exactly what the calm-interface default was protecting against — so this has to be opt-in, not a global flip. Every recipe is tuned for unit mass, so a mass-on field needs recalibration. And the moment bodies can be moved by matter, layout stability becomes something the system has to actively defend rather than get for free. None of these is disqualifying; all of them are real.

Why this one

Of everything on Fundamental’s list of honest limitations, momentum is the one a single, clean mechanism would resolve — and the one that pays the system’s own story back the most. Most of the others are deliberate (energy decay keeps things calm; designed forces are kept distinct from natural laws on purpose) or aren’t the engine’s job at all (whether any of this helps a user still needs studies that haven’t been run). Momentum is different. It’s a capability the field almost has, that the recipes don’t yet use, that the roadmap already gestures at, and that would make the reciprocity at the heart of the whole project something you could feel in the motion — the way you could feel it in a swarm of bats over Gotham.

The first domino

Here’s the part that makes momentum more than a feature. You can’t add it cleanly without touching the floor of the engine — and that turns out to be the point.

To give bodies recoil, the integrator has to stop being a quick v += F with a hard speed cap and become something that actually conserves (a symplectic, fixed-timestep solver — the kind of thing d3-force has used for graph layout for years). To give matter mass, bodies and particles stop being two different kinds of thing — an immovable source and a weightless dot — and become one kind of matter that can both push and be pushed. And for any of it to be measurable, forces have to produce forces instead of secretly editing velocity, so the integrator can do the physics in one place.

None of that is momentum, exactly. It’s the foundation momentum needs. And it’s the same foundation a serious substrate needs anyway — which is why the right way to read this isn’t “add bounciness,” it’s momentum is the first domino. Knock it over deliberately and the engine stops being a calm decoration that reads the page and starts being a small physics that the page is one view of.

What that unlocks is already written down. Fundamental’s possibility space has long described matter that isn’t particles — fluid that pools, fabric that sags under a heavy heading, sand that piles where attention accretes. It describes the field as something an AI agent could query instead of scraping the DOM, and as a shared space where a room full of readers’ attention sums into a visible weather. Almost all of it is waiting on the same thing: a field whose physics is real enough, and independent enough of the browser’s layout loop, to carry that weight. Momentum is where you start paying that down.

The bats had it in 1992. It’s time the field did too.

Update — the first domino is falling (2026-07-01)

Written above as a proposal; since then, most of the foundation it argues for has actually shipped in the 0.9.x substrate line. The three prerequisites this note names —

• “forces have to produce forces instead of secretly editing velocity, so the integrator can do the physics in one place” → shipped as the dimension-aware impulse accumulator: every force’s contribution now flows through one canonical capture path (applyAndRecord) and is attributable per force, across five channels (linear, thermal, angular, temporal, semantic).
• “a symplectic, fixed-timestep solver” → shipped as an opt-in fixed-timestep integrator (createField({ integrator: 'fixed' })) — the frame-rate-correct path, alongside the calm v += F default (partial progress on #658; #659 velocity-Verlet is still open).
• bodies that can be “pushed by the matter it disturbs” → shipped as body-authority modes: a data-authority="dynamic" body is engine-owned and recoils under the net field (there’s a live wiring of it, plus a reading of it as agent-consumable JSON, on the /demo page).

So the “first domino” is no longer a proposal — it’s largely knocked over. What remains genuinely unshipped is momentum proper: the specific Newtonian own-emission reaction (a body recoiling from what it emits, not just from other bodies’ fields) and first-class mass by rendered area. A dynamic body already moves under the field; making it feel the equal-and-opposite of its own pushes, and giving it real inertia, is the next push — and it’s the one this note was really about. The foundation is built; the reciprocity-in-motion payoff is what’s left to claim.

Measured — the clump hypothesis, tested (and it lost)

Since the update above, momentum proper shipped too, both opt-in and experimental: first-class body inertial mass ∝ rendered area (mass: true, so a heading settles slowly and a tag snaps — measured: a light body recoils >1.5× a heavy one) and own-emission reaction (reaction: true, so a directional emitter recoils like a rocket). The whole proposal is now buildable.

Which let us run the one experiment this note leaned on hardest — “the clump might ease.” It was worth being honest that it needed testing, because the test says no. A central attractor with friction freezes matter into a stalled clump; turning first-class mass on made it, if anything, slightly tighter (−8% spread), less mobile (−22% residual motion per frame), and markedly more oscillatory (+77%) — not looser. The guess that “matter carries through a balance point” didn’t hold for the swarm; the caution that momentum “could equally introduce new oscillation” did. (Caveat: one scenario, particle-mass only — own-emission reaction is body-level and orthogonal to the particle clump.)

So the honest ledger: momentum’s body-level story is real and measured (weight reads as inertia; reciprocity closes through motion). Its clump-easing story — the one soft claim in this note — is not supported by measurement, and shouldn’t be sold. That distinction is the whole point of insisting on the measurement rather than the vibe.

And there’s a twist worth stating plainly: the clump is solvable — just not with momentum. The same frozen clump, with the separation force turned on (separation: 1, short-range particle repulsion the engine ships specifically “to break up clumps in high-density areas”), loosens +100–157% and gets 49–73% less oscillatory — the ease this note reached for, from the tool actually built for it. The lesson isn’t “momentum was a bad idea”; it’s that a plausible mechanism deserves a measurement before it gets credit for a problem another mechanism already solves.

The full comparison this grew out of — Fundamental against Reynolds’ Boids, with the per-axis verdict and the explainability-vs-arbitration result — is in the research paper Substrate, Not Spectacle: Behavioral Models After Boids. The original argument below the update was written 2026-06-29 as a proposal; momentum proper (own-emission reaction + first-class mass) is still unshipped, and everything framed as “expected results” remains a hypothesis, not a measurement.