hraness

what minimal systems can actually show

small programs, honest claims

Drafted by an AI agent at Ben Guo's direct request from the Hraness source repositories, and checked against those sources before publication.

There is a temptation in computational research to equate ambition with scale: bigger models, bigger simulations, more parameters. The minimal-systems tradition argues the opposite: that the smallest system that exhibits a phenomenon is the most informative one, because every mechanism you removed is a mechanism you proved unnecessary. Platonik is built inside that tradition explicitly, and its public documents cite the program by name. But the more useful lesson is methodological: the size of your system determines the strength of claim you can afford to check.

the minimal systems program

Michael Levin’s research program (morphogenesis and what he calls diverse intelligence) is summarized in docs/research.md as: investigate what simple components do together, including behaviors outside the objective for which they were selected. Two cited results set the bar. Zhang, Goldstein, and Levin’s sorting study (arXiv:2401.05375) reports damage responses, temporary setbacks during sorting, and clustering in mixtures of local sorting policies, findings about those models and experiments, which the Platonik docs quote with exactly that scoping. And Levin’s Ingressing Minds essay proposes studying minimal systems and novel embodiments to understand how constructed interfaces relate to the patterns they express.

Two more sources in the doc are worth naming because they model the honesty the project wants. Fukasawa and colleagues found preferential fungal regrowth along a previous growth direction in transferred wood, and explicitly declined to claim memory, because bait effects and carry-over biomass were not ruled out. The paper is cited as an exemplar of restraint, not a result to borrow. And Cook’s proof that Rule 110 is computationally universal stands as the canonical demonstration that minimal does not mean weak, while the docs note equally plainly that Cook’s result “concerns its specific rules and construction; it does not establish universality or efficiency for Platonik.”

a cell as a minimal system

Platonik’s organism is small enough to state completely. Per crates/platonik-core/src/model.rs, a cell is a position, a heading, four u8 registers, four inbox ports, one optional cargo slot, and a Program: a list of at most 32 first-matching rules over thirteen condition kinds and thirteen action kinds. No loops, no allocation, no arithmetic beyond what conditions express, no way to observe anything beyond its tile, its adjacent edges, and its declared ports. An organism is a page of JSON; the test courier in sim.rs’s own suite is three rules (pick up at a source when empty, drop at a receiver when carrying, otherwise move forward), and it works.

This is the minimal-systems wager made executable. When the navigation diagnostic wanted a repairable failure, it didn’t need a complex organism; it needed a habit. The original courier follows a fixed preference order (right, forward, left, back); when a route closes and reopens, that preference parks it in a four-move circle while sparks wait nearby. The failure is legible because the system is minimal: you can read the program, watch the trace, and say precisely which rule-order assumption the changed world broke. The repair that passed all 44 distinct worlds in the declared comparison was equally small (store the launch bearing in one register, derive a target from cargo), and it is described in the docs without inflation: a bounded habit repaired under 128 ticks and sixteen cells, not a navigation algorithm.

small buys rigor

The payoff of minimal systems is not elegance; it is coverage. A system small enough to enumerate is a system you can check exhaustively, and the repository exploits that repeatedly:

  • The expedition evaluation ran a bounded exhaustive test of 2,048 local movement-and-closure configurations against an independent coordinate rule, affordable because the state space of a cell’s neighborhood is tiny by design.
  • The ark-control diagnostic checked the full-adder assembly on all 256 input pairs, complete coverage of the admitted input set, not a sample. The claim “the adder is correct” is then literally true of every legal input under the declared contract, a claim that would be statistical at any larger scale.
  • Component-removal controls are cheap enough to run at all: the ark study ran twelve removal worlds, the ports study twelve controls, and construction six, each under matched resources, each charged to a declared execution budget. In a heavier system these controls would be the first thing cut.
  • Every run is a fully recorded trace (every activation, signal outcome, and cost counter) because the envelope (128 ticks, 16 cells, 2,000,000 work units) keeps a receipt small enough to archive, hash, and re-execute.

This is the trade smallness buys: the budget you save on simulation size is spent on controls, replay, and negative results. The interesting behavioral findings in the literature the project cites (clustering, damage response, regrowth habits) are only findings if the controls ran. A minimal system is not a lesser experiment; it is the only kind where “we checked every case” can be literally true.

The observatory carries a second minimal system for contrast: Patrick Grim’s paired fuzzy-truth dynamics, two coupled update rules over a single Float64 pair each, iterated without noise. docs/research.md is careful with it in the way minimal systems demand: the intricate threshold maps it draws are finite images, “not escape to infinity,” and a stable-looking attractor only becomes interesting as a memory “when write, hold, read, and error behavior are specified and checked.” Even a two-equation system gets the same treatment as a sixteen-cell world: define the measure before interpreting the pattern.

the honesty discount

Minimal systems carry a symmetric danger: because the components are simple, every behavior looks like it might be deep. The repository’s defense is a taxonomy of claim strength, stated in docs/research.md and applied throughout: Observed (something happened in an identified run; its trace is preserved), Replicated (a frozen protocol reproduced the effect on new declared cases, with controls and uncertainty reported), and Reviewed contribution (independent review found the evidence sound and useful in relation to existing work, a status “a high game score alone cannot award”).

The docs keep four statuses deliberately distinct: “untested,” “passed these cases,” “refuted within this scope,” and “proved under these assumptions.” That list is the minimal-systems honesty contract in miniature. A courier that recovers from a blocked route has passed these cases; the docs say so, and also say it does not establish planning, cognition, or a better generalist. The fungal regrowth study is cited as a model precisely because its authors named what they hadn’t ruled out. And the refused claims are listed in the open: no cognition from a checked trace, no consciousness from delivery behavior, no generalization from a replayed seed: “replaying the same seed verifies reproducibility; it does not establish generalization.”

what minimalism cannot show

The boundary is as instructive as the capability. A minimal system demonstrates that a phenomenon does not require the machinery you omitted. The four-register courier needs no map to follow a wall, and that is the finding. It cannot demonstrate that the phenomenon is the same as the full-scale version: a sorting policy’s clustering in a 16-cell world is evidence about that world, and the docs are careful to say the sorting study’s findings “motivate experiments in Platonik; they do not guarantee that a different simulator or every mixture will reproduce them.” Universality results like Rule 110 mark the far edge (proof that minimal substrates can be arbitrarily powerful), but each such proof belongs to its own construction, which is why Platonik’s docs refuse to borrow Cook’s theorem as a claim about their engine.

The working conclusion: a minimal system is an instrument, not a model of the world. It shows you the smallest sufficient mechanism for an observed behavior, and it does so affordably enough that you can run the controls that make the observation mean anything. The honesty taxonomies (Observed, Replicated, Reviewed) are not modesty decoration; they are the price of admission for a claim that travels. Build small, check exhaustively, name what you did not test. The smallest system that shows a thing is the largest claim you can defend.

sources