Playable theory · a companion to Self-Improving Agents (Manning)
The creatures never get smarter — their policy cards do. Each forager runs a fixed SENSE→PLAN→ACT→LEARN loop; its behaviour comes entirely from a readable policy card. Around the population, the improvement loop — ACT→SIGNAL→SEARCH→APPLY — evolves those cards. You are the harness engineer: set the dials, gate what ships. The loop, the dials, and every term on this page are from Self-Improving Agents by Micheal Lanham (Manning).
—
—
—
—
—
Inside each creature: sense → plan → act → learn, every tick. The code that runs that loop — the creature's "model" — never changes. Outside, once per generation: act → signal → search → apply (ch. 1). Everything the loop improves is an artifact — here, only the policy card. Within a life a creature also keeps working memory (the hazard and cluster chips in the inspector); that is the raw material the book's L2 layer would persist and improve across tasks, which this demo deliberately does not do.
Signal — the judge is available for every run but noisy, and biased toward busy-looking lives (watch the judged line float above the true line). Ground truth is exact but costly: alone, it halves how many candidates you can afford per generation. The default reranks — judge everything, ground-truth the shortlist — a budget trick of this demo; the book's composite signal instead pairs a score with a reflection critique (ch. 3).
Search — hill-climb is the ratchet (ch. 3): cheap, single-lineage, parks on the first peak.
The population (GEPA-style, ch. 4) breeds from its top four with crossover, and reflection reads each life before
mutating. The archive (DGM-style, ch. 4) never discards: quality-diversity sampling taxes every reuse so strong
neglected branches keep earning turns — and because nothing is thrown away, the first time a branch is bred
from, its judge-flattered score is grounded to a real measurement. The book's evidence gate is the
accept rule itself: a card ships only on a better measured score, which is exactly this demo's
promote() bar.
Artifact — fixed here: the whole policy card, which bundles the rule text (the L1 analogue) with three body traits (speed and sense ranges) folded into one artifact to keep the demo simple. The same loop, pointed at memory, scaffolds, or code, is the rest of the book.
promote() is the offline deploy gate: you review the winning card's diff before it
ships. Switch to auto-apply and you are running the online loop — watch what a noisy judge does to it.
The Ratchet: an open meadow — hill-climbing works. The Valley: the ratchet parks on the near patch; crossing needs several genes to move together, which is what a population's crossover can do. The Mirage: a dense field that runs dry mid-life traps converged populations; the archive's payoff is recovery — reaching back to a low, structurally different branch and growing past the plateau. And sometimes the cheaper search wins: spend by the landscape.
Set the signal to judge-only and let it run: the judged line floats far above the truth (roughly 130 judged for 52 real on the meadow) because the judge rewards busy-looking lives. The loop optimizes what the signal sees, not what is good — chapter 1's ceiling, and the reason the signal dial is the most consequential one.
New to the ideas? The Proof Gate is the five-minute story version — proof, measurement, and the Darwin Gödel Machine — and a good warm-up for this sandbox.
A companion toy for Self-Improving Agents (Manning, in progress). Helix is the book's name for the L0 toolkit — the harness you are operating here; the garden is just what it is pointed at. Deterministic per trial — same dials, same story. No model calls: the "judge" is simulated noise with a bias, which is the point.