Pattern organs · #5 core · sovereign

A pattern survives only if it generalises

Give the forge a domain's cases. It proposes simple patterns from a training split, then keeps only the ones that still predict a held-out split — graded by balanced accuracy, never an LLM judge. Overfit patterns die on data they never saw. Re-derived live in your browser from the gated kernel.

Domain A — a real signal buried in noise

Feature 0 truly drives the label; features 1 & 2 are noise. The forge grades every proposed pattern on held-out data. keep = beat the bar; drop = failed to generalise.

pattern	train BA	held-out BA	verdict

Domain B — no real pattern (an overfit trap)

A feature that matches the label early (training) then flips late (held-out). It looks great on train and dies on held-out — exactly the trap pattern-finding usually falls into.

pattern	train BA	held-out BA	verdict

Improves itself — breeding compound patterns

Domain C: the signal is a conjunction (x0 high AND x1 high) that no single feature catches. Breeding ANDs patterns across generations — selecting on train, judged on held-out — and finds the true two-term rule.

The local model proposes — the gate keeps only what generalises

Domain D: y = (x0 > x1) — a feature comparison no single threshold can catch. A sovereign local model (qwen2.5:7b, on your own metal) proposed these rules from 12 sample cases; the held-out gate graded each. keep survives; drop fails to generalise.

proposed rule	train BA	held-out BA	verdict

The observer loop — it learns which forms generalise

The observer watches the run above, learns that comparison rules generalise where stumps don't, and reprioritises the proposals. The generalising rule then surfaces earlier — fewer held-out checks before a winner.

The pattern book

The deliverable: a catalog of every pattern that survived across the domains, in plain English with its held-out evidence and where it came from. Compiled live from the kernels — not typed. (Shippable as book.md.)

domain	pattern	held-out BA	found by

Mint it — own the model

The validated classifier is a tiny, owned model. Minting it is gated by fallforgemint's own rule — a node mints only on a certified BEATS receipt — and its economics are fallforgemint's own own-vs-rent calculator.

The sealed claims

Pre-registered in predictions.json before measurement; checked live against the kernel now.

Honest scope

This is the core of Pattern organs #5: the held-out, rule-graded survival test that stops pattern-finding being theatre. The fuller vision — an LLM proposing richer candidate patterns, kard-evolve breeding the survivors, the observer loop rewriting them, and shipping a pattern book + a minted small model + a Deck card — are stages built on top of this, flagged, not yet claimed. What's proven here: patterns only survive if they generalise.