intuitive-block-design

Work through eight ordered questions: define concrete input-output episodes, cost the dumbest baseline solution, find a domain that solves the problem by construction, name quantities in plain words before equations, separate forced moves from free parameters, distinguish learning and inference clocks, state capacity and failure modes, and pre-register a readout probe.

Use it before writing a block contract, adding a registered candidate, opening an invention block, or when a block underperforms and you are tempted to add machinery. It is designed to run before implementation, not after.

Every candidate must be exactly one delta from an existing candidate, and each delta ships with the capability it buys and the cost it adds. Two deltas at once makes regressions untraceable, so changes ship one at a time against a runnable baseline.

The method warns against borrowed vocabulary without imported dynamics, formulas chosen because they are standard, and validation by loss curve alone. A block whose failure mode or capacity you cannot state is not yet understood, and hard-coding hoped-for emergent representations is flagged as a smell.

The references file runs all eight questions across four derivations: Hopfield networks, Boltzmann machines, the Tolman-Eichenbaum Machine, and attention. Reading one column shows a full derivation's shape; reading one row shows one question answered in four domains.