This is a seedbank before it is a doctrine.
The list came from a larger breeding pool: search algorithms, HCI laws, optimization heuristics, distributed-systems protocols, control theory, testing methods, information retrieval, compression, game theory, and the already-live design-breeding branch of this graph. Most famous algorithms stay trapped in their home domains. The useful question is which ones become obvious human moves once generation gets cheap and judgment gets scarce.
These ten are the first shelf: small enough to remember under pressure, rich enough to become future nodes.
Lineage: BUGSY, Bayesian optimization, bandits, optimal stopping.
Every extra pass spends time reality could have spent teaching you. Search quality includes delay. Keep exploring while the expected learning beats the cost of waiting; move when waiting has become more expensive than imperfection.
Seed: Search Time Is Utility.
Lineage: particle filters, Kalman filters, Thompson sampling, best-arm identification.
A builder often collapses too early into one story: this is the user, this is the product, this is the design, this is the market. Carry multiple live hypotheses with weights. Let each new signal resample the population instead of forcing one premature truth.
Seed: The Particle Filter Mind.
Lineage: Picbreeder, interactive evolutionary computation, genetic algorithms.
A winning artifact is an ancestor. It is too large to copy. The useful move is to name the living trait, preserve that gene, and reopen the rest of the space. A click becomes powerful when it becomes inheritance.
Seed: The Click Becomes a Gene.
Lineage: MAP-Elites, novelty search, design of experiments, quality-diversity search.
Taste learns from separable differences. Four palette variants teach less than one topology child, one timing child, one density child, and one route child. The batch should make the next reaction informative.
Seed: Orthogonality Buys Taste Bandwidth.
Lineage: simulated annealing, tabu search, GRASP.
Early search should be allowed to look worse. It needs permission to cross valleys, keep strange candidates, and avoid cycling back into familiar almost-good answers. Freeze late, after the landscape has had time to reveal its traps.
Seed: Permission to Be Worse Early.
Lineage: Grothendieck's rising sea, colimits, morphological analysis, abstraction-as-search.
When local search keeps producing competent wrong answers, change the abstraction level. Ask what larger object the pieces are trying to become together. The form problem may be a trust problem. The copy problem may be a promise problem. The onboarding problem may be a self-abstraction problem.
Seed: Abstraction Must Return.
Lineage: cybernetics, PID control, model predictive control, OODA.
Correction has more than one clock. There is present error, accumulated error, and the speed at which error is changing. A good loop does not merely react. It notices drift, remembers residue, and updates the next move before the miss becomes identity.
Seed: The PID Theory of Correction.
Lineage: GOMS/KLM, Fitts's law, Hick-Hyman law, information scent.
Interface friction is arithmetic before it is taste. How many primitive operations? How far is the target? How many choices slow the hand? What scent tells the user the next click is worth it? A surface that hides action cost hides the product's real shape.
Seed: Interface Friction Is Arithmetic.
Lineage: QuickCheck, fuzzing, delta debugging, metamorphic testing, premortems.
When the oracle is hard, generate failures. Mutate inputs. Stress assumptions. Shrink the break until the cause fits in the hand. A builder who cannot produce counterexamples is still negotiating with the happy path.
Seed: QuickCheck for Reality.
Lineage: DAGGER, feedback-as-process-signal, corrections-as-training-data, provenance trails.
Feedback compounds only when future work can reach it. A correction should become a rule, memory, test, branch, edge, example, or changed generator. Otherwise the lesson becomes a mood.
Seed: Lessons Are Machinery or Mood.
This shelf is deliberately incomplete. The next shelf is already visible: mechanism design for truthful systems, rate-distortion for lossy summaries, Shapley value for credit, Little's law for work in progress, max-flow/min-cut for bottlenecks, real options for waiting, Markov blankets for boundaries, active learning for question choice.
The test for an algorithm to build by is plain: after the human reacts, does the system know what to do differently next?