discovery

Type: kb/types/tag-readme.md

A learning operation distinct from constraining and from source-derived reshaping: positing a new general concept and simultaneously recognizing existing particulars as instances of it. Discovery produces theories — the highest-explanatory-reach items accumulation can store. A child of learning-theory.

The operation

• recognition, not linking, is the hard problem in knowledge systems — recognizing shared structure is the expensive step, articulating a seen link is cheap; naming a recognized structure amortizes later recognition
• discovery lifecycle — definition: the staged path from conjecture to accepted discovery (observe → conjecture → consequences → test → accept → integrate); ampliative traffic enters at the conjecture stage, and the co-arising insight is the degenerate case with the phases telescoped
• candidacy evidence licenses escalation to assessment, not acceptance — conjecture: cheap evidence can route a hypothesis to expensive assessment without gaining verdict authority; Peirce's economy of research, witnessed by pricing-versus-adequacy and source-grounding cases
• automated synthesis is missing good oracles — why discovery resists automation: no cheap verifier for whether a posited generalization is good
• known-target discovery benchmarks show reachability, not discovery — benchmark critique: recovering a planted generalization measures search, not the open-ended act

Reach — what discovery produces

• first-principles reasoning selects for explanatory-reach — Deutsch's adaptive-vs-explanatory distinction: explanatory knowledge transfers because it captures why, not just what works
• Theory warrant should be tracked at the finest granularity evidence licenses — separates structural candidacy from earned warrant: support attaches to the finest claim, bundle, and scope the evidence identifies, and joint warrant distributes only through entailment or attribution
• warranted reader update is the objective of substantive writing — the contribution criterion: substantive writing selects a nontrivial audience-relative update and earns it with evidence and reasoning
• brainstorming: how explanatory-reach informs KB design — working notes applying the explanatory-reach concept to KB design decisions
• theory refinement may improve sample efficiency under structured shifts — conjecture: the measurable payoff of a discovered theory is fewer target observations after a shift that preserves the structure it names, conditional on reach-assessment having earned it
• revision guided by rationale needs faithfulness, not just legibility — when repair uses a recorded rationale, misleading dependencies can direct it to the wrong premise; test that guide by intervention
• Derivation and inheritance give starting warrant; discriminating evidence or proof earns scope — provenance grades a posited carve's starting warrant and relocates its untested part to a statable place; an underdetermined choice gets none, and scope is earned only over the domain evidence or proof actually covers

Conditions for discovery

• ad hoc explanation can be rational when error is cheap and local — prices the lifecycle's heavy phases: derive/test/integrate controls errors that can propagate, while disposable guesses can select cheap local probes before the reach toll is paid at promotion
• short composable notes maximize combinatorial discovery — the artifact-shape argument: small claims compose into more candidate generalizations
• information value is observer-relative — the gap discovery (and consumer-directed reshaping) bridges: structure exists but is inaccessible to the bounded observer until transformed
• Epiplexity by example — worked examples: encrypted messages, shuffled textbooks, CSPRNGs, and chess notation make observer-relative extractable structure concrete
• reverse-compression is when LLM output expands without adding information — the failure inverse: expansion that adds no extractable structure, where productive transformation makes structure accessible
• minimum viable vocabulary — naming as the discovery lever: the vocabulary that most reduces extraction cost for an observer entering a domain