Recognition, not linking, is the hard problem in knowledge systems

Type: kb/types/note.md · Tags: learning-theory, discovery

Connecting two knowledge items takes two steps with very different costs. Recognition is seeing that the items share a structure at some level of abstraction. Articulation is stating the connection once it is seen — writing the link with its reason. Articulation is comparatively cheap: once someone sees that two proofs make the same move, saying so is straightforward. Recognition is where the cost concentrates, so it is where a knowledge system's effort and tooling should go.

Recognition cost grades with what must be recognized

What is shared Typical recognition cost
Shared feature — an observable similarity Cheap. Embedding search, keywords, and filenames surface candidates.
Shared structure — a reusable relation, pattern, or proof shape Expensive. Requires understanding what each item is about, then comparing them semantically.
Generative model — one process capable of producing the cases Most expensive. Requires proposing the dimension along which the commonality becomes visible, then testing whether it explains more than the cases that suggested it.

The grading is a heuristic, not a law. The forms are not strictly nested: a generative model can unite cases that look different on the surface. And the costs are observer-relative — they depend on the reader's prior vocabulary, tools, and access to the cases, since information value is observer-relative. Recognition also requires co-presence: a structure shared by items that never appear in the same context cannot be seen, which is why short composable notes maximize combinatorial discovery.

Naming amortizes recognition

The mathematical tradition supplies the main cost reducer: name the recognized structure. A mathematician who notices that two proofs make the same move can state that move as a lemma — a reusable intermediate result. The earlier proofs become applications of the lemma, and later proofs can recognize the move by matching against the name instead of re-deriving the similarity. One expensive act of recognition becomes many cheap lookups.

The amortization has conditions. The name must carry usable membership criteria: a name without a test for what counts as an instance does not make boundary judgments cheaper. And the structure must recur: naming a structure that never reappears adds vocabulary without saving recognition.

The mechanism-note heuristic

For a knowledge base this yields a heuristic. When several artifacts look similar, consider writing a note that names their shared mechanism rather than only linking them pairwise. Pairwise links record each recognition separately; a mechanism note converts future recognition into matching against a stated structure. The new note earns its place when its expected reuse, and its power to discriminate instances from lookalikes, exceed the context cost of one more abstraction.

Naming a shared mechanism is itself a conjecture. It enters the discovery lifecycle at the conjecture phase and earns acceptance as later instances confirm, refine, or refute the proposed structure.

Scope

The claim compares recognition with articulation, not with every linking activity: a link still needs a reader-relevant reason, and judging that reason is real work. It covers similarity-based connections; contrastive, causal, and temporal links are judgments of a different kind. And "hard" locates where cost concentrates, not a fixed ordering — with strong prior vocabulary, deep recognition can be cheap, and without co-presence, even surface recognition fails.

Relevant Notes:

Operationalized into:

  • /connect skill — the articulation test operationalizes the cheap half: a proposed link must state the specific relation that was recognized