map-analogy-to-concept

Map Analogy to Concept

Table of Contents

• Workflow
• Mapping schema
• Worked example
• Guardrails

Related skills: Called by the Intuition Builder per framing, after generate-analogy-set and before stress-test-analogy. Gentner's structure-mapping theory is the theoretical spine: good analogies map relations, not just objects.

Workflow

For one framing (source → target):
- [ ] Step 1: Enumerate the source domain's key components (entities + relations)
- [ ] Step 2: For each source component, propose the target component it maps to
- [ ] Step 3: Check systematicity — do the relations carry across, or only objects?
- [ ] Step 4: Flag any source component that maps to nothing concrete (vague mapping = reject)
- [ ] Step 5: Return the mapping table

Step 3: Systematicity check

A strong analogy preserves the pattern of relations, not just object-level similarity. Example:

• Weak: "a neural network is like a brain" — both have "neurons", but the relation "neurons fire" doesn't carry over in a useful way.
• Strong: "V/D/J gene recombination in B cells maps to multi-agent diversity" — the relation "small vocabulary generates exponential combinatorial space" carries.

If the framing only matches on objects (nouns), reject or downgrade.

Step 4: Unmapped flagging

Every source component must map somewhere. "It's like a brain" fails because "brain" is unmapped to anything specific in the target (neuron? cortex? entire NS?). Flag and reject.

Mapping schema

source_domain: "library card catalog"
target_concept: "KV cache"
mapping:
  - source: "library"
    target: "the KV cache data structure"
    relation: "contains"
  - source: "drawer"
    target: "cache slot"
    relation: "capacity-bounded container"
  - source: "card"
    target: "(key, value) projection pair"
    relation: "indexed entry"
  - source: "lookup by drawer then card"
    target: "retrieval by position in key tensor"
    relation: "indexed retrieval"
  - source: "eviction when drawers fill"
    target: "LRU / FIFO eviction under context-length pressure"
    relation: "replacement under capacity constraint"
systematicity_score: 4/5  # how well relations carry over
unmapped_source: none
unmapped_target: "the attention operation that reads this cache"  # flagged — see stress-test

Worked example

Framing: "Dropout is antibody diversity for weights."

Source components:

• immune system
• antibody population
• pathogen recognition
• diversity generation (V/D/J recombination)

Target components:

• neural network
• weights under dropout
• generalization to unseen examples
• implicit ensembling via sub-networks

Mapping:

Source	Target	Relation
immune system	the trained neural network	generates patterns from a small genome/parameter set
antibody population	ensemble of thinned sub-networks	many variants tested in parallel
pathogen recognition	generalization on test data	performance on unseen inputs
V/D/J combinatorial generation	random dropout masks produce sub-network diversity	small seed → many variants

Systematicity: 4/5 — the relation "small number of building blocks → large functional diversity" carries across. The one break: actual biological V/D/J has selection (negative selection in thymus), which dropout doesn't do. Flag.

Guardrails

1. Every source component maps to a concrete target component. No source is left unmapped.
2. Rate systematicity on a 1–5 scale. Below 3 = the analogy is object-level, not relation-level; flag for downgrade.
3. Do not generate the analogy itself — that's generate-analogy-set. This skill only maps.
4. Flag unmapped target components as candidates for the stress-test boundary.
5. Mapping output is structured (table or yaml) — don't write it as prose.

Quick reference

• Input: one framing (source + target).
• Output: structured mapping + systematicity score + unmapped flags.
• Theoretical basis: Gentner's structure-mapping; relations > objects.