lyndonkl/claude/check-analogy-novelty

Check Analogy Novelty

Table of Contents

• Workflow
• Match tiers
• Worked example
• Guardrails

Related skills: Called by the Intuition Builder after generate-analogy-set and before presenting framings. Reads shared-context/analogy-catalog.md. Does NOT write to the catalog — that's update-analogy-catalog's job, fired on publish.

Workflow

For each of the 5 framings:
- [ ] Step 1: Extract the analogy's source domain (e.g., "library card catalog")
- [ ] Step 2: Extract the analogy's target concept (e.g., "KV cache")
- [ ] Step 3: Grep analogy-catalog.md for exact or near-exact source × target match
- [ ] Step 4: Classify: new | reused-from-catalog | adjacent-to-catalog
- [ ] Step 5: For reused/adjacent, cite the catalog entry with its post

Match tiers

• New: neither the source nor a close variant of it appears in the catalog for this target concept.
• Reused-from-catalog: exact source × target match — writer already used this analogy in a published post. Flag strongly; either justify reuse or swap.
• Adjacent-to-catalog: source matches a nearby analogy in the catalog (e.g., "drawer" vs. "card catalog") or target is in the catalog but source differs. Flag softly — writer should see the adjacency.

"Close variant" = same source domain (library, immune system, bucket brigade) even if different specific instance.

Worked example

Proposed framings for KV cache:

1. Everyday: "a library's card catalog with a fixed drawer count"
2. Physical metaphor: "a ring buffer with index"
3. Contrarian: "the cache is not the table — it's the index into what's been already paid for"
4. Historical: "before caches, every lookup was recomputation"
5. Counterfactual: "remove the cache and decode becomes quadratic"

Catalog check:

• Entry in catalog: "Agents as departments reading each other's memos but never sitting in on each other's meetings" — source: institutional. Target: multi-agent. Not related to KV cache.
• Entry: "Bucket brigade" — source: bucket brigade. Target: sequential multi-agent. Not related.
• No entry for library → KV cache, or ring buffer, or index.

Output:

1. Everyday: new
2. Physical metaphor: new
3. Contrarian: new
4. Historical: new
5. Counterfactual: new

But: if the catalog had an entry "Library index → tokenizer vocabulary", framing 1's adjacency would trigger: "adjacent-to-catalog — library appears once for a different target". Flag softly.

Guardrails

1. Read-only. Never writes to the catalog.
2. Match on both source domain and target concept. Same source for different target is an adjacency, not a reuse.
3. Flag adjacencies, don't auto-reject — the writer decides if reuse is intentional (a recurring analogy can be a feature).
4. Cite the catalog entry + post when flagging. "Reused from catalog" without pointing to the entry is useless.
5. Don't expand near-matches to other analogy directions (biology vs. organizational are different domains, never flag across).
6. If the catalog is empty (early-stage publication), all framings are new by definition. Note this once in the output.

Quick reference

• Input: list of 5 framings + analogy-catalog.md.
• Output: per-framing classification {new | reused-from-catalog | adjacent-to-catalog} with citation.
• Side effect: none. Read-only.