Analogical Reasoning
TL;DR
Analogical reasoning solves a new case by mapping it onto a settled one: retrieve the precedent, align the structure — not the surface — and transfer the conclusion along the mapping. It is how case law works, how policy is applied, and how experienced adjudicators actually think. It is also the easiest strategy to do badly, because surface similarity masquerades as structural similarity. The graph makes the mapping an explicit artifact and puts its weight on a disanalogy critique: a dedicated node hunts for the differences that break the transfer, and the gate admits the conclusion only if the mapping survives them. Like cases alike — and a recorded account of “like.”
“We had one just like this”
A new dispute lands and the reviewer relaxes: we had one just like this in March. Same merchant category, same dollar range, same claim of an unauthorized recurring charge. March was a denial; this will be too. The write-up nearly authors itself — which is exactly when it goes wrong. The March consumer had e-signed an enrollment at account opening; this one enrolled through the merchant, with nothing on the bank’s side. Same surface. Different spine.
The instinct wasn’t bad — precedent is how consistency happens, and consistency is a regulatory virtue. The failure was that the resemblance was never inspected: what actually made March a denial, and does this case have that load-bearing feature, or only the costume?
An unguided model matches on surface similarity — that is literally what next-token statistics reward. The analogical graph must state the mapping and then survive a dedicated search for the differences that break it.
How the graph works, step by step
Retrieve the source case. The precedent arrives as evidence, not memory: the settled case with its facts, its outcome, and — critically — its rationale. What made the outcome right is the thing to be transferred; without the rationale, the outcome is just a costume to copy.
Generate the structural mapping. A dedicated node aligns the cases feature by feature, separating structure from surface: which facts in the source did the rationale actually rest on, and what corresponds to each in the target? The mapping is an explicit artifact — each correspondence stated, each gap visible.
Critique the disanalogies. The strategy’s center of gravity. A critique node searches for the differences that break the transfer: features the rationale depended on that the target lacks, and features the target has that the source’s rationale never contemplated. Surface matches with structural breaks — the March enrollment signature — die here.
Evaluate the transfer and gate it. What survives critique is transferred along the mapping, with adjustments the disanalogies force. The gate admits the conclusion only with the mapping and the survived critique attached: the trace records not just “like March” but which features carried the likeness and which differences were weighed.
The gates, operationally
| Gate | What it checks | Fails when |
|---|---|---|
| Source gate | The precedent is retrieved with facts, outcome, and rationale, and its own status is sound (not overturned, not distinguishable on its face) | The “precedent” is remembered rather than retrieved, or arrives as an outcome with no rationale to transfer |
| Mapping gate | Correspondences are stated feature-by-feature and tied to the source’s rationale, not its surface | The mapping rests on category, amount, or phrasing resemblance that the rationale never depended on |
| Disanalogy gate | A genuine search for breaking differences ran, and each found difference is either answered or reflected in an adjusted conclusion | The critique lists only differences that don’t matter — the strawman version of skepticism — or a load-bearing difference is noted and ignored |
Where it fits — three use cases
1. Precedent-consistent adjudication
Identical facts resolving differently across cases is its own exam finding. Analogical structure is how consistency becomes inspectable: every determination that leans on a prior one records the mapping, so “we treat like cases alike” is a claim with evidence rather than a slogan.
2. Policy application to novel fact patterns
Policies are written for the cases their authors imagined. When a new pattern arrives — a payment type, a channel, a product the policy predates — analogical reasoning is the honest tool: map the new pattern to the nearest contemplated one, and let the disanalogy critique say whether the policy’s rationale actually reaches it or whether this is a gap to escalate.
3. Vendor and model-risk comparability
“This system is like the one we validated last year” is an analogical claim, and due diligence teams make it constantly. Running it as a graph forces the question the shortcut skips: which properties did last year’s validation actually rest on, and does the new system share those — or just the vendor’s category and the deck’s vocabulary?
When to reach for it
Reach for analogical reasoning when settled cases exist and consistency with them is itself part of correctness — adjudication, policy application, anything with precedential weight. Be suspicious of it when the domain shifted under the precedents, and pair it with abstention: the honest output of a mapping that fails critique is “this case is genuinely new,” routed to a human as first-impression — not a forced fit to the nearest old answer.
- Structure vs. surface
- A mapping is licensed by the features the source’s rationale rested on — not by category, amount, or phrasing resemblance.
- Disanalogy critique
- A dedicated search for the differences that break the transfer — the step that separates precedent-following from pattern-matching.
- First-impression routing
- When no mapping survives, the case is declared genuinely novel and escalated — the analogical graph’s form of abstention.