RESEARCH PROJECT / STRUCTURE ABSTRACTION

Solve problems across fields
with the same structure.

Remove the concrete names and keep only the roles and relations.
Then knowledge proven in an entirely different field
becomes a clue to the problem in front of you.

What is structure abstraction? ↓

ONE STRUCTUREMANY FIELDS

Electric circuit Water tank Room temperature

STRUCTUREA difference drives a flow,
resistance slows it, capacity stores it. τ = R × C

Three different-looking phenomena share one shape of relations.

A research topic LMR pursues on its own computing infrastructure.

01 / CONCEPTWhat is structure abstraction?

Structure abstraction Drop the names, keep the roles and relations.

Structure abstraction removes the concrete names from a case and extracts only the roles each part plays, the relations between those roles, and the conditions under which the relations hold. The extracted structure can be compared across fields.

What we extract: Roles (what resists flow, what stores), relations (what rises when something else rises) and conditions (where they hold).
Not enough on its own: Tagging, summarising or turning text into vectors can all help, but if the relations and conditions a question needs are lost, it is not structure abstraction.
Theoretical basis: Structure-mapping theory from cognitive science, which treats analogy as carrying over the structure of relations, and qualitative physics and bond graphs, which describe electrical, thermal and fluid systems with common quantities.

Analogy Structure mapping Qualitative reasoning

FOUR LAYERSInformation kept in four layers

Layer	Example (electric circuit)	Why keep it
Concrete facts	Resistance 2, capacitance 3, a step change in input	The baseline for citing and re-checking
Roles & relations	A part that resists flow, a part that stores, a lag in response	The core for comparing fields
Conditions	Component values are constant; change settles exponentially	Where a transfer still holds
Prediction	Double only the resistance and the time constant doubles	To test whether it is right

Roles and relations are what we compare; facts, conditions and predictions are kept alongside. Values are illustrative.

KEY IDEA

What matches is the shape of the relations, not the appearance.

02 / EXAMPLEElectric circuit and water tank

The same structure Predicting a water level with circuit knowledge.

A circuit charging a capacitor and a tank filling with water over a drain look nothing alike. Extract the structure, though, and both take the same form: a difference drives a flow, resistance slows the flow, and capacity stores it.

Correspondence: Voltage difference and water level difference, current and flow rate, electrical resistance and drain resistance, the capacitor and the tank’s capacity each play the same role.
Transferable knowledge: The circuit’s well-established time constant, τ = R × C, directly describes how fast the tank’s water level settles.
Prediction and test: The circuit prediction “double only the resistance and the time constant doubles” becomes, for the tank, “double the drain’s resistance and the time to settle doubles.” We check this by calculation.

Role matching Formula mapping Testing predictions

MAPPINGMatch by role

Electric circuit	Structure (role)	Water tank
Voltage difference	Driving difference	Water level difference
Current	Flow	Flow rate
Resistor	Resists flow	Drain resistance
Capacitor	Stores	Tank capacity
Time constant R × C	Response lag	How fast the level settles

Matched by role, the circuit’s formula carries over to the tank. Heating a room (temperature difference, heat flow, insulation, heat capacity) has the same structure.

KEY IDEA

A formula proven in one field becomes a prediction in another.

03 / LENSESEight lenses for extracting structure

Eight lenses Questions that prevent blind spots.

When extracting structure from text, we ask questions through the following eight lenses. Relations that fit none of them are captured as “open”, and we never force a lens without evidence.

Structure & dependency What, if missing, makes the whole fail?
Cause & effect Change what, and what changes? Is that a fact or an assumption?
Flow & accumulation What comes in, goes out, accumulates or stalls?
Time & change How do delay, order, cycles and lifetimes matter?
Information & observation What is unseen, and who is not informed?
Control & feedback How are goals, measurement and correction linked?
Constraints & trade-offs What narrows the options?
Actors, goals & incentives Who wants what, and what do they respond to?
Open Which relations do the eight lenses miss?

04 / HOW IT WORKSExtract, map, verify

Five stages Not plausibility, but predictions you can check.

01Extract
Roles, relations and conditions are extracted from cases and papers through eight lenses. This reading is done by AI.
02Map
The problem is matched against stored cases by the structure of their relations. When relations correspond, their elements are paired one to one.
03Hypothesise
The matched case’s formula is rewritten through the correspondence to produce the prediction mechanically. AI never writes the formulas.
04Test
Predictions are checked by independent calculation, from consistent units to numbers.
05Record
We keep the cases, facts and correspondences each prediction came from as its evidence.

MODE 01

Judgment

Decides which known law applies to a new problem. When the given facts cannot settle it, it defers and returns questions or observations that would tell the options apart.

MODE 02

Proposal

From a numeric table and a description, it lists candidate mechanisms (saturation, power laws, queues and more) with formulas and fit evidence, and suggests observations that tell the candidates apart.

05 / WHY IT MATTERSWhat it is good for

How it helps Wisdom from other fields, for the problem at hand.

Borrow insight from other fields

Even for a new problem, if the structure matches, a model proven over many years in another field can be applied. It gives a sound starting point faster than building hypotheses from scratch.

Propose mechanisms from data

From tables of numbers gathered in experiments or operations, it proposes the likely mechanisms behind them, with formulas, and shows what to measure next to tell the candidates apart.

Say “not sure” when unsure

When the given facts cannot settle a problem, it does not force an answer. It defers and asks a question that would tell the options apart, preventing a wrong transfer.

Traceable evidence

Every prediction records the cases, facts and correspondences it used. It is a hypothesis you can check by calculation, not merely plausible text.

Where we see it helping

Forming hypotheses in R&D and planning the next experiment
Diagnosing the causes behind how products and equipment behave
Reading the mechanisms hidden in survey and operational data
Reusing the cases and knowledge an organisation has built up

These are envisaged uses, not a record of deployments.

06 / STATUSWhere we are now

Where the research stands Step by step, checking as we go.

Structure abstraction is a research topic LMR pursues on its own computing infrastructure. Results are checked with experiments whose judging criteria are fixed in advance.

What we have checked: Judgment, which applies known laws, has been checked on purpose-built test problems: both that its decisions are correct and that it defers problems it cannot settle.
In progress: Validating proposals on real research data, and growing the library of cases.
Computing: Local AI models and embedding models run on our own GPU servers and Mac Studio.

REFERENCES