Learned representations first
Language and multimodal models increasingly reason, use tools, execute code and call specialised quantitative systems.
prompt → representation → reasoning → tool → answerLargeQuant Research
Quantitative General Intelligence
Intelligence designed to reason across measurable reality.
Quantitative General Intelligence (QGI) is LargeQuant's research framework for intelligence that can represent, model, simulate, predict, optimise and experimentally interrogate quantitative systems across domains while preserving uncertainty, reproducibility and evidence.
Research claim boundaryLargeQuant is developing and measuring the QGI architecture. It does not claim that general intelligence has been achieved.
LIVE QUANTITATIVE FIELDQUANTITATIVE COGNITIONObjective → measurable state → evidence-bearing decision
The architectural distinction
Quantitative-native cognition with language as an interface.
Quantitative cognition first
State, models, constraints, simulation, uncertainty, experiments and evidence are first-class parts of the reasoning loop.
objective → state → model → experiment → evidence → next stateA different epistemic contract
The output is not principally an answer. It is a defensible quantitative result.
STATEWhat is numerically true now?
DYNAMICSHow does the system change?
UNCERTAINTYWhat remains unknown?
CONSTRAINTSWhat is physically or operationally permissible?
OBJECTIVEWhat outcome are we seeking?
EXPERIMENTWhat should we measure or simulate next?
EVIDENCEHow can the conclusion be reproduced and checked?
MEMORYWhat should the system retain for the next cycle?
Where LargeQuant fits
LQMs are quantitative engines. QGI is the intelligence architecture. LargeQuant is the operating layer.
Large Quantitative Models provide domain-specific numerical cognition. Solvers and simulators provide execution. Evidence, uncertainty and scientific memory provide the control boundary. QGI coordinates those capabilities across quantitative systems.
QGI
Quantitative General Intelligence
orchestratesQuantitative General Intelligence
Large Quantitative Models
ModelsSolversSimulatorsExperiments
acts onMeasurable reality
Generality must be measured
QGI should be falsifiable, benchmarkable and cross-domain.
QGI-0Specialised
QGI-1Adaptive
QGI-2Integrated
QGI-3Cross-domain
QGI-4Autonomous scientific
QGI-5Open quantitative intelligence
These levels are a research measurement scale, not LargeQuant product certifications.
Explore the framework
Architecture, measurement, evidence and generalisation.
01
QGI Architecture
The quantitative cognitive and execution stack.
02QGI vs. AGI
A precise distinction without underestimating frontier AI.
03Measurement & Benchmarking
How generality can be measured across quantitative domains.
04Research Programme
The research agenda, evidence programme and open questions.
Measurement
Generality should be measurable before it is marketable.
LQBench-QGI defines capability dimensions, quantitative domain families, explicit task manifests and evidence requirements for evaluating quantitative generality.
NO UNIVERSAL QGI SCORE
The framework does not certify a system as generally intelligent. Numeric thresholds require reproducible calibration across benchmark families.
Calibration
Turn measurement vocabulary into executable evaluation methodology.
The calibration framework defines task families, reference baseline classes, held-out transfer regimes, resource accounting and machine-readable evaluation manifests.
MEASUREMENT BEFORE CLAIMS
Methodology is not evidence.
The framework defines how evidence should be produced without manufacturing a score, level, certification or benchmark victory.
Demonstrations
Evaluation should produce evidence-bearing demonstrations.
Pre-registration, cross-domain demonstration tracks and public evidence packets make results inspectable without converting a demonstration into a general-intelligence claim.
Evidence registry
Real evaluation evidence can be published through a controlled, verifiable path.
Operator-controlled publication and a read-only hash-verified result registry preserve the distinction between measured evidence and product claims.
Independent validation
Independent evidence can support, question or contradict a result.
Signed review statements bind to exact result and evidence hashes while keeping challenge, contradiction and uncertainty visible instead of collapsing review into a trust score.
Comparative evaluation
Compare systems under the same task contract and preserve the evidence behind the comparison.
Comparison cohorts bind to exact result and evidence hashes while uncertainty, resources, failures, validation and contradictions remain inspectable.
Cross-domain generalisation
Transfer should be recorded as an inspectable evidence relation—not inferred from a leaderboard.
Source and target results remain bound to exact manifests, transfer regimes, held-out conditions and adaptation budgets so cross-domain evidence can be inspected without inventing a universal generality score.
LargeQuant
Advancing the infrastructure for Quantitative General Intelligence.
Models, simulation, uncertainty, evidence, provenance, scientific memory, experiments and cross-domain quantitative workflows are already part of the operating system that this research builds on.
Explore LargeQuant