Representation
Can one architecture infer useful state representations across unrelated quantitative systems?
LargeQuant research programme
Building the foundations of Quantitative General Intelligence.
The programme is deliberately staged: define the architecture, make generality measurable, build the quantitative control plane, test transfer across domains, and publish evidence before making stronger intelligence claims.
01DEFINE
Category & architecture
Formalise QGI, its native primitives, its relation to LQMs and its claim boundary.
02MEASURE
Benchmarking & transfer
Develop cross-domain tasks, capability levels and evidence requirements with LQBench.
03BUILD
Quantitative cognition stack
Federate models, solvers, simulators, uncertainty, evidence and scientific memory.
04AUTONOMISE
Research loops
Bounded hypothesis → experiment → evidence → next-experiment cycles.
05VERIFY
External challenge & reproducibility
Expose results to independent recomputation, provenance inspection and challenge.
Research questions
The programme is organised around questions that can fail.
Transfer
What capabilities genuinely transfer from finance to engineering, biology or physical science?
Uncertainty
Can uncertainty remain calibrated as a system changes models, tools and domains?
Experimentation
Can a system choose experiments that reduce uncertainty rather than merely generate more outputs?
Memory
What must be retained so failed hypotheses and contradictory evidence improve future reasoning?
Verification
What evidence is sufficient for an autonomous quantitative conclusion to be trusted?
Public research doctrine
Definition → measurement → demonstration.
DEFINE
→Publish what QGI means and what it does not mean.
MEASURE
→Create tasks that can falsify claims of generality.
DEMONSTRATE
Show cross-domain quantitative work with evidence.
Research boundary
LargeQuant will not use QGI as a synonym for “advanced AI.”
The term should earn meaning through architecture, cross-domain performance, uncertainty, reproducibility, scientific memory, autonomous experimentation and evidence. Until those standards are demonstrated, QGI remains a research programme and proposed destination.
PUBLIC QGI R2 / Measurement
The measurement layer is now public.
LQBENCH-QGI/0.1 defines the proposed capability dimensions, domain families, task contract, evidence contract and level policy. It does not award QGI levels or publish a universal intelligence score.
PUBLIC QGI R3 / Calibration
The evaluation methodology is now public.
LQBENCH-QGI/0.2 defines task families, baseline classes, held-out transfer regimes, calibration methodology, resource accounting and evaluation manifests. No level threshold or result is fabricated.
PUBLIC QGI R4 / Demonstrate
Turn the public protocol into evidence-bearing demonstrations.
LQBENCH-QGI/0.3 defines pre-registration, reference cross-domain tracks, evidence packets and publication states while keeping public result arrays empty until real runs exist.
PUBLIC QGI R5 / Execute
Publish real evidence without publishing a certification claim.
LQBENCH-QGI/0.4 adds a controlled operator publisher, immutable result IDs, evidence packet hashes and a read-only public result registry.
PUBLIC QGI R6 / Validate
Independent review becomes signed evidence, not a popularity score.
LQBENCH-QGI/0.5 binds independent reproduction, method review, corroboration and contradictory evidence to exact published result hashes.
PUBLIC QGI R7 / Compare + Federate
Compare under one task contract. Federate validation without voting on truth.
LQBENCH-QGI/0.6 adds immutable same-task comparison cohorts and cross-result signed-validation evidence coverage.
PUBLIC QGI R8 / Generalize
Test transfer across materially different quantitative domains.
LQBENCH-QGI/0.7 binds source and target evidence, transfer manifests, held-out regimes and adaptation budgets into inspectable cross-domain transfer records.