Researchbeyondthe norm.beyondthe norm.

Cumulant Research

Core positioning

“Researchers define the questions, challenge the assumptions, interpret the evidence, and control what becomes public. AI agents search literature, analyze data, test robustness, criticize findings, and reproduce results.

Neural network · visualizationThe two intelligences Cumulant pairs: human judgment and machine reasoning.

Transparent computational research.

Cumulant Research develops transparent computational methods for difficult questions involving capital, risk, intelligence, and human systems.

Its work combines human research direction, specialized AI agents, deterministic analytical pipelines, and public accountability.

The goal is not to automate judgment. The goal is to give researchers greater reach while preserving traceability, criticism, reproducibility, and human responsibility.

Tissue cross-section · microscopyThe fine detail computational methods are built to read.

Pigment in water · macroRaw evidence before it resolves into a finding.

Marbled pigment · macroStructure emerging from noise.

From question to auditable result.

Cumulant's workflow is designed to make research inspectable. Software agents help retrieve evidence, write and test code, challenge findings, and preserve research state.

Deterministic systems verify estimates, citations, artifacts, and progression gates. Human researchers define the question, review the design, interpret the evidence, and decide what becomes public.

Define

01Frame

Define one specific and falsifiable research question.

02Map

Review prior literature, methods, disagreements, and unresolved gaps.

Build

03Retrieve

Acquire data and record its source, coverage, and provenance.

04Validate

Check missingness, duplicates, alignment, transformations, and sample construction.

05Declare

Specify the main outcome, assumptions, model, exclusions, and robustness plan.

Interrogate

06Analyze

Run deterministic code and save machine-readable outputs.

07Challenge

Use critical review to search for leakage, weak identification, alternative explanations, and unsupported claims.

08Stress-test

Test alternate samples, variables, models, windows, costs, assumptions, and out-of-sample behavior.

Confirm

09Verify

Trace numerical claims and citations to saved evidence.

10Reproduce

Rebuild results from preserved inputs, code, seeds, and configurations.

11Review

Require human evaluation of the design, interpretation, limitations, and release decision.

12Release

Publish a versioned output with methods, code, data status, limitations, and correction history.

Different strengths. One accountable process.

The partnership is designed to increase research capacity without outsourcing judgment.

Stained cells · light microscopyHuman researchers frame the questions and own the judgment.

Microprocessor · macroAI agents extend the work: search, drafting, and iteration at scale.

Data-centre infrastructureDeterministic systems keep every step fixed and reproducible.

Current research.

Cumulant studies difficult questions across finance, risk, statistical inference, model reliability, humanitarian allocation, and AI-assisted research systems.

Financial inference and model risk

Statistical inference and model reliability in financial data, where limited decision counts and flexible analysis can make findings fragile.

FOCUS

Backtest inference

Decision timing

Researcher degrees of freedom

Model selection

False discovery

Robustness

Event studies

Statistical fragility

Financial inference and model risk — Network telemetry · monitoring viewFinancial inference: separating signal from false discovery.

Featured projects.

Current and archived work. Each project links to its question, methods, data status, and limitations. Statuses are honest: in progress, working paper, research note, experimental, replication, or null result.

Program

Status