Written down first. Checked after.
HGE, the Hypothesis Generation Engine, is the evidence engine behind our research: it forms testable hypotheses, runs them against instruments and signs the evidence. Every result can be replayed from the same data.
The claim is written down before the test.
HGE states the claim first, stores it where it cannot be changed, and lets the outcome decide whether it held.
CERES for famine risk, Laplace for Gaia DR3, MARVIS at sea and the quantum evaluation in Osaka all run on HGE.
What it reads.
- Instruments
- Any source with an adapter: star catalogues, satellite archives, humanitarian data, a quantum computer.
How it works.
- Data
- Sources are harmonised, checked for outliers and drift, and given a provenance event as they arrive.
- Hypotheses
- Each claim is a machine-readable JSON-LD hypothesis with bounds, priors and the test that would prove it wrong.
- Experiments
- Experiments are ranked by expected information gain within what the instrument and the budget allow, and pass a policy gate before they run.
- Evidence
- The outcome, the updated confidence and the provenance chain are packed into a signed evidence bundle that replays to the same result.
Technical foundation.
- Hypotheses
- JSON-LD schemas, Bayesian prioritisation of experiments.
- Integrity
- Ed25519 signatures, SHA-256 hashes, provenance aligned with W3C PROV.
- Governance modes
- Disabled, logging or enforced, so an institution can move from watching to enforcing.
- Integration
- An instrument interface, a REST API for experiments, webhooks for slow instruments, containers, PostgreSQL or MongoDB.
In figures.
- 1.8 bn
- Gaia DR3 objects in the validation run
- 6
- Attack vectors in the red-team tamper test
- 26
- Experiments on quantum hardware
- 4
- Operations in the HGE Core API
One engine under every system.
The HGE Core API went into production in January 2026 with four operations: submit a hypothesis, score it, move it to a new state and read the ledger. Laplace, CERES, ORION, MARVIS, AION and ATHENA are adapters on it. The evaluation on the OQTOPUS quantum computer in Osaka ran on it too.
We publish the evidence and keep some of the recipe: the internal ranking heuristics, the operating parameters tied to a deployment and the security mechanisms that would weaken tamper resistance if disclosed.
Papers, code and data.
Who it serves.
Research institutions that want to run a domain on the engine, or review how it reaches a conclusion, can ask for a technical briefing.
Write to us- Research institutions that need results others can replay.
- Public bodies and regulators reviewing how a conclusion was reached.
