Data protection
Governs access
Tells us who may touch the data. Never asks how it was constituted, or whose interests shaped its production.
The Public Reason Project builds open infrastructure for warranted trust: reliance grounded in evidence anyone can inspect, fork, and verify — not in reputation or authority. Our first standard is the OpenReason Protocol.
Data does not fall from the sky. Every dataset is the product of decisions — what to measure and what to ignore, which populations to include and which to leave out, what counts as a finding and what gets filed away. Those decisions are made inside institutions with interests, funding relationships, and incentives that shape what they choose to see.
None of it requires fraud. It requires only that the people who fund knowledge have interests — and that the people who produce it need funding. This is not an edge case. It is the default condition of data production in the modern world.
We are automating the distortions of the past and calling it progress.
Tells us who may touch the data. Never asks how it was constituted, or whose interests shaped its production.
Examines what AI systems do with data. Cannot detect distortions that entered before the model ever saw a record.
Makes data available to everyone. Democratising a distorted dataset doesn't correct the distortion — it spreads it.
Checks equity across groups in the evaluation set. Cannot measure the people who were never in the data at all.
Each is genuine. Each addresses something real. And each begins downstream of where the problem lives.
Five commitments about method, not conclusions. They don't tell you what to believe — they define what it takes for a belief to be warranted.
Claims of benefit or harm must be testable with evidence — not asserted.
Assumptions are hypotheses to be stated and stress-tested, not settled facts.
Model downstream effects before committing — and make the model re-runnable by others.
No affected party is invisible. Minorities are tested, not averaged away.
Every decision is logged, attributable, and auditable.
Infrastructure, not regulation — the epistemic equivalent of double-entry bookkeeping.
An open standard that turns the reasoning behind a decision into a structured, inspectable, verifiable object — five layers, from data provenance to lineage, with cryptographic attestation.
Explore OpenReason ↗The academic case for provenance-grounded governance — the constitutive problem in data, and REE as its philosophical foundation.
Apache-2.0. No single owner. Disagreement is resolved by forking and showing better reasoning — not by vote. Designed to function without its founders.
Putting ORP to work on real consequential decisions — beginning with environmental and tax law, documented end to end.
We reconstructed four of the most consequential AI training datasets as full OpenReason documents — more than 8,000 lines of documented provenance. Each surfaced constitutive decisions that Model Cards and Datasheets never capture. Every figure links to the full reconstruction, open to fork and check.
Shaped a decade of vision models.
Billions of generated images a day.
The web corpus under most LLMs.
Suggestions drawn from everyone's code.
ORP attests to provenance and reasoning — who said what, when, citing which sources and assumptions. It never issues a verdict on whether a claim is good or correct. The moment a standard grades content, it loses the neutrality that makes an attestation worth anything.
Every claim should be checkable without trusting the author — by hand for a person, by signature for a machine. Trust is earned by evidence that holds up, not by who is making the claim.
The protocol, the spec, and the tools are open. Read it, implement it, challenge it — and if you can reason better, fork it and show us.