Where the four fields meet, and where we test what comes next.
Five research directions and four tracked research projects, including publicly evidenced work and an internal research prototype. Each project shows the evidence behind its status, and what is still missing.
Our research is internal, sealed and exploratory. Read every result on this page with these limits.
We distinguish research, demonstrated capabilities and production-ready technology through documented evidence.
An open question, stated before any work starts.
Experiments are running. Results are not public yet.
It exists and is documented. Nobody outside has checked it yet.
Public evidence that we, or anyone, can check again.
When the evidence is not enough to place something, it is marked Status to confirm, and it stays there until the evidence exists.
Each project also names the kind of evidence behind it, so a simulation is never read as a hardware result, or a failed test as a success.
Each direction lists its tracked projects. Where there is none yet, the page says so.
New methods for AI systems, explored before anything is built on them.
AI, Data
How autonomous systems decide, act and stay accountable, online and offline.
AI, Blockchain
Ways of computing beyond the usual cloud setup, from edge devices to unconventional models of computation.
No published project yet.
Quantum hardware and algorithms, and what quantum computing will mean for today's cryptography.
Exploratory runs on IBM quantum hardware exist; the documented kernel experiment produced a negative result. It is described, with its limits, on the Quantum & Scientific Computing page. The project listed here is post-quantum cryptography.
Data, Blockchain, Quantum
The algorithms behind verification, search and decision-making, studied on their own terms.
No published project yet.
Four tracked research projects: three with public evidence and one internal research prototype. Each states what was checked, when, and which evidence is still missing.
Research questionCan layered checks catch false statements about regulation that a single model check would let through?
Why it mattersAI-written statements about rules end up in reports, filings and decisions. A check against the source text makes errors visible before they spread.
UVO checks statements about regulation against the regulation text and returns a signed receipt. A nightly test, in which a model writes new false statements, publishes its run history as open data.
Evidence limitationsNo run compares the layered checks with a single check, so the hypothesis is untested. No independent evaluation; the test statements come from the project's own pipeline.
Evidence classDocumented implementation with public run telemetry
Public evidence
Run history retrieved on 31 October 2026; the latest run was that night.
Research questionCan a stranger confirm, offline and without an account, that an agent action was authorized, ran as authorized and was paid for?
Why it mattersWhen an agent acts on someone's behalf, a third party should be able to confirm what was authorized and what ran, without trusting the operator.
One record binds an agent's request, its authorization, the execution, the result and, where it applies, the payment. Anyone can check it offline.
Evidence limitationsNo independent third party has validated the reference yet. Replay protection cannot be confirmed by a stateless check.
Demonstrated forVerification of the published reference receipt
Evidence classPublicly demonstrated capability, reference case only
Public evidence
We verified the published reference receipt on 31 October 2026 against the separate execution JWKS.
Research questionCan records carry a post-quantum signature next to a classical one without breaking the verifiers that use them?
Why it mattersRecords signed today may need to be checked after quantum computers can break the signature schemes in common use.
feedoracle.io publishes an ML-DSA-65 public key next to its classical ES256K keys. ML-DSA is a post-quantum signature scheme: it runs on ordinary computers and is designed to resist attacks by future quantum computers.
Evidence limitationsProduction-wide post-quantum signing and the validation of dual signatures are not demonstrated.
DemonstratedML-DSA-65 public key published, checked 31 October 2026
Evidence classResearch; key publication demonstrated separately
Public evidence: key only
Key set retrieved on 31 October 2026.
Research questionCan a device decide whether an action may run while it has no connection, keep a signed record, and have that record checked once it reconnects?
Why it mattersDevices in the field cannot always reach a server before they act, yet their actions still need rules and a record.
An internal research prototype for authorization and signed records on devices that lose their connection.
Evidence limitationsNo physical hardware tests, no production deployment, no certification. Evaluation material is internal and not published.
Evidence classSimulation-only result, internal
Internal research prototype
Evaluated in a synthetic environment only.
Projects appear here when their evidence can be shown and their publication is approved.
Each hypothesis is a provisional research question, not a confirmed finding.
Several independent checks catch false statements that a single model check lets through.
Unverified hypothesis
Nightly runs against statements the system generates itself, each logged publicly.
A public record of each nightly run against the system's own test statements.
A comparison with a single-check baseline, and an evaluation by someone outside FeedOracle.
A stranger can check, offline and without an account, that an action was authorized, ran as authorized and was paid for.
Supported for the published reference record
Published reference records for one permitted and one refused action, with a stateless verifier.
The receipt digest recomputes, the Ed25519 signature verifies, and a changed field makes the check fail. An engineering check on one reference record, not a statistical result.
Validation by an independent party; replay protection a stateless check can confirm.
Records can carry a post-quantum signature next to a classical one without breaking existing verifiers.
Unverified hypothesis
So far, only the public key is published. No dual-signature verification has been run here.
Key publication only.
Responses carrying both signatures, checked by an independent verifier.
Authorization and signed records can work offline and be reconciled safely after reconnection.
Unverified hypothesis
A synthetic evaluation of offline authorization, local signing and queued recovery.
Internal results, not published.
Tests on physical hardware and evaluation material we can publish.
Research starts in the fields and its results go back into them: execution-bound proofs into AgentNOMOS, claim verification into FeedOracle's services.
Primary fieldAlso draws on
This space is reserved for the interactive Research Graph. It will show fields, projects, directions and evidence as one connected picture, and how results flow back into the fields.
Until projects are published as machine-readable records, it shows a static preview, drawn only from the four projects on this page.
Drawn in the previewWritten on this page, not yet linked as data
PlannedInteractive graph
We do not list partners we do not have. These are the ways in today.
Bring a question that sits between two fields, or between a field and one of the five directions. The collaboration page explains what a useful proposal contains.
Projects with public evidence publish what you need to check them yourself. If your result differs from ours, we want to know.
Every project draws on at least one of the four technology fields. Their pages show what is in production, in research or still to confirm.