KeyAI · verification workspace for AI research
Today, KeyAI is a public reference deployment—not a self-serve or hosted multi-project product. It preserves hypotheses, evidence, bounded experiments, counterexamples, formal results, barriers, and provenance in one inspectable state—so the next researcher or agent can continue from the last known boundary.
Demonstrated, not implied
sorry / custom axiomsWhat KeyAI is
Research agents can generate claims, proofs, experiments, and code faster than a team can establish what is true, why a decision was made, and what should happen next.
Give a research team one inspectable state for claims, evidence, decisions, tasks, candidate outputs, verifier results, and provenance.
Models can propose proofs, experiments, and code. KeyAI preserves the research program around those attempts: what was asked, what was tried, what a verifier accepted, what failed in scope, and what should happen next.
See the research loopHow Research OS works
The canonical workflow moves from pinned inputs to a retained outcome. Open each stage to see what it contributes; the full explanation remains available without JavaScript.
Pin the corpus, sources, target, and verifier contract.
Convert material into claims, dependencies, barriers, and threat models.
Select, park, or reject routes under explicit evidence gates.
Give a human or model one bounded task with a falsifiable exit condition.
Run the declared verifier and an independent result validator where needed.
Promote accepted results and preserve negative evidence, provenance, and rollback.
Research map · generated projection
This small public map is generated from the same decision, formal, result, and engine state that drive the technical workspace. It intentionally leaves the full internal graph in the repository.
The target, inputs, output, threat models, and promotion gates are pinned before work begins.
4 are ruled out for the exact target; 4 remain open and parked. Every status is scoped.
Inspect every route and stop conditionFormal results establish their encoded statements and assumptions; they do not automatically establish an attack or a product claim.
Browse source-linked formal resultsThe consumed bounded run was independently validated on frozen toy instances. It is empirical evidence, not a secp256k1 result, route promotion, or rerun authorization.
Inspect the retained outcome8 common promotion requirements preserve the boundary between a plausible idea and an authorized research route.
Open the canonical decision contractECDLP reference deployment
secp256k1 is precise, technically demanding, and rich in formal proofs, experiments, threat-model boundaries, failed routes, and unresolved gaps. That makes it a useful test of whether research state can remain inspectable over a long-running program.
The boundary is explicit: no secp256k1 break, shortcut, or validated subgeneric route is claimed.
Exact declarations, source anchors, proof methods, and disclosed trust labels.
VERIFIED.mdRoutes retain their scope, evidence, stop conditions, and reasons to reopen.
repo/ECDLP_DECISION_SUBSTRATE.jsonSelected outputs bind pinned inputs to a separate validation path.
experiments/framework/candidate_run.schema.jsonTasks, hypotheses, outcomes, provenance, and generated views persist across sessions.
tasks/NEXT.mdCurrent stage
For researchers and AI labs
KeyAI is recruiting one formal-research team to test orientation in the current workspace, map one repeated research-state problem, and reach an explicit build, change, stop, or pending decision.
No external pilot session has been completed or recorded. Interest is not counted as adoption, retention, or product validation.
The next product milestone
A non-owner research team can connect a second project, obtain a trustworthy initial map, run one candidate through its verifier, and understand the resulting decision without editing KeyAI's generator code. A technical MVP still would not establish repeatable demand or willingness to pay.
A new collaborator identifies current state, blockers, and next action in 10 minutes or less.
Every promoted result links to its source, task, verifier result, and trust boundary.
Zero stale generated or public artifacts after a canonical state change.
At least one external team completes the core loop and returns for a second session.