verification workspace for AI research

Turn AI research into verifier-scoped, reusable state.

KeyAI turns model-generated research into evidence-linked, verifier-scoped assets. It keeps claims, evidence, decisions, verifier outcomes, and provenance in one inspectable state.

Bring a research workflow Open the reference workspace

Reference deployment The repository demonstrates the full research-state loop on one difficult domain, but it is not yet a self-serve hosted product.

Live decision graph 17 routes evaluated / 0 selected RS-2026-07-22-001

Live reference: secp256k1 ECDLP The environment maps and verifies the research boundary. It does not solve the plain secp256k1 discrete logarithm problem.

296

verified ledger rows

~257

distinct results

routes evaluated

routes selected

The missing layer

AI can propose. KeyAI keeps the research state.

Individual agents already write proofs and code. A long research program needs a durable answer to a different question: what should the next agent trust, challenge, or stop doing?

Before execution

Bind each task to a source, exact scope, route, and falsifiable exit condition.

At verification

Record what the declared verifier accepted and what remains a semantic or empirical assumption.

After the attempt

Retain accepted results, negative evidence, stop conditions, and a reproducible rollback path.

Product loop

One state from source material to a governed result.

The current repository implements this loop through machine-readable contracts. The next product step is to make the same loop configurable for an external team.

Ingest

Pin the corpus, sources, target, and verifier contract.

Structure

Convert material into claims, dependencies, barriers, and threat models.

Decide

Select, park, or reject routes under explicit evidence gates.

Execute

Give a human or model one bounded task with a falsifiable exit condition.

Verify

Run the declared verifier and an independent result validator where needed.

Retain

Promote accepted results and preserve negative evidence, provenance, and rollback.

Active validation / TASK-011

The next result must come from another team.

We are recruiting one formal-research team to test the current workspace, map one repeated workflow, and make an evidence-based build, change, stop, or pending decision.

Status: recruiting
Session: 60 minutes
Completed discovery: 0 sessions

Read the pilot contract Apply on GitHub

Reference deployment

A difficult research boundary, represented honestly.

secp256k1 is the test case, not the product claim. It forces KeyAI to distinguish a theorem, an experiment, a threat model, a failed route, and a practical attack.

Current route decision

RS-2026-07-22-001 · 2026-07-22

Monitoring

SELECT_NONE

No current route clears the proposal gate.

No audited route currently satisfies every proposal-level requirement for a new experiment against the primary plain single-target secp256k1 objective.

The generic lower bound and generic algorithms are guardrails or baselines, not non-generic attack mechanisms.
GLV supplies a verified constant-factor structure; Pohlig-Hellman, low-degree pairing transfer, anomalous lifting, and extension-field descent fail target-specific applicability screens.
The open prime-field algebraic, GLV-Semaev, Petit-style, EDS/division-polynomial, and transfer directions do not yet provide an exact nonredundant mechanism plus a justified subgeneric cost bridge.

Inspect all 17 route dispositions

What exists now

Evidence, not a product demo made of placeholders.

Each capability below links to a live artifact in the reference repository.

Kernel-checked result ledger

Inspectable in the reference repository and checked by the repository gates.

VERIFIED.md

Evidence-gated route decisions

Inspectable in the reference repository and checked by the repository gates.

repo/ECDLP_DECISION_SUBSTRATE.json

Reproducible candidate and independent validation contract

Inspectable in the reference repository and checked by the repository gates.

experiments/framework/candidate_run.schema.json

Tasks, hypotheses, graph, provenance, and generated views

Inspectable in the reference repository and checked by the repository gates.

tasks/NEXT.md

Current capability

Reference system

Kernel-checked result ledger
Evidence-gated route decisions
Reproducible candidate and independent validation contract
Tasks, hypotheses, graph, provenance, and generated views

Not yet

Hosted product

Self-serve repository or corpus import
Hosted multi-project workspaces
Authentication, collaboration, and organization controls
A verifier adapter beyond the current repository contracts
Validated external users, retention, or willingness to pay

The next product milestone

We will call it an MVP when another team can run the loop.

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 does not establish a repeatable buyer or willingness to pay.

orientation-time

A new collaborator identifies current state, blockers, and next action in 10 minutes or less.

provenance-completeness

Every promoted result links to its source, task, verifier result, and trust boundary.

state-drift

Zero stale generated or public artifacts after a canonical state change.

external-pilot

At least one external team completes the core loop and returns for a second session.