OneForge
Build intelligence. Keep humans in control.
OneForge is the AI capability control plane — one operating surface that governs the complete path from versioned data and evaluated training to reviewable releases, production traffic and sandboxed engineering.
Three guarantees
One operating surface for the AI capability lifecycle.
Most teams ship models faster than they can govern them. OneForge closes that gap: every artifact is traceable, every release is approved, and every run is auditable.
Reproducible
Every artifact has lineage. Data versions, training runs and releases trace back to exactly what produced them.
Governed
Every release has an approver. Nothing reaches production without a designated sign-off.
Observable
Every run leaves evidence. Executions generate auditable records for support, compliance and review.
One control plane, three capability lines
OneForge doesn't just govern models. It governs a factory for each kind of capability.
Version, evaluation, approval, canary, rollback, audit and tenant isolation live in one control plane. Model, Agent and Evolution are three factories that register into it — not three separate products with three separate release paths.
Model factory
Verified for single-machine release and real cloud GPU training
Idea, dataset and training run through evaluation, checkpoint and registry into the same governed release path — shadow, canary, production, rollback.
Agent factory
A real single-repository task path is verified
Agents are managed the way models are: their own data, evaluation, version and release. OneCoder — a real coding agent — clones a repository, opens a branch, edits in an isolated sandbox, tests its own change, hashes the evidence and opens a reviewable pull request.
OneCoder: clone → branch → sandboxed edit → test → evidence hash → reviewable PR
Evolution
The loop runs end to end; no candidate has cleared its own quality gate yet
Real usage becomes a failure record, a candidate, a real training run and a real independent evaluation — gated by the same production baseline as everything else. Weak candidates get rejected, not shipped.
The lifecycle
From versioned data to governed production.
Version data
Track and version the data behind every model, with full lineage.
Evaluate training
Run and evaluate training so releases are backed by evidence, not vibes.
Review releases
Package reviewable releases that a designated approver signs off before promotion.
Manage traffic
Promote to production, manage traffic and keep sandboxed engineering environments.
Capabilities
Control built for teams that answer for their AI.
Lineage tracking
Full artifact lineage across data, training and releases.
Dual-principal approval
A second principal with separate credentials must sign off. Not a setting — an invariant: the requester can never approve their own promotion.
Audit evidence
Every run produces records you can hand to support, security or compliance.
Sandboxed engineering
Isolated environments to build and test without touching production.
Release management
Reviewable releases and controlled production traffic in one surface.
Secure workspace
Approval tokens held in memory only and cleared on reload.
Where it stands today
M2, moving toward M3 and M4.
Every milestone below can only be Verified, Implemented — pending acceptance, or Planned. Code existing is not evidence; a milestone only moves up when a specific environment produces a repeatable trace, log, hash or drill report.
Real and simulated execution stay permanently distinguishable.
VerifiedProves the governance control plane itself: real training, independent evaluation, second-principal approval, shadow, canary and rollback all traced.
VerifiedReal vLLM inference, AWS two-node T4 training and a governed cloud GPU release have passed; EFA and gVisor isolation are still open.
Partially verifiedHA, backup and budget code exists; 30-day SLO observation and drilled recovery are not yet accepted.
Implemented — pending acceptanceThe loop runs end to end and correctly rejects weak candidates — its best real score, 0.5417, still hasn't cleared its own 0.6 gate.
Implemented — pending acceptanceOneCoder's single-repository path is real; multi-team, multi-tool, 90-day production use is not yet proven.
Implemented — pending acceptanceEnterprise SSO, tenant isolation penetration testing and the cross-product flywheel.
PlannedCommercialization has not started, and is planned to follow — not precede — this internal verification. The flywheel with OneAI Labs' other products also depends on those products reaching real usage, which is outside what OneForge itself can prove.
Who it is for
For organizations that have to answer for their AI.
AI platform teams
Give model and data workflows a governed home instead of scattered scripts and notebooks.
Compliance & risk
Get lineage, approvals and audit evidence built into the lifecycle, not bolted on after.
Regulated industries
Meet human-oversight and traceability expectations for production AI.
Where it fits
The governance layer of the OneAI stack.
OneAI Core runs AI as a commercial product. OneForge governs how that AI capability is built, released and audited — so speed never comes at the cost of control.
OneAI Core
Gateway, routing, cost guards and usage — the commercial operating layer.
OneForge
Lineage, evaluated training, reviewable releases and approver sign-off.
Governed production AI
Capability that ships fast and stays reproducible, governed and observable.
Your AI should not stop at a demo.
Route intelligence through Core, govern it with TheOne, persist the work in OneMission, execute through OneClaw and keep verified outcomes in OneField.