Build your evidence base
Register candidate routes, import held-out articles, and review their outputs. Names and markup are checked automatically. Meaning is reviewed by a person.
TOKENS.IO / CONTROL PLANE
No live model callsConnect your workspace
Better outcomes.
Less wasted inference.
Choose routes by measured quality and total cost per acceptable task.
Keys stay in this tab’s memory. No browser storage or third-party analytics.
ROUTING INTELLIGENCEQuality is the constraint.
Quality is the constraint.
Cost is what we optimize.
A shared decision layer for AI workloads. Every recommendation starts with evidence.
One task. A defensible choice.
- QualifyLanguage, format, privacy, capacity
- MeasureMatched articles, reviewed outcomes
- OptimizeTotal cost per acceptable article
- AbstainWhen evidence does not meet the bar
SHADOW RECOMMENDATION
No publishing authorityAsk the evidence
No decision requested. The system will not select a model without sufficient evidence.
Evaluation lab
One article is one sample. Keep source IDs, product names, links, and meaning intact.
1. Add an article contract
2. Record a candidate result
Imports do not make a model call. Include all generation, repair, validation, and failed-attempt costs.
3. Review outcomes
Review text is displayed safely as text, never executed as HTML. Reviewers must assess omissions, language, meaning, quantities, and clinical distinctions.
Model routes
A model is not globally “best.” Every route has a fixed provider, model revision, and capability contract.
Register a candidate
Decision ledger
Trace changes, approvals, costs, and the reasons behind each recommendation.
Connect a workload.
Keep control of execution.
Evaluation integration
Your application retains control of its queue, retries and downstream actions. This release accepts evaluation copies and returns evidence-based recommendations. Live provider routing is not connected yet.
- Submit an immutable source article to
POST /v1/cases. - Reserve a trial using
POST /v1/jobsand anIdempotency-Key. - Import its result, then record bilingual review.
- Request
POST /v1/decisions. An abstention is an intentional outcome.
Provider connections come next
No model credentials are requested or stored in this build. No live inference or production publication is enabled.
All-in costs use integer millionths of a US dollar. Unknown outcomes hold their reservations until explicitly reconciled; the system never silently retries them.
Private workspace access
Admin keys register routes and import results. Reviewer keys submit semantic reviews. Client keys submit cases and request recommendations. Each key is scoped to one tenant.
Pilot deployment requires HTTPS, server-configured keys, restricted ingress, persistent storage, backups, and a restore check. This is not yet a public multi-tenant billing service.