jev claim check

Current Claim and Evidence Ledger

Different Sources Prove Different Things

Citations should preserve this level in the sentence. “TypeSafe reports 70–500 ms” is supportable; “Jev responds in 70 ms” removes both the range and attribution. “A project reports X” is not the same as “Jev achieves X.” The social evidence roundup applies this rule to launch commentary and community experiments.

Latency Claims Need Workload and Network Context

TypeSafe reports a 70–500 ms end-to-end range and says its published evaluations were generally run from West Coast laptops near the service. Preserve that attribution and context. State length, question count, provider route, region, load, retries and client work can change observed latency.

Measure p50, p95 and p99 end to end in the deployed architecture. Do not shorten a first-party range into “70 ms latency” or compare it with model-only timing from another system.

Price Facts Require a Route and Date

TypeSafe’s direct model page listed $42 per billion input tokens—$0.042 per million—and free output tokens when checked. Provider pricing may differ, and retries, gateways, storage and human review add cost. The pricing guide gives worked examples.

Rate-limit figures are explicitly dynamic. Treat 250k tokens per second and 1,200 requests per minute as dated listed limits, not contractual capacity.

Calibration Is a Training Objective, Not a Deployment Guarantee

TypeSafe describes RLCD as reinforcement learning for calibrated decisions. That supports the design objective, not the claim that every probability is calibrated for every criterion or population. Validate reliability bins, Brier or log loss and action coverage on target labels.

LangChain’s evaluator experiment is independent evidence for one configuration. The reranking post is a useful negative result for another. Together they argue for task-specific testing, not an overall winner.

Availability Claims Need Both Announcement and Interface Evidence

TypeSafe’s cited September 2026 post says Jev is generally available with no waitlist. That answers the broad access question at that date. A developer should still verify the current sign-up flow, supported region, credentials and account limits before planning a deployment.

Provider announcements need an additional check. Cloudflare’s announcement is backed by a published model page and request contract. OpenRouter announced Jev beta, but its public models catalog did not expose a TypeSafe/Jev entry when checked September 22, 2026. The accurate statement includes both observations rather than selecting the more convenient one. See the platform comparison and launch tracker.

No-Training and No-Retention Are Different Claims

The official model page states that requests and responses are not used for training. Retention duration, regional processing, abuse logs, backups and third-party routes require separate evidence. See the privacy and retention review.

Maintain the Page as a Dated Correction Record

• Link the primary source beside each claim.
• Record verification date, provider and resolved model.
• Label vendor, independent, community and anecdotal evidence.
• Preserve prior wording when a factual correction materially changes a conclusion.
• Downgrade stale claims to unknown rather than carrying them forward.
• Invite reproducible counterevidence and link the artifact.