jev on cloudflare workers ai

Cloudflare Is a Serving Route, Not a Different Jev Primitive System

Cloudflare’s model page describes the same state and Noul, Choice and Score question shapes, but wraps them in Workers AI’s runtime and HTTP conventions. The model ID is typesafe/jev, not TypeSafe’s direct jev-1.13.0 identifier. Provider model names and resolved upstream versions are separate fields that should not be conflated in logs.

The listing calls Jev a third-party model and points to TypeSafe’s legal and model pages. Review both providers’ current terms and the Cloudflare dashboard price before production; TypeSafe’s direct price and rate limits do not automatically apply.

Choose the Workers Binding or REST Boundary Deliberately

Both Cloudflare paths reach the same listed typesafe/jev route, but they expose different credential and deployment boundaries. Keep the semantic request in provider-neutral application types and map it at one adapter. This makes it possible to compare or replace routes without rewriting evaluation logic.

Workers Binding Example

This request form follows Cloudflare’s published model page as checked September 22, 2026. Bind Workers AI to env.AI through the project’s Cloudflare configuration. Do not place the direct TypeSafe key in the Worker merely to use this binding; Cloudflare owns authentication and account configuration for its route.

export default {
  async fetch(_request, env) {
    const response = await env.AI.run("typesafe/jev", {
      state: "Help! My payouts have been failing for 3 days.",
      questions: {
        is_urgent: {
          type: "noul",
          instructions: "Does this convey urgency?",
          criteria: {
            true: "Explicitly time-sensitive",
            false: "No urgency expressed",
          },
        },
      },
    });

    return Response.json(response);
  },
};

Cloudflare HTTP Envelope Differs from TypeSafe Direct HTTP

The direct TypeSafe endpoint accepts state, model and questions at the top level; Cloudflare’s generic run endpoint places Jev’s state and questions inside input. Copying one payload into the other is not portable. Validate Cloudflare’s response envelope against its published schema before reading the nested Jev result.

curl "https://api.cloudflare.com/client/v4/accounts/$CLOUDFLARE_ACCOUNT_ID/ai/run" \
  --header "Authorization: Bearer $CLOUDFLARE_API_TOKEN" \
  --header "Content-Type: application/json" \
  --data '{
    "model": "typesafe/jev",
    "input": {
      "state": "Help! My payouts have been failing for 3 days.",
      "questions": {
        "is_urgent": {
          "type": "noul",
          "instructions": "Does this convey urgency?"
        }
      }
    }
      }'

Normalize the Transport Without Flattening Jev Semantics

The types and helper are an application architecture sketch; only the env.AI.run("typesafe/jev", { state, questions }) invocation is the documented Cloudflare surface. Build the validator from the current provider schema and preserve the native response for debugging under the project’s data policy.

Do not normalize a missing or malformed answer into false, an empty Choice or score zero. Those are plausible semantic values and would hide a provider failure as a model decision. Return a separate no-result status to the application.

type JevRequest = {
  state: string | Record<string, unknown>;
  questions: Record<string, unknown>;
};

type ProviderResult = {
  provider: "cloudflare";
  routeId: "typesafe/jev";
  requestedAt: string;
  latencyMs: number;
  attempts: number;
  nativeResponse: unknown;
  answers: Record<string, unknown>;
};

async function runThroughCloudflare(ai: Ai, request: JevRequest) {
  const started = performance.now();
  const nativeResponse = await ai.run("typesafe/jev", request);
  return normalizeAndValidate(nativeResponse, performance.now() - started);
}

Re-Evaluate Provider-Specific Behavior

Provider parity is a hypothesis. Run the same labeled fixtures through direct TypeSafe and Cloudflare if portability matters, compare full distributions, and measure latency from the actual Worker region. A schema match does not establish identical operational behavior.

Record Both Route Identity and Model Identity

This is an application log schema, not a Cloudflare response. Cloudflare’s route ID identifies the provider surface; it should not be rewritten as proof that a particular TypeSafe version ran unless the response or current provider documentation exposes that version. Thresholds calibrated on a known direct model may need a fresh provider-specific study.

{
  "provider": "cloudflare-workers-ai",
  "provider_route": "typesafe/jev",
  "upstream_model": "<record only when exposed>",
  "question_version": "urgency-v3",
  "projection_version": "ticket-state-v2",
  "response_status": "scored",
  "attempts": 1,
  "latency_ms": 184,
  "native_response_ref": "secure://eval/9f2"
}

Production Architecture

An adapter makes provider differences explicit and lets tests replay the same application request against multiple transports. Keep fallback behavior in the application, not hidden inside provider switching. See the broader Jev platform comparison and HTTP contract.

Treat Cloudflare Failures as Provider Outcomes

Provider failure and model uncertainty have different owners and remediation. A timeout says no result arrived; it does not say the condition is false. Bound the complete request deadline, make retries visible and include their cost and latency. The errors and retries guide supplies the general failure budget.

Run a Provider Parity Study Before Switching Traffic

A successful request proves connectivity, not behavioral parity. Compare at the action boundary: how many cases move from allow to review or review to decline, and which labeled errors change? The regression-testing guide defines the fixture manifest and calibration defines probability comparison.

1. Freeze Choice, Score and Noul fixtures with independent labels and full expected answer sets.
2. Send equivalent semantic state and questions through direct TypeSafe and Cloudflare adapters.
3. Record route IDs, any resolved model metadata, distributions, errors, latency and usage.
4. Compare paired answers, probability movement, calibration and action-band changes.
5. Load-test representative state lengths and concurrency within current provider limits.
6. Fit route-specific thresholds if required; canary and retain the previous route as rollback.

Verification Checklist

Cloudflare publishes separate setup, pricing and platform-limit pages. Capture their review dates in the deployment record because catalog presence, pricing and account limits can change independently. Re-run the checklist after a route, Worker runtime or model update.

• Confirm typesafe/jev still appears in Cloudflare’s model catalog and inspect the schema links.
• Check dashboard pricing, account quotas and regional availability.
• Run a credentialed smoke test and store a redacted response fixture.
• Replay labeled semantic tests and compare with the previously approved provider.
• Inject timeout, quota and malformed-response paths before enabling an action.