Typesafe system one models and new /v1/decisions endpoint

Bifrost now supports TypeSafe's jev judgment models behind a unified POST /v1/decisions endpoint and a 1:1 /typesafe drop-in for TypeSafe's official SDKs. This is available in v2.2.2 for both OSS and Enterprise.

TypeSafe's jev family takes a piece of state - a support ticket, a model's answer, a review, any string or JSON blob - and returns calibrated judgments: a probability, a pick from your options, or a rubric score, each with confidence and a full probability distribution. That makes them a natural fit for the decision points scattered through every AI product: content moderation, ticket routing, eval-in-the-loop grading of LLM outputs, churn and escalation triage, guardrail checks. Anywhere you force a chat model to answer "just say yes or no", a judgment model does it faster, cheaper, and with a number you can actually threshold on.

TypeSafe support, two ways in

Bifrost now supports TypeSafe's System One models (jev-1.13.0, jev-latest, jev-preview) as a first-class provider, with a new operation type built for judgment workloads.

POST /v1/decisions is the unified Bifrost format. Send state plus a map of named questions - noul for probabilities, choice for classifications, score for rubrics - and get one typed answer per question, with confidence, probabilities, and legend metadata normalized into a single value field. Model routing, virtual keys, governance, logging, and cost tracking all apply, and every request is strictly validated before it costs you anything.

The /typesafe integration is the 1:1 drop-in. If you already use TypeSafe's Python or JavaScript SDK, point TYPESAFE_BASE_URL at your Bifrost deployment's /typesafe prefix and you are done - one environment variable, zero code changes. Native request and response shapes are preserved byte-for-byte on the success path, and Bifrost adds what upstream does not have, including a models endpoint and automatic retries on rate limits and overloads.

LLM fallbacks

If jev is down or for some reason API calls fail - we have added support for LLM fallbacks on both /v1/decisions and /typesafe/v1/systemone endpoint.

How it maps to a normal fallback call

A decision request already carries a fallbacks list, exactly like chat and responses requests. The trick is what happens when a fallback provider has no native decision endpoint (only typesafe does today). Instead of returning "unsupported", Bifrost emulates the decision through that provider's Responses API:

1. The decision dispatch calls provider.Decision(...).
2. If the provider returns unsupported_operation (every non-typesafe provider), one shim at the dispatch layer takes over - no per-provider code.
3. The shim rebuilds the question set as a single forced emit_decision function tool (tool_choice: "required"), sends it via provider.Responses, and maps the tool-call arguments back to the exact DecisionResponse shape - values, per-answer confidence, choice probabilities, and score legends.

Because emulation runs through the same fallback loop, both the primary path (name an LLM as the decision model) and the fallback path (LLM after jev fails) flow through that one shim.

A decision request already carries a fallbacks list, exactly like chat and responses requests. The trick is what happens when a fallback provider has no native decision endpoint (only typesafe does today). Instead of returning "unsupported", Bifrost emulates the decision through that provider's Responses API:

1. The decision dispatch calls provider.Decision(...).
2. If the provider returns unsupported_operation (every non-typesafe provider), one shim at the dispatch layer takes over - no per-provider code.
3. The shim rebuilds the question set as a single forced emit_decision function tool (tool_choice: "required"), sends it via provider.Responses, and maps the tool-call arguments back to the exact DecisionResponse shape - values, per-answer confidence, choice probabilities, and score legends.

Because emulation runs through the same fallback loop, both the primary path (name an LLM as the decision model) and the fallback path (LLM after jev fails) flow through that one shim.

We did a 30 question benchmarking keeping jev as ground truth to ensure it doesn't impact the quality of the output. We have picked a few models in random order as an example : overall testsuite runs across 100+ models.

If jev is reachable, jev answers. If not, gpt-5.6-luna emulates the same decision and returns the identical response shape - the caller sees no difference beyond the resolved model name.

Client request (jev primary, gpt-5.6-luna as fallback)

jev is down, so the request falls to openai/gpt-5.6-luna

Bifrost catches the failure, moves to the fallback, and since openai has no native decision endpoint it emulates the decision through the provider's Responses API. This is the call it actually makes to gpt-5.6-luna:

gpt-5.6-luna is forced to call emit_decision, and Bifrost maps its tool-call arguments back to the standard decision shape.

Response the caller receives (identical shape to jev)

Benchmarking

Here are the results

Combination	Score	Latency (ms)	Tokens	Status
typesafe/jev-1.13.0	30/30	920	2144	REF
openai/gpt-5.6-luna	30/30	6763	3534	PASS
openai/gpt-5.6-terra	30/30	11452	3596	PASS
openai/gpt-5	30/30	25795	5901	PASS
openai/gpt-5-mini	29/30	36482	6487	PASS
openai/gpt-4.1	30/30	3963	3336	PASS
openai/gpt-4.1-mini	29/30	13113	3365	PASS
openai/gpt-4o	30/30	7630	3470	PASS
openai/gpt-4o-mini	30/30	5852	3292	PASS
anthropic/claude-opus-5	30/30	15667	8203	PASS
anthropic/claude-sonnet-5	30/30	11768	8172	PASS
anthropic/claude-fable-5-1	30/30	17789	8095	PASS
anthropic/claude-haiku-4-5-20251001	30/30	7842	6784	PASS
anthropic/claude-sonnet-4-5	30/30	16429	6784	PASS
xai/grok-4.5	29/30	24335	6618	PASS
xai/grok-4	30/30	11472	5958	PASS
xai/grok-3	30/30	13600	5976	PASS
xai/grok-3-mini	29/30	12291	6147	PASS
groq/openai/gpt-oss-120b	30/30	4208	5194	PASS
deepseek/deepseek-v4-pro	30/30	10299	5841	PASS
cohere/command-a-03-2025	30/30	9047	6487	PASS
vertex/gemini-2.5-flash	30/30	11127	1513	PASS
vertex/claude-sonnet-4-6	30/30	16266	6477	PASS
vertex/claude-opus-4-8	30/30	16091	8207	PASS
gemini/gemini-2.5-flash	30/30	9078	5307	PASS
gemini/gemini-2.5-pro	30/30	14811	5505	PASS

Try it

• Decisions quickstart - curl to first judgment in a minute
• API reference - full /v1/decisions and /typesafe/* schemas
• TypeSafe provider guide - complete field mapping to the native API
• TypeSafe - the jev models and System One