Albatross
An agent runtime can ask Jev to answer narrow error-classification questions before using the main model.
Added to Jevfast
How it uses Jev
In active or shadow Jev mode, Albatross sends only the current eligible plain-text request. Jev first routes it to retryability, error category, or the main LLM, then answers the matching classification. The direct answer is used only when route and answer meet both configured confidence and selected-choice probability thresholds; shadow mode records what it would answer but uses the main LLM.
What Jev decides
Illustrative direct-answer routing and error classificationExample answers · not a recorded Jev response · Source ↗
Question 1 · route
YOUR APP
INSTRUCTION
Choose a direct handler only for one self-contained English question about an included error that can be fully answered by a fixed classification. Otherwise choose main_llm.
STATE
A single self-contained text request, represented as {"request": text}.
JEV · CHOICE
- retryability
- error_category
- main_llm
Question 2 · retryability
YOUR APP
INSTRUCTION
Classify the supplied error as transient, persistent, or uncertain. This is not permission to retry.
STATE
A single self-contained text request, represented as {"request": text}.
JEV · CHOICE
- transient
- persistent
- uncertain
Question 3 · error category
YOUR APP
INSTRUCTION
Classify the supplied error; use uncertain for missing, conflicting, or insufficient evidence.
STATE
A single self-contained text request, represented as {"request": text}.
JEV · CHOICE
- authentication
- permission
- rate_limit
- network
- input
- server
- uncertain
App workflow
- Check eligibility
Skip Jev for disabled mode, multipart or empty requests, oversized text, or unavailable credentials.
- Route and classify
Send the current request to Jev. Require a confident direct-handler route and a confident non-uncertain classification.
- Answer or fall back
Active mode returns a fixed classification string. Shadow mode records a possible direct answer but continues to the main LLM. Other requests also go to the main LLM.
Why it is interesting
Jev can answer a narrow, fixed-output error question before the main model runs, while an explicit router sends ambiguous or broader requests to the main LLM.