MetaCog
A model-agnostic metacognition loop that generates candidate reasoning paths and uses a System One judge such as Jev to select or score them.
Added to Jevfast
How it uses Jev
In best-of-N mode, the thinker generates multiple candidate solutions. Jev sees the problem and labeled paths in one Choice request, then MetaCog selects the highest-probability candidate. Other controller modes score each candidate independently with a Noul question so the judge does not compare paths directly.
What Jev decides
Illustrative candidate-path selectionExample answers · not a recorded Jev response · Source ↗
Question 1 · best path
YOUR APP
INSTRUCTION
Compare candidate paths for the same problem and select the best path.
STATE
Representative two-candidate request: one problem and candidate paths labeled A and B. The implementation supports up to 26 labels.
JEV · CHOICE
- A
- B
Illustrative isolated correctness scoreExample answers · not a recorded Jev response · Source ↗
Question 1 · is correct
YOUR APP
INSTRUCTION
Judge whether the candidate path final answer is correct for the stated problem.
STATE
A problem and one truncated candidate path, sent in a separate request for each candidate.
JEV · NOUL
YesNo
App workflow
- Generate candidates
The thinker produces candidate reasoning paths for the supplied problem.
- Judge candidates
MetaCog asks Jev to compare paths in best-of-N mode or score paths independently in a scoring mode.
- Commit a candidate
The controller chooses the highest judged path and returns its answer.
Why it is interesting
It separates candidate generation from candidate judging and supports both comparative selection and blind per-candidate correctness scoring.