NewSkylos Jev Review
An opt-in Jev review of ambiguous dead-code findings in the Skylos scanner, with uncertain results left visible.
Added to Jevfast
NewHow it uses Jev
For a static dead-code candidate, Skylos builds state from a bounded, complete snapshot of eligible repository files and asks a Choice whether the target is retained, unreferenced, or unsupported by the available evidence. A confident unreferenced answer can preserve the static finding or skip the broad LLM check, depending on mode; a confident retained answer can suppress it in judge mode. Uncertain or unavailable answers fall back to the LLM path when enabled; Jev does not authorize fixes.
What Jev decides
Illustrative Jev decisionExample answers · not a recorded Jev response · Source ↗
Question 1 · q0000
YOUR APP
INSTRUCTION
For this finding’s target kind, file, and symbol, determine whether a concrete path in the supplied repository retains it.
STATE
Bounded, complete repository_files snapshot and fixture_scope describing the supplied source.
JEV · CHOICE
- retained
- unreferenced
- insufficient_evidence
App workflow
- Select static candidates
Choose findings for dead-code verification.
- Build the evidence state
Read the bounded complete project snapshot and construct one typed Choice question per valid candidate.
- Validate and classify answers
Check response shape and compute status using the selected choice probability and confidence threshold.
- Route findings
Skip broad verification for confident unreferenced candidates; in judge mode, suppress confident retained candidates.
- Continue verification
Send uncertain or unavailable candidates to the general LLM verifier unless Jev-only mode is selected.
Why it is interesting
The system treats abstention and source-snapshot completeness as part of the decision contract, and uses separate confidence thresholds for routing and final judge mode. It connects model output to review outcomes while keeping provenance distinct from an LLM verdict.