SKILL QA / SEMANTIC EVALUATION
Skill Loop.
Real package snapshot · staged three-file revision · baseline and candidate reviewed by typesafe/jev
◆
CHECKINGCloudflare Workers AI
Human approval boundary. Jev adds semantic evidence; it cannot approve, install, or modify a skill.
SYSTEM MAP
Installed skill evaluation & staged repair
Select a node to inspect its boundary.
01
CONCEPTUAL
The complete Skill Loop
Every candidate stays isolated until a human approves it.
→→→→→
02
EXECUTED
Real package revision + live Jev review
Seven package files, three staged edits, three paired comparison cases.
→→→→→
RECORDED EVENT 1 / 12 · INSTALLED PACKAGE FINGERPRINTED
⊘ Jev supplies advisory semantic evidence. Package facts and approval remain separate.
RECORDED REPLAY
Evidence timeline
Installed package fingerprinted.
TEST CASES
Cases and results
Select a case
JEV-GATEWAY · /v1/select-skill · RUBRIC STAGE: SHADOW
Picking one skill instead of loading all of them
Task + manifests67 installed skills, name and description only
→
Policy gate0 candidates → error · 1 → returned, no Jev call
→
Jev · noul per skillyes/no per skill · batches of 20 → 4 calls via AI Gateway · 24h KV cache
→
Thresholds≥0.75 and not tied → select · ≥0.40 → load top 3 · else no match, load all
→
Load one bodyprivileged skills keep requires_gate · every call written to D1
chosen skillother candidates0.75 select threshold
REAL PACKAGE SHAPE
Loading package…
The installed package is fingerprinted; every proposed edit lives in an isolated working copy.
BASELINE —CANDIDATE —
PACKAGE MANIFEST
Instructions, references, script & assets
STAGED DIFF
What changed
JEV COMPARISON CASES
Score movement with evidence
SELECTED COMPARISON
Loading comparisons…
The score is the pass probability for the atomic requirement question. It is advisory and never changes exact checks or approval eligibility.