An awesome list that knows the difference between cited and tested

What it does Jev is TypeSafe AI’s decision model: hand it state plus typed questions and it returns typed answers with calibrated confidence — it does not write text, and it is fast and cheap enough to sit in an agent’s inner loop. This repo indexes 1,207 public examples of it in use, sorted into 18 decision patterns (tool selection, safety gating, context compaction, retry control…) rather than by stars or framework. Every row names its source, with call-site citations and caveats recorded where they exist; the README is generated from catalog.json, which ships with a JSON schema and English/Chinese versions.

The interesting bit The “Measured, not claimed” section gives negative results top billing. Its self-declared most credible row is Hermes Agent’s evaluation: Jev compaction recalled worse than their shipping summariser and merely tied plain recency ordering at a matched context budget — cost was genuinely far lower, but they published the decision not to adopt. The catalog turns the same skepticism on itself: its counts describe saved records, not current link availability or passed tests.

Key highlights

• All 18 decision patterns have entries; tool selection (230) and content scoring (164) are the busy ones, recommendation has exactly one.
• Everything reduces to three primitives — choice, score, noul — and the README calls the asymmetry in the noul row the single most common source of bugs.
• 1,204 dated HTTP link records and 1,121 call-site citation records back the index.
• Standout rows include worldmonitor holding Jev in shadow mode after a blind eval found it merely tied the incumbent, and no-mistakes deleting its Jev pre-brief after two measurements showed no wall-clock gain.
• The “start here” list explicitly rejects “most starred” as a reading order.

Caveats

• Runtime and performance are not independently tested here; inclusion is a source record, not an endorsement — the repo repeats this like a mantra, correctly.
• Jev has no published weights, so nothing runs locally; input is text-only within a 64k-token-per-request context.
• The sharpest engineering caveat — reversing option order flipped a 0.9 threshold in one independent report — is unreplicated, and the catalog says to treat the magnitude as indicative.

Verdict Browse it if you are wiring Jev into an agent and want to know what has actually been measured, including where it lost. Skip it if you need official docs or independent benchmarks — this is a third-party bibliography, and the first to say so.