jev-architecture-research
Black-box reverse-engineering archive studying how the Jev decision model behaves, with evidence matrices and calibration analysis.
Overview
Jev Architecture Research is an evidence-first archive of experiments that study Jev as a black-box decision system. It does not claim access to Jev's weights, training data, source or internal traces, and it is not a leaderboard. Every conclusion is marked as observed, inferred or speculative, synthetic labels are not treated as production accuracy, and API limits are kept as limits rather than turned into model claims.
Its current behavioural attribution is a shared state and option representation with dynamic semantic option encoding, global semantic competition between options, typed choice, score and noul heads, a ranking-like decision layer, and batch-aware inference. A final 720-call controlled experiment favoured global joint semantic scoring over cluster-first filtering, while provider batching, top-k reduction, reranking, the backbone and training objective remain unresolved. The archive includes a master report, evidence matrix, hypothesis tracker, failure-surface notes on negation, contradiction, collisions and language, latency and calibration analyses, a machine-readable knowledge base and experiment receipts. Code is MIT and research notes CC BY 4.0.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Other | How does Jev internally compare options and produce typed answers? | — |
How it works
- Design controlled black-box experiments
- Run calls and save receipts
- Record claims with evidence and caveats
- Update hypotheses and attribution
Strengths
- Careful separation of observed, inferred and speculative claims
- Documents failure surfaces and calibration behaviour
- Machine-readable knowledge base
Caveats
- Black-box inference cannot establish internal architecture
- Several key questions remain unresolved
At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- research scripts
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Black-box reverse engineering research archive for the Jev decision model
logicrw/awesome-jev-projects
Frequently asked questions
Does this reveal Jev's architecture?
It offers a behavioural attribution from black-box experiments, not internal access.
What failure modes are documented?
Negation, contradiction, option collisions and language effects, among others.
How are claims handled?
Each is marked observed, inferred or speculative with evidence and caveats.
Related projects
Pre-registered field test of Jev, GLiNER2.5-Decide, Laya and other decision models against small LLMs on fresh September 2026 data.
ProjectresearchPythonIndependent API-only audit of what Jev's probabilities mean: abstain removal, Korean, coherence, option order, interference and caching.
ProjectresearchPythonResearch framework (DART) testing whether Jev's probabilities are a reliable signal to answer, escalate or abstain.
ProjectresearchPythonEvidence ledger auditing Jev claims, recomputing published calibration and running an 800-item difficulty-gradient study.
AppresearchPythonRunnable yardstick that tests Jev's three headline claims on a small local model: typed output, no hallucination, calibration.
ProjectresearchPythonReproduction and stress test of a paper asking whether Jev-style probabilities stay calibrated after routing on selected examples.
ProjectresearchPython