jev-xray
Decision-forensics library that explains a Jev answer by ablating parts of the input and measuring how far the probability moves.
Overview
jev-xray provides decision forensics for System One models. Jev returns a typed decision with a calibrated probability but, by design, no explanation, so a wrong or surprising answer like 0.62 offers nothing to inspect. TypeSafe's own jaggedness notes say accuracy falls and attribution gets harder as state grows with unrelated detail. jev-xray measures the why instead of asking for it: it ablates segments of the state, re-asks the identical question, and reports how far the answer moved.
A single explain call returns the baseline value, the top segments that moved the answer with signed effects, an interaction residual showing whether effects are additive, and a cost ledger; the README's example used 7 requests, 592 input tokens and about $0.000025. The author argues occlusion attribution, normally too expensive on frontier models, becomes practical at Jev's prices and latency because ablations run concurrently, and that a fixed answer axis and calibrated probabilities make the deltas trustworthy. A published report shows a real run, and tests run in CI.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Example: The customer is asking for a refund (re-asked under each ablation) | — |
How it works
- Ask the original question for a baseline
- Ablate each segment of the state and re-ask concurrently
- Measure signed changes and interaction residual
- Report top contributing segments with a cost ledger
Strengths
- Explains decisions without a text channel
- Explanation costs a fraction of a cent per the README
- Works for Noul and other typed questions
Caveats
- Attribution depends on how state is segmented
- Ablation measures sensitivity, not true causal reasoning
At a glance
- Type
- Project
- Status
- beta
- Language
- Python
- Runtime
- Library
- Has code
- yes
- Has tests
- yes
- Stars
- 0
Use cases and integrations
How curated lists describe it
Decision forensics for System One models. Jev returns a calibrated probability and cannot explain it — jev-xray measures which part of the input caused it, by ablation.
wh000wh000/awesome-jev-live
Frequently asked questions
How does it explain a Jev answer?
By removing parts of the input, re-asking the same question and measuring how much the probability changes.
How much does an explanation cost?
The README's example used 7 requests and about $0.000025.
Why does this work better on Jev than on generative models?
The answer axis is fixed and probabilities are calibrated, so changes are directly comparable.
Related projects
Nautilus Compass agent memory layer whose jev-trust middleware logs Jev decisions and measures calibration in your domain.
ServicebetaPythonAdapter that turns existing LLMs into Jev-style decision functions with option scores, optional calibration and review thresholds.
ProjectprototypePythonDecision sidecar and Copilot CLI pre-turn hook that mimics Jev choice, noul and score using Azure Foundry Cohere Rerank.
HookprototypePythonExact-target audit finding Jev's Choice probabilities are a step function on stated base rates while per-option Noul tracks them closely.
ProjectresearchPythonResearch replicating Jev-style typed decisions with open LLMs via logprobs; Qwen3.5-9B reached 94% with good calibration in ~1 s.
ProjectresearchPythonRunnable yardstick that tests Jev's three headline claims on a small local model: typed output, no hallucination, calibration.
ProjectresearchPython