jev-deferred-crispification
Position paper arguing per-hop calibration does not compose in decision pipelines and proposing deferred crispification.
Overview
This repository is a position paper on what Jev and other System One models are missing. It argues that the typed Choice, Score and Noul interface lacks two primitives: a hidden-Markov belief over latent regimes and fuzzy membership for borderline predicates. Two lemmas state that calibration does not compose across pipeline hops and that thresholding at each hop causes cliff cascades.
It proposes a design law, deferred crispification (keep structured uncertainty through the pipeline and collapse to an action only once), a reference architecture called BSF-S1, and two weight-free audit metrics, TCE and AMS. A key claim is that per-hop ECE is permutation-invariant and so cannot detect clustered regime-shift errors that harm trajectories. Everything is said to reproduce on a laptop CPU in about fifteen minutes.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Other | Does per-hop calibration guarantee reliable multi-step decision trajectories?(inferred) | — |
How it works
- State the missing primitives and failure modes
- Prove two lemmas on composition and cascades
- Propose deferred crispification and the BSF-S1 architecture
- Provide TCE and AMS audit metrics with CPU-reproducible scripts
Strengths
- Highlights a blind spot of per-hop ECE dashboards
- Audit metrics need no model weights
- Reproducible on a laptop CPU
Caveats
- Theoretical position paper; not peer reviewed per the sources
- Does not provide a drop-in replacement for Jev
At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- CPU scripts
- Has code
- yes
Use cases and integrations
How curated lists describe it
Position paper: the Hidden-Markov and fuzzy primitives missing from TypeSafe…
wh000wh000/awesome-jev-livePosition paper: the Hidden-Markov and fuzzy primitives missing from TypeSafe AI's Jev and System-One decision models. Two lemmas, one principle (Deferred Crispification), one architecture (BSF-S1).
hellogumbo/awesome-jev
Frequently asked questions
What is the main claim?
Calibrated per-step decisions can still produce unreliable trajectories, and ECE cannot detect this.
What is deferred crispification?
Keeping structured uncertainty through a pipeline and converting to a hard action only at the end.
Can I run the metrics?
Yes, the README says everything reproduces on a laptop CPU in about fifteen minutes.
Related projects
- cbjev★ 4
Self-hosted encoder model giving Jev-compatible choice, score and noul answers in one pass, pitched as a better-calibrated Laya successor.
ProjectresearchPython Joke decision model trained to answer wrongly with total confidence: 1.9% accuracy at 0.957 mean confidence, via a Jev-style API.
ProjectresearchPythonPre-registered field test of Jev, GLiNER2.5-Decide, Laya and other decision models against small LLMs on fresh September 2026 data.
ProjectresearchPython11 MB CPU-only model that answers choice, noul and score questions with fitted calibration and no LLM at inference.
ProjectresearchPythonTiny pre-registered experiment testing whether a confidence-scoring reward (RLCD-like) yields confidences safe to automate on at 0.9.
ProjectresearchPythonResearch code for an image-and-text System One model that reads option-letter logits from a multimodal LLM, with training and calibration.
ProjectresearchPython