Skip to content

Jev Deferred Crispification

Preprint plus reproducible experiments arguing that per-hop calibration does not compose across a decision pipeline: two lemmas, a Deferred Crispification design law, a BSF-S1 reference architecture, and two weight-free audit metrics.

Stars
0
Language
TeX
Published by
Community
Author
dnakhoa
Last verified
Oct 2026

About this project

Your decision pipeline is calibrated per hop and blind per trajectory. This repo contains the math that proves it, the architecture that fixes it, and two audit metrics you can run today — no model weights required. Per-hop Expected Calibration Error is permutation-invariant: shuffle your decision stream and ECE does not move. But regime-shift errors are clustered — their cluster length is the latent regime's sojourn time. Therefore:

Quoted from the project's README. Read the full README.

Highlights

  • Per-hop ECE is permutation-invariant, which is the argument that it cannot see clustered regime-shift errors.
  • TCE compares implied against observed errors per window with a shuffled-stream control, while AMS measures action mass moved by a perturbation.
  • Six experiments plus lemma checks run on CPU in about fifteen minutes through `python3 experiments/run_all.py`.
  • A memory ablation is reported and lands on the memoryless head, so the gain is attributed to the transition matrix alone.

Watch out

This is a paper repository with simulation experiments, not an evaluation of Jev. The claims concern what happens when typed decision heads are composed, and the reported tables compare a memoryless typed head, a Bayes-optimal memoryless head, the learned BSF-S1 and an exact filter on synthetic regime-shift setups. The experiments are weight-free and CPU-only, so nothing here measures a real model. The integrity changelog deserves as much attention as the results: an earlier cascade headline was withdrawn because the comparison used unmatched base rates, and a later ablation bug was caught by a referee pass that had made the ablation a disguised prior correction. One reported row also cuts against the thesis, since a thresholded mean matches the single-collapse flip mass. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

24GitHub stars
Unofficial laptop study of parallel constrained decoding: prefill once, share the KV cache across fields, and decide all schema fields in one batched forward pass without writing JSON.
Jev Practices & Patterns#community#jevbest#catalog-import
23GitHub stars
Live race rig that labels 1,000 real app reviews with Jev typed questions against Gemini JSON output on identical batches, then replays recordings and re-ranks rows locally without extra model calls.
Jev Practices & Patterns#community#jevbest#catalog-import
22GitHub stars
Offline evaluator that checks whether classifier confidences match reality and converts mistake costs into a deployable threshold, emitting an HTML report plus machine-readable threshold files.
Jev Practices & Patterns#community#jevbest#catalog-import
16GitHub stars
Vendor-neutral harness that tests whether Jev reranking improves a small RAG stack, using frozen hybrid BM25 plus BGE-M3 candidates and isolating retrieval, rerank, and generation stages.
Jev Practices & Patterns#community#jevbest#catalog-import
37GitHub stars
Minimal Python recipe that turns any causal LLM into a Jev-style classifier by prefilling choice_index and constraining sampling to legal indexes in one forward pass.
Jev Practices & Patterns#community#jevbest#catalog-import
39GitHub stars
Open replica of the typed decision API on openly available base models, answering choice, score, and true-probability questions in one forward pass with no decoding.
Jev Practices & Patterns#community#jevbest#catalog-import

From the Jev guides

Original write-ups on what this entry is for.

Most Jev accuracy problems are state problems. What belongs in a state, what is a pre-made judgment to leave out, and why naming the slice you ask about matters more than the wording.
Jev Practices & Patterns#state#design#primitives
Read article
A Noul returns one number and no separate confidence. How to read the value, pin its boundary with criteria, split compound questions, and set a per-action threshold that matches the cost of being wrong.
Jev Practices & Patterns#noul#primitives#confidence
Read article
A Score is a weighted position, not a ruler. Why the number can land between levels, why levels are judged separately, and how to combine several atomic scores in code instead of one crowded rubric.
Jev Practices & Patterns#score#primitives#scoring
Read article
Back to all resources

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai