Skip to content

Latest commit

 

History

History
119 lines (96 loc) · 7.19 KB

File metadata and controls

119 lines (96 loc) · 7.19 KB

Background: Jev, and the other open recreations

Moved here from the README so the front page leads with the result. Nothing below was cut.

What Jev is, and what this reproduces

Jev, launched September 15, 2026, is a closed hosted model that answers three kinds of typed question against a shared state: Choice (one option from a set), Score (a level on an ordered scale), and Noul (is a statement true). Every answer carries probabilities. TypeSafe trains it with a method it calls RLCD, Reinforcement Learning for Calibrated Decisions, and has published neither the architecture, the weights, nor a paper.

Jev's architecture is undisclosed. It is not a fixed-label classifier, because it accepts arbitrary instructions, option descriptions, and JSON state at request time. A black-box analysis by Archer Hume probes the hosted API and concludes it is a causal decoder language model post-trained for this contract: the state is shared across questions, each question is scored in its own branch that cannot see its siblings, options inside one question influence each other, and the answer is read as numbers rather than generated as text. TypeSafe has not confirmed that, but it matches every behavior we can measure from outside. The mechanism that buys the speed is the same either way: encode the input once, read the model's scores at the answer position, restrict them to the declared options, normalize. ruling implements that mechanism on a stock instruction-tuned open model. On the two fixtures the other open recreation published, ruling scores higher than it reported — 0.920 against 0.813 on authored144, 0.684 against 0.637 on WANLI (see the comparison at the end). On TypeSafe's own published evaluation rows, a 19 GB open model running through ruling answers 91 of 102 questions the way TypeSafe's reference does, against Jev's 90, with no training of any kind.

Jev (hosted) ruling
Primitives Choice, Score, Noul Choice, Score, Noul, same request and response shape
Many questions per call, one state Yes Yes. The state is encoded once; every question suffix runs in batched forward passes against replicas of that cache
Probabilities and confidence Yes Yes. Confidence uses the formulas from TypeSafe's MIT-licensed client, so the number means the same thing: for Choice, how far the leading option stands above uniform; for Score, how tightly mass sits around the modal level
Calibration RLCD training, undisclosed Temperature scaling fit on your labeled data, bound to the model and settings it was fit for
Option-order robustness Unknown Averaging over option orderings, on by default
Max options per Choice 255 255. Above the tokenizer's 62 single-token codes a Choice runs in two stages
Structured instructions and descriptions JSON accepted JSON accepted
TypeSafe's official SDK Yes Yes. Point TYPESAFE_BASE_URL at a ruling server and their unmodified client works; there is a test for it
Context ~32k tokens per question branch, ~64k per request Same two limits by default, both configurable
Weights Closed Any MLX chat model on Hugging Face
Latency, 3 questions, short state 70 to 500 ms, vendor claim 80 to 136 ms on a 4B, 141 to 204 ms on the 19 GB model, see Measurements

Not reproduced: RLCD training, Jev's parallel sampler, and frontier-level judgment. Read the Measurements section before trusting any number here.

ruling versus TheoLeeCJ/openjev

TheoLeeCJ/openjev is the other open recreation, released the same day as Jev. It asks "can a home GPU do what Jev does?" and answers with a benchmark. ruling is the thing you call from software. Both read option logits from a Qwen3.5-4B model; everything above that differs.

TheoLeeCJ/openjev ruling
Form Research CLI: JSONL of questions in, results file out Service: HTTP endpoint with Jev's request and response shape, Python and TypeScript clients, playground, one-shot CLI
Question types One: pick from up to 16 lettered options Three, matching Jev: Choice to 255 options, Score, Noul, each with probabilities and confidence
Questions per call One per row Any number against one state, one request
Shared-state reuse Special mode; every row must carry the identical state Always: state encoded once, all question suffixes in one batched forward pass
Option-order bias Measured it, 10 of 36 flips, left it Averaged over orderings; 4 of 144 flips
Calibration None Temperature scaling, bound to the model it was fit on
Hardware CUDA GPU, BF16 Apple Silicon Mac, 4-bit, no GPU box
authored144, their fixture 0.813 0.920 at 4B, 0.964 at 35B
WANLI 256, their row selection 0.637 0.684
TypeSafe's public evals, agreement with reference 0.85 0.814 at 4B, 0.892 at 35B
Still theirs Committed raw predictions, a WebGPU demo

Same rows, same metrics, and on the 4B rows ruling runs a quantized version of the same base model. Procedures and caveats are in the comparison section at the end; the three fixtures in datasets/ are theirs, converted and credited in THIRD_PARTY.md.

Comparison with TheoLeeCJ/openjev, procedures and caveats

The summary table is near the top. This section shows where each number comes from and what does not compare cleanly.

Same rows, both systems

Dataset Metric TheoLeeCJ, 4B BF16 ruling r=1, 4B 4-bit ruling r=3 (default), 4B 4-bit ruling r=3, 35B 4-bit
authored144 mean family balanced accuracy 0.813 0.884 0.920 0.964
perturbations108 mean family balanced accuracy 0.850 0.909 0.954
WANLI 256 balanced accuracy 0.637 0.684 0.672

Same rows, same metric definitions, their numbers read from their published results. ruling runs a 4-bit quantization of the same base model, which if anything favors them.

Order stability, related but not identical procedures

System Procedure Flip rate
TheoLeeCJ 36 base cases, each compared with its reversed-option twin 10 of 36, 28%
ruling r=1, 4B every row of authored144 rescored with options reversed 15 of 144, 10.4%
ruling r=3 (default), 4B same 6 of 144, 4.2%
ruling r=3 (default), 4B every row of perturbations108 rescored with options reversed 4 of 108, 3.7%
ruling r=3, 35B every row of authored144 rescored with options reversed 4 of 144, 2.8%

Both procedures ask the same question, does the argmax survive reversing the options, but on different row sets, so compare the rates loosely.

Throughput, not a controlled comparison

For 21 binary decisions against one state they report 1,023 ms with batched suffixes on an RTX 3090 in BF16. ruling measures 216 ms at one ordering and 495 ms at three on an Apple Silicon Mac in 4-bit, on a different state. Hardware, precision, and inputs all differ; the numbers show both are in the same range, nothing finer.

Where they are still ahead

  • Committed raw per-row predictions with checksums. ruling reports are regenerated by command, not committed.
  • A browser-only WebGPU demo.