Aleria Judge (MORJ)

Last updated:

MethodApache-2.0~17 ★

⚠️ Why not a model

Training-free: it reads stock Gemma-4-31B next-token probabilities over rotated option orders. No decision weights of its own, so no model page (precedent: simple-jev, AnyJev).

Technical specs

Base LLMs
google/gemma-4-31B-it (stock), nvidia/Gemma-4-31B-IT-NVFP4 (TensorRT-LLM, text only)
Decision types
choice, score, boolean
Features
training-free, option-order-debiasing, prefill-only-engine, images, systemone-api, playground
License
Apache-2.0

Aleria Judge (GitHub, Apache-2.0, 17 stars) turns stock Gemma 4 31B into a typed decision server on one 80 GB GPU. It was published as MORJ (the GitHub URL still redirects); the README now calls it Aleria Judge 1.1.0. You send a state (text, any JSON, up to eight images) and typed questions (a choice of up to 255 options, yes/no, an ordinal score) to POST /v1/systemone and get probabilities back. There is no fine-tuning and no generated text.

How it works

Each question becomes a few prompts in the state-first layout of Mapika/decider, and the engine reads the softmax over the option-label tokens in one prefill:

  1. Two option orders: the original and a half-rotated one, averaged, to cancel position bias. Rotations by a quarter and three quarters are added when the two disagree on a question with three or more options (about 10% of requests).
  2. Nine-digit rating for every two-option question, in both orders, so near-certain yes/no answers get spread probabilities (which matters for Brier and ranking).
  3. Option-count temperature for questions with three or more options, fitted on the authors’ training data.
  4. Effort per request (low, medium, high); the default and the Decision Index kit use high.

The default engine is a prefill-only “token tree” engine (each request’s shared state and images are encoded once); TensorRT-LLM remains available, including an NVFP4 checkpoint (text only, about 1 point lower). A playground at port 30104 shows what the engine did for any request.

Why it is not a model page

Aleria Judge adds no trained weights: the decisions come from a prompt layout, rotation averaging and readout logic over the stock google/gemma-4-31B-it checkpoint. That is the same class as simple-jev and AnyJev. The board entry is evidence for the method, not for a model.

Board position (official)

Decision Index 0.3.2, data generated 16 Oct 2026 19:00 UTC. Full = 0.2 public + 0.5 same-skills + 0.3 new-domains.

MORJ 1.0.1 (Gemma 4 31B) Jev 1.13
Full 60.96 (#14) 60.11 (#18)
Public 60.24 (#21) 57.96 (#37)

The Full gap to Jev is 0.85, inside the board’s 0.9 tie band, so the board lists them tied (tie-group rank 5). The board row is version 1.0.1; the README describes 1.1.0, whose own public run is 60.23 text (all 140,620 requests answered; median 97 ms, p80 393 ms, mean 312 ms on one RTX PRO 6000) and 73.38 on the vision suite. Those are the authors’ runs, not maintainer-reviewed. The same GPU measured 173 ms median for the earlier TensorRT-LLM engine. There is a dataset repository (Aleria-AI/morj-decision-index) and no model repository on Hugging Face.

Fit / anti-fit

Fit: a team that wants a typed decision endpoint from an unmodified Gemma 4 31B, fully on-premise or air-gapped, with images, and a transparent engine.

Anti-fit: a trained-weights model, a smaller footprint (80 GB GPU), or a hosted API.

Limits

  • Needs an 80 GB GPU for the BF16 default; the NVFP4 route is text only and needs a Blackwell GPU.
  • The board row (1.0.1) and the README (1.1.0) are different releases.
  • Gemma 4’s license terms apply to the base model.