Primus Decision 0.1 is open source. Apache-2.0, weights and protocol on GitHub · Hugging Face

Primus Decision 0.1 · Architecture

Two small encoders, one typed answer

The released Decision 0.1 component, end to end. No transformer and no pretrained encoder anywhere: both encoders are trained from scratch on the benchmark's training part, and the only attention pools one question's token states.

JSON stateone workflowflattened textpath: value lines + relationsS4D member2 diagonal state-space blocksGRU member2 bidirectional GRU layersLSA case vectorTF-IDF → 256-d SVD, per memberoption scorer + softmaxper member, over offered optionsaverage 0.5 / 0.5raw distribution; optional profilequestion-conditioned attention pooling in both members · no token-to-token attention · no transformer · no pretrained encoder

The path of one decision

Every prediction runs both members in full. Neither is a router or a fallback: their two probability distributions are averaged with equal weight.

The data path turns the JSON state into text. The question path scores each offered answer separately, so the output always covers exactly the answers the caller offered.

  1. Request
    A JSON state, a typed question and the answers it allows.
  2. Flatten and relate
    One path: value line per leaf, plus comparison, membership, sign and null sentences.
  3. S4D member
    1,932,609 parameters
    Diagonal state-space encoder, plus its LSA case vector.
    GRU member
    1,782,465 parameters
    Bidirectional GRU encoder, plus its LSA case vector.
  4. Option scoring
    Each offered answer scored separately; one softmax per member over exactly those answers.
  5. Ensemble
    0.5 × S4D + 0.5 × GRU
    The two members' distributions averaged.
  6. Typed probability distribution
    Raw by default; an optional temperature profile ships off.

Parameters

partneural parameters
S4D member (2 diagonal state-space blocks per direction, d = 192)1,932,609
GRU member (2 bidirectional GRU layers, d = 192)1,782,465
Total, all active on every prediction3,715,074

Counted from the frozen float32 safetensors: 15 MB of weights. M, measured by us Source: MODEL_SIZE_AND_PARAMETERS.md §1

Fitted tables

Each member carries an LSA featurizer: word 1–2-gram and character 3–5-gram TF-IDF with a 256-d truncated SVD, fitted on the training split. These learned feature tables are reported separately from the neural parameter count. They are 71.4 MB each, which is why the package is 157.9 MB installed and why peak memory is about 1.5 GB. M, measured by us Source: MODEL_SIZE_AND_PARAMETERS.md §2–3

How a state becomes text

The JSON state is flattened to one path: value line per leaf and extended with derived-relation sentences: numeric comparisons within a key family, list membership, signs and nulls. The first 640 tokens reach the encoders; at most 40 relation lines are added. Question instructions and option texts enter as bags of words.

Each option is scored separately, so a choice question with K options gives a K-way softmax over exactly those options. Callers supply any subset of the schema's trained option keys in any order, together with their descriptions. The runtime validates the schema before inference.

Decision 0.1 in the Primus programme

This component computes typed decisions from the supplied state. The broader programme explores persistent memory, recurrent reasoning, planning and tool interfaces, with separate release and evaluation milestones. The Road to Primus