AAME: Primus Decision 0.1, an open-source model for typed probabilistic decisions
but in the right next step.”
How these numbers were measured
- [1] 3.7M neural parameters: 3,715,074 neural parameters (1,932,609 S4D + 1,782,465 GRU, both run on every prediction); the two LSA featurizers are fitted tables and are not counted. M, measured by us Source: MODEL_SIZE_AND_PARAMETERS.md §1
- [2] 75.1% accuracy on the official test split of LocalLLaMA/typed-decisions (400 cases, 2,000 decisions), one sealed run after freezing, raw probabilities. M, measured by us Source: SEALED_RESULT.json
- [3] ~140 ms p50 CPU inference: one five-decision case with the model loaded, 4-vCPU Intel Xeon 2.10 GHz, 2 threads; 115–162 ms p50 across six recorded runs with two methods on the same machine class. M, measured by us Source: EXPERIMENTS.md §4
- [4] Brier 0.059 and Raw ECE 0.127, raw probabilities, official test split. M, measured by us Source: SEALED_RESULT.json
- [5] 90.9% accuracy on the most-confident 50% of decisions: raw confidence, this test split. M, measured by us Source: EXPERIMENTS.md §3
Open weights.
Open protocol.
Primus Decision 0.1 is public under Apache-2.0: the model files, the runtime, the evaluation protocol and the sealed result, with SHA-256 hashes for every file. Download it, check the hashes, and run the example on a CPU. The public research boundary explains how released evidence and protected work fit together.
- Open source · Apache-2.0
- Local CPU inference
- Non-transformer
Download, verify and run
CPU only, Python 3.11+. No GPU, no API key, no sign-up.
$ python3 -m venv .venv && . .venv/bin/activate && pip install -U huggingface_hub
$ hf download The-Aame/primus-decision-0.1 --local-dir primus-decision-0.1 && cd primus-decision-0.1
$ grep -vE ' (README\.md|\.gitattributes)$' SHA256SUMS | shasum -a 256 -c - && \
python -m pip install -r requirements.txt --index-url https://download.pytorch.org/whl/cpu --extra-index-url https://pypi.org/simple && python examples/predict_example.pyYou get three probability distributions for one invoice, 20 % over its purchase order. The manifest covers the published documentation and evidence. The check skips the Hub-specific README.md and .gitattributes; frozen model and runtime artifacts match tag primus-decision-0.1. No Git LFS is needed.
On Linux, sha256sum works in place of shasum -a 256. On Debian and Ubuntu, python3 -m venv needs the python3-venv package.
~140 ms p50 CPU inference per five-decision case, measured by us (M) on a 4-vCPU Intel Xeon 2.10 GHz, 2 threads, model loaded; 115–162 ms p50 across six recorded runs. The CPU build of torch is about 200 MB.
From state
to decision.
Give Primus a structured state, a question and a set of valid answers. It maps the state to a compact representation and returns a probability for each answer. The released interface supports 20 question schemas across invoice processing, support, security and agent traces, with explicit schema validation.
Replay 25 casesSealed evaluation.
Measured results.
The frozen model records 75.1% accuracy on 2,000 Typed Decisions test decisions, with its probability quality, runtime and per-workflow results published together. The separate invoice case study extends the evidence to structured reconciliation.
View full benchmarksExplore accuracy, probability quality, model size and CPU inference by task. The tables identify results measured by AAME, checkpoints reproduced by AAME, and figures reported by other publishers.
The comparison in one sentence
- [1] Primus records 75.1% accuracy, NLL 0.637 and raw ECE 0.127 with 3,715,074 neural parameters. AAME's reproduction of the 421M-parameter Laya checkpoint on the same harness records 76.6% accuracy, NLL 0.707 and ECE 0.213. Primus uses about 113× fewer neural parameters; its separate feature tables bring the installed release to 157.9 MB. Full task results and measurement conditions accompany the comparison. M, measured by us R, reproduced by us Sources: SEALED_RESULT.json, EXPERIMENTS.md §2, §4, §6
Local decisions.
Inspectable evidence.
The released weights answered 26 of 32 structured reconciliation questions correctly (81.25%) without retraining in a broader synthetic invoice study, following 32/32 in the initial pilot. The full study includes narrative inputs, duplicate-ID checks, a classifier, Qwen3-1.7B and explicit rules.
Read the case studyOne public component. A broader research programme.
We are researching how memory, recurrent reasoning, learning, grounding and planning can connect around decision components. Each release documents its demonstrated capabilities while unreleased implementation remains protected.
Explore the research directionHow to check us
- Sealed protocolThe protocol was fixed before the test split was opened. One sealed release evaluation records the original result; later verification and diagnostics are identified separately.
- Reproduced to 1e-16A separately written metric implementation used by AAME reproduces the sealed result within approximately 1e-16 on a fresh run of the verified model.
- Bit-identical packageThe public package produces bit-identical outputs to the internal frozen package on 600 decisions.
- Every number labelledMeasured by us, reproduced by us, or reported externally, on every page.
- Check it yourself
grep -vE ' (README\.md|\.gitattributes)$' SHA256SUMS | shasum -a 256 -c -In the release folder; on Linux, sha256sum works too.
Releases and papers.
Primus for invoice decisions
Primus Decision 0.1 — Research Alpha
Introducing Primus Decision 0.1
Continuous Memory Consolidation Without Retraining: Solving the Edge Bloat Problem Through Computational Sleep
Causal Imagination Without Language Generation: Forward Simulation Through Graph Traversal
Primus: Cognition From First Principles — The Vyan Dual-Graph Cognitive Architecture Without Language Models
Cross-Domain Transfer Learning in Developmental Knowledge Graphs Without Training
Smaller models
for a more capable world.
We believe useful intelligence should not depend on massive models, but on better structure and calibration, measured in the open. Primus is our first step toward AI systems whose decisions can be checked.
Learn about AAME