Crivo-Prumo-400M
A typed-decision model: an encoder that reads every question in one pass, plus exact rule capabilities that claim what they can solve exactly, routed by capability.
Crivo-Prumo-400M answers the Decision Index /
systemone wire format: a state plus typed questions (choice, noul) in, a probability for every option out.
- Crivo-400M (395,838,465 parameters), a typed-decision encoder, reads every question in one pass.
- Prumo (527 hand-set constants, no learned weights) recognises structures it can solve exactly (code to run, a chess position, a Boolean formula, a date, an arithmetic problem, a tool catalogue…) and claims those questions.
- A fixed routing table decides, per claiming capability, whose answer is kept. It was chosen on development data only.
Highlights
- Full score 24.86 on the Decision Index 0.4.1 (maintainer-run private tests: same skills 32.55, new domains 9.48).
- Public index 45.50 on the Decision Index 0.3 (shown on the board, not counted in the Full score).
- Size: 395,838,992 numbers in total (395,838,465 learned parameters + 527 Prumo constants).
- No retraining, no new data. The weights are Crivo-400M's, byte for byte; Prumo's code and constants are those of its release.
- No truncation: Crivo reads inputs whole up to 32,768 tokens.
What it is
| Kind | hybrid: encoder (Crivo-400M) + rule-based capabilities (Prumo v1), routed by capability |
| Size | 395,838,465 learned parameters (model.safetensors) + 527 hand-set constants (prumo.safetensors) |
| Input / output | Decision Index wire format: state + questions (choice, noul) → a probability for every option |
| Runtime | PyTorch + transformers (Crivo), CPU Python + chess (Prumo); GPU recommended for the encoder |
| Language | English |
| Author | Tardelli Stekel |
How it decides
For every request, with no benchmark names or item ids:
- Crivo scores every
choice/noulquestion (one encoder pass per question; the state is tokenized once). - Prumo runs its twelve capabilities in a fixed order; each either claims a question and returns probabilities, or passes it on.
- Routing by the claiming capability only (
route.json). A question no capability claims keeps Crivo's answer.
| capability | what Prumo does | kept answer |
|---|---|---|
code |
runs Python code from the request in a sandbox | Prumo's answer |
chess |
node-limited alpha-beta search on a FEN | Prumo's answer |
tools |
BM25 tool relevance + schema validation | product of both |
retrieval |
BM25 over the request's candidates | Crivo's answer |
grounding |
lexical support of a claim by its evidence | product of both |
consensus |
centroid of candidate statements | product of both |
labels |
IDF overlap between text and label names | Crivo's answer |
music |
chord templates | Prumo's answer |
logic |
exact Boolean logic / SAT | Prumo's answer |
puzzles |
exhaustive search, claimed only if the solution is unique | Prumo's answer |
dates |
calendar arithmetic | Prumo's answer |
arith |
exact-fraction arithmetic | product of both |
"Product of both" is the normalised product p_Prumo · p_Crivo.
How the table was chosen.
- Crivo-400M and Prumo were both run on development sets that are not part of the suite: internal panels for same-skill and new-domain questions, train rows of the suite's source datasets, sibling datasets of the suite's skills, and generated chess, arithmetic and logic sets. That is about 41,000 requests.
- A capability leaves the default (Prumo's answer) only if Crivo's answer, or the product, is better with a bootstrap 95% interval above zero, in both halves of the data, and without clearly hurting any group.
- Two capabilities had no development evidence (
music,consensus). They follow the type of the capability: exact solvers keep Prumo, heuristics take the product. - The public suite was never used to choose the table.
Results
| Crivo-Prumo-400M | |
|---|---|
| Decision Index 0.4.1, Full score (board, private tests) | 24.86 |
| private tests of the same skills | 32.55 |
| private tasks from new domains | 9.48 |
| Decision Index 0.3, public index (not counted) | 45.50 (raw 59.80) |
| Knowledge & Reasoning | 27.1 |
| Language Understanding | 44.6 |
| Retrieval & Classification | 51.3 |
| Tools & Automation | 71.6 |
| Arts & Human Taste | 36.2 |
(Area rows: skill ×100 on the public suite.)
Decision Index 0.4.1 — Full score (board)
Scored by the maintainers on the board (Decision Index 0.4.1, board data of 2026-11-17). The Full score is 0.60 × private tests of the same skills + 0.40 × private tasks from new domains, both mapped onto the public scale. The public index is shown next to it but does not count.
| Full score 0.4.1 | Private same skills | Private new domains | Public index (not counted) | Board latency, median / p95 (1× RTX PRO 6000) | Rank group |
|---|---|---|---|---|---|
| 24.86 | 32.55 | 9.48 | 45.50 | 44 / 178.7 ms | 45 of 64 |
Models of similar size (board, Decision Index 0.4.1):
| Model | Parameters | Full 0.4.1 | Same skills | New domains |
|---|---|---|---|---|
| GLiDE 0.4B | 394.8M | 33.84 | 46.92 | 10.00 |
| Crivo-Prumo-400M | 395.8M | 24.86 | 32.55 | 9.48 |
| GLiClass v4.1 edge (0.4B) | 439.8M | 32.47 | 44.09 | 10.70 |
| GLiDE 0.8B | 752.4M | 35.91 | 49.40 | 11.14 |
| Adjudex-0.8B v5 | 853.0M | 32.81 | 44.61 | 10.75 |
| PKU-KCL-Meteor-0.8B | 853.0M | 30.94 | 44.04 | 7.67 |
| OneJev1.6-0.8B | 853.0M | 29.58 | 40.49 | 9.28 |
Decision Index 0.3 — public index (self-run)
How the run was produced. The Crivo-400M answers are those of its official complete run (results, kit 9eb2dbe, edition 0.3); Prumo v1 was run on the same 140,620 requests; the two were combined with this package's route.json. This offline composition is what the package computes: on Crivo-Prumo-68M, where the package itself ran the whole suite, the offline composition agrees with the package's run on 99.998% of 323,259 questions (38.78 vs 38.75). 140,178 requests ok, "complete": true.
Results (compact: no payloads, no suite text): https://huggingface.co/datasets/tardellirs/crivo-prumo-results/tree/1acb239df9a6913f6098767715062c7f4cf5e88a/runs/crivo-prumo-400m.
The family
| Model | Parameters | Full 0.4.1 | Public 0.3 index |
|---|---|---|---|
| Crivo-Prumo-17M | 16.9M | 14.08 | 29.79 |
| Crivo-Prumo-32M | 32.0M | 16.02 | 31.92 |
| Crivo-Prumo-68M | 68.4M | 20.88 | 38.75 |
| Crivo-Prumo-400M | 395.8M | 24.86 | 45.50 |
| Crivo-Prumo-1B | 1,031.3M | 30.49 | 52.88 |
Crivo-150M is not in the family: its hybrid would sit below Crivo-Prumo-68M.
Latency
Board-measured (1× NVIDIA RTX PRO 6000, Decision Index 0.4.1): median 44 ms, mean 64.5 ms, p95 178.7 ms.
The package adds Prumo's CPU time to Crivo-400M's encoder time (Crivo-400M alone, kit-measured: median 23.4 ms, p95 73.6 ms). On Crivo-Prumo-68M the whole run measured median 22.7 ms and mean 42.6 ms (shared GPU). Prumo's slowest paths are the chess search (about 0.1–0.3 s per question, bounded by nodes) and the code sandbox. The board's admission limit is 1,000 ms (median, mean or p80).
Quick start
pip install "decision-index[transformers] @ git+https://github.com/apolinario/decision-index@9eb2dbe" chess==1.11.2
hf download tardellirs/crivo-prumo-400m --local-dir crivo-prumo-400m && cd crivo-prumo-400m
python -m decision_index pipeline --engine decision_index_engine:DecisionEngine --edition 0.3 --out runs/crivo-prumo-400m
| File | What it is |
|---|---|
decision_index_engine.py |
the hybrid engine (Crivo + Prumo + routing) |
route.json |
the routing table above |
crivo_engine.py, model.py, crivo.py, model.safetensors, config.json, tokenizer*.json |
Crivo-400M, unchanged |
prumo_engine.py, prumo_rules/, prumo.safetensors |
Prumo v1, unchanged except that its constants file is named prumo.safetensors |
Intended use and limitations
- Built for the Decision Index wire format and English inputs. Crivo's own limitations apply (see its card).
- Prumo executes Python code taken from requests in a sandbox with resource limits (see the Prumo card). Do not expose it to untrusted input without further isolation.
- Several Prumo capabilities target structures that also appear in the public suite (chess positions, code, chord sequences). On the private tests the scores are lower (same skills 32.55, new domains 9.48, against a public index of 45.50); take the Full score as the guide for unseen inputs.
- The routing table is fixed. It was chosen on development data and was not tuned on the suite.
License
The weights and the package are released under CC BY-NC 4.0, as Crivo-400M. Prumo's code (prumo_engine.py,
prumo_rules/) is under the Apache License 2.0 (LICENSE-prumo).
See also
Crivo-400M · Prumo · Crivo collection
Citation
@misc{stekel2026crivoprumo,
title = {Crivo-Prumo-400M: a Crivo encoder and the Prumo rule engine in one decision engine},
author = {Tardelli Stekel},
year = {2026},
url = {https://huggingface.co/tardellirs/crivo-prumo-400m}
}
- Downloads last month
- 73
Model tree for tardellirs/crivo-prumo-400m
Base model
jhu-clsp/ettin-encoder-400mEvaluation results
- Public index (complete public suite) on Decision Index 0.3self-reported45.500
- Full score, Decision Index 0.4.1 (maintainer-run private tests) on Decision Index 0.4.1self-reported24.860
- Private tests of the same skills on Decision Index 0.4.1self-reported32.550
- Private tasks from new domains on Decision Index 0.4.1self-reported9.480