smalldecideGet the code
Evidence / Public-data checkpoint

A separate run.
Its complete results.

This later experiment uses a different checkpoint (f64a7763…) from the v0.3.0 release (ec832471…). Its results and recordings are shown together here.

This experiment remains a draft, labelled v0.3.0-public. The report download begins with a correction to its original release wording. “Public-data” describes its training sources; the weights still require repository access.

Read the report ↗Download report data ↗Release results ↗

Task accuracy

Frozen test and held-out splits. The reference is the pretrained model before this run's training. These tasks do not establish reliability for autonomous financial approvals.

EvaluationTrainedReferenceScope
CLINC: 8-way routing96.7%54.0%150 · test
Emotion presence88.7%50.7%150 · test
CommonsenseQA56.0%23.3%150 · test
Synthetic threshold100.0%52.7%150 · test
Synthetic ordinal rubric100.0%31.3%150 · test
WinoGrande [unseen]50.0%56.0%150 · heldout
Agent trace [synthetic, unseen]77.3%12.0%150 · heldout
Security policy [synthetic, unseen]41.3%68.7%150 · heldout

Full benchmark and local latency data ↗

The model receives health, ammunition, and visible-enemy coordinates from ViZDoom. It does not infer from video pixels. Recordings use game time and do not show inference speed. Complete mission metrics ↗

Game artwork is from Freedoom. Artwork license and attribution ↗

WIKIPEDIA / TWO RECORDED ROUTES

Both routes cycled.

Each decision selects a real outgoing link. Candidate probabilities do not represent the probability of reaching the target.

Route and chosen pathOutcomeData
Bicycle → PhysicsBicycle → Bicycle (disambiguation) → Bike (disambiguation) → Bik (disambiguation) → Bike (disambiguation)CycleTrace ↗
Cat → Computer scienceCat → Cats and the Internet → CatCycleTrace ↗
LOCAL TIMING / APPLE M4 MAX

27 questions, 3.32 seconds median.

Recorded on an idle machine. No chat-model comparison was recorded, so this does not establish a speed advantage over another model.

Recorded answers and timing ↗Input request ↗

SmallDecide was previously named OpenAuditor. Original recordings and reports use the earlier name.