arc-decide
Quadrant scores
See the full quadrantScored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.
- Maturity6.3/10
MIT weights (0.6B v0.1, 0.4B v0.2) with int8 ONNX for CPU, a community GGUF, a detailed card, canonical question list and DATA.md. Python Decider class, no /v1/systemone server. Used inside DocsGPT; no SLA.
- Capability6.7/10
noul, choice (2–16 options) and 3- or 4-level score heads, temperature-scaled per question on held-out rows. Author tables on public held-out sets for RAG relevance, injection, grounding, routing and intent. No latency figure. Not on any board.
- Adoption3/100
HF on 14 Oct: 58 downloads (v0.1) and 35 (v0.2), 1 like; 489 downloads on mradermacher's GGUF. The card says it powers the decisions in DocsGPT.
Vendor claims
- arc-decide-0.4b-v0.2 fits in 1 GB of memory on 4 CPU cores and keeps v0.1's DocsGPT relevance (nDCG@10 0.916); three of its pre-set bars were missed and accepted[Vendor claim — not independently verified]
- Same held-out sets, zero-shot baselines: deepset injection AUROC 0.982 vs Laya 0.901; AggreFact balanced accuracy 0.701 vs Laya 0.595; Banking77 85.7% vs Laya 86.5%[Vendor claim — not independently verified]
Read this first. arc-decide is a specialist for the small decisions a RAG or agent app makes on every request. It is trained on a fixed list of canonical questions (
questions.json) and the card says answers get worse far from them. All numbers are Arc53’s (UnverifiedClaim).
What it is
Arc53, the team behind DocsGPT, published arc-decide-0.6b-v0.1 on 30 September 2026 and arc-decide-0.4b-v0.2 on 11 October, both MIT. Give it a state (text or JSON) and typed questions; it returns calibrated probabilities in one forward pass, with no generated text. The question shapes follow TypeSafe’s System One: noul (P(yes)), choice (2–16 options you pass) and score (3 or 4 levels). It covers English, Spanish, Chinese and Russian.
v0.1 is Qwen3-0.6B-Base with typed linear heads read at the last token. v0.2 keeps the first 16 of its 28 layers (about 407M parameters) and is retrained on v0.1’s data, with every target mixed half and half with v0.1’s own answers. Two int8 ONNX builds target hosts without a GPU. The questions about one state share its encoding.
Author results (UnverifiedClaim)
| Held-out set | v0.2 (0.41B) | v0.1 (0.6B) | Laya (421M) | Qwen3-4B-Instruct |
|---|---|---|---|---|
| Injection, deepset (AUROC) | 0.982 | 0.992 | 0.901 | 0.817 |
| Grounding, AggreFact (bacc) | 0.701 | 0.712 | 0.595 | 0.656 |
| Intent, Banking77 (acc) | 85.7 | 87.5 | 86.5 | 88.9 |
| Routing, needs retrieval (AUROC) | 0.915 | 0.915 | 0.528 | – |
Laya and Qwen run zero-shot; arc-decide was trained on these kinds of question, not on these sets. Against v0.1, v0.2 missed three of its own bars (BoolQ AUROC 0.901 vs 0.912, deepset injection 0.982 vs 0.992, one analytics goal) and the team shipped it anyway.
Score working (14 Oct 2026)
- Maturity = availability 9 × 0.30 + docs 7 × 0.25 + integrations 5 × 0.25 + support 3 × 0.20 = 2.70 + 1.75 + 1.25 + 0.60 = 6.30. Integrations 5: DocsGPT plus a community GGUF.
- Capability = types 9 × 0.30 + calibration 7 × 0.25 + latency 4 × 0.25 + accuracy 6 × 0.20 = 2.70 + 1.75 + 1.00 + 1.20 = 6.65 → 6.7. Types 9: choice capped at 16 options, score at 4 levels. Latency 4: no figure beyond “fits on 4 CPU cores”.
- Adoption = engagement 2 × 0.35 + ecosystem 3 × 0.35 + production 3 × 0.30 = 0.70 + 1.05 + 0.90 = 2.65 → 3.