Granite 3.3 2B Instruct
15 wins · 0 lossesTwo-sided exact test: p=0.00006103515625. Status: AUDITED.
PRIMARY EVIDENCE
LLM alone vs LLM + Nexus Safe
Accuracy is directly labelled. No hover is required.
47audited questions / questions auditées
2/4comparisons classified as GAIN
+68.1 ppdescriptive aggregate gap
LLM aloneLLM + Nexus Safe
PAIRED DECISION
Gain, loss or inconclusive result
A higher score is not enough: the decision accounts for discordant pairs and sample size.
Gemma 3 12B IT QAT
12 wins · 0 lossesTwo-sided exact test: p=0.00048828125. Status: AUDITED.
Llama 3.1 8B
5 wins · 0 lossesTwo-sided exact test: p=0.0625. Status: AUDITED.
Granite 3.3 2B Instruct · code
0 wins · 0 lossesTwo-sided exact test: p=1. Status: AUDITED_SMALL_SAMPLE.
| Model | Domain | n | Alone | Nexus | Δ | Decision | exact p |
|---|---|---|---|---|---|---|---|
| Granite 3.3 2B Instruct | mixed verifiable | 18 | 3/18 (16.7 %) | 18/18 (100 %) | +83.3 pp | GAIN | 0.00006103515625 |
| Gemma 3 12B IT QAT | mixed verifiable | 18 | 3/18 (16.7 %) | 15/18 (83.3 %) | +66.7 pp | GAIN | 0.00048828125 |
| Llama 3.1 8B | mixed verifiable | 6 | 1/6 (16.7 %) | 6/6 (100 %) | +83.3 pp | INCONCLUSIVE | 0.0625 |
| Granite 3.3 2B Instruct | code | 5 | 1/5 (20 %) | 1/5 (20 %) | +0.0 pp | INCONCLUSIVE | 1 |
AUDIT LEDGER
Informative evidence that is not a product claim
These campaigns remain visible, but are excluded from primary averages.
Excluded or running campaigns
HOW TO READ THIS
What these numbers do — and do not — show
They show a reproducible gain on some verifiable problems when Nexus has an appropriate executor or verifier.
They do not show universal improvement in reasoning, creativity, code or open-domain knowledge.
Publication rule: previously observed answers are never reused to score a new campaign; sealed sets are not used for training.