What We Publish (and What We Withhold)

Swarmgram's validation program depends on trust. We publish aggregate benchmark results, evaluation protocol, failure modes, and forward forecasts. We do not publish proprietary training pipelines, model weights, agent internals, calibrator parameters, or inference routing details.

Public

• Pooled and panel-level MAE with sample sizes
• Cross-validation and held-out protocol descriptions
• Category-level failure cases (e.g., national census-weighted items)
• Pre-registered forward forecast items before ground truth release
• Scale invariance results (n=1,000 vs n=10,000)

Available under NDA / partner review

Researchers, investors, and enterprise buyers can request materials at hi@swarmgram.com. We prefer evidence over architecture tours in public channels.

• Full item-level benchmark tables
• Swarmgram City protocol summaries
• Private panel architecture for enterprise customers
• Due diligence data room materials