Haven Learning Lab: fictional generated data only. dl_arrays: 40 two-feature observations; first24 train/next8 validation/last8 reserved. Vision: 32 generated 1x5x5 strokes; 20 train/6 validation/6 reserved. Text CSV: 24 invented phrases, label0 billing/1 access; 16 train/4 validation/4 reserved. Vocabulary train-only. handbook.json: five fictional passages with stable H01–H05 IDs; retrieval-only baseline, no external API. Never interpret toy metrics as real-world performance.
