Evaluation

Every live submission is scored against hidden packs derived after receipt. The seed binds the epoch secret, a future Base blockhash, epoch id, patch hash, parent root, miner address, corpus root, and bundle hash.

CoreTex uses two samples:

For each query, the evaluator decodes the substrate, builds candidates, renders Memory-IR where enabled, reranks query/document pairs, scores against graded qrels, and compares the patched substrate to the parent substrate on the same pack.

nDCG@10 is the main retrieval metric. Secondary signals cover temporal freshness, relation recall, abstention, structural validity, and policy-atom effects.