Protocol Rules
CoreTex is built around four protocol rules:
- The reward law is retrieval-native. The evaluator scores whether the substrate retrieves answer-bearing documents for hidden queries.
- Corpus growth drives difficulty naturally. There is no
CorpusHardnessIndex, operator hardness knob, or credit-weighted rate limiter. As the corpus expands, the same 1024-word substrate must compress more useful retrieval structure. - Major corpus deltas use baseline grace. If a large
eval_hiddendelta lands,minImprovementPpmfreezes for one epoch while the parent substrate baseline is recomputed on the new query distribution. - Live eval uses per-patch on-chain randomness. Each submitted patch gets its own hidden gate/confirm packs derived from a future Base blockhash, patch hash, miner address, parent root, corpus root, bundle hash, and epoch secret.
The production bundle pins:
| Component | Role |
|---|---|
| Bi-encoder | BAAI/bge-m3, dense mode, CPU-only deterministic runtime |
| Reranker | Qwen/Qwen3-Reranker-0.6B, production cross-encoder |
| Label audit model | IAAR-Shanghai/MemReranker-4B, offline audit/reference path |
| Primary metric | nDCG@10 over hidden query packs |
| Secondary metrics | MRR@10, Recall@10, temporal accuracy, relation recall, abstention, structural validity |
| Bundle binding | bundleHash == coreVersionHash for the active epoch |
| Randomness binding | Per-patch future Base blockhash with launch targetBlockOffset = 30 |
Model revisions, file hashes, runtime pins, evaluator profile, corpus root, query-pack policy, and replay tolerance are all part of the client bundle. A coordinator or verifier that sees a bundle mismatch must refuse canonical evaluation.