Validator protocol for empirical quality verification (CP-015).
Any Canon node can opt into validation. Validators download content from
Arweave, index it locally, run test queries against it, measure retrieval
quality, and publish OpenSkill Bayesian ratings for content contributors.
These ratings help other nodes prioritize what to download from Arweave.
Validators also test live search peers by connecting over Tor, sending test queries, and rating peer quality for circuit selection.
The approach adapts Templar/Covenant (tplr.ai), which uses empirical
quality verification for decentralized pre-training. OpenSkill Bayesian
ratings provide a robust, game-resistant ranking that converges even
with sparse observations. Window-based coordination ensures validators
test the same content without requiring blockchain synchronization.
Data flow:
- Load test query corpus (from Arweave or local file)
- Compute current validation window
- Select test queries deterministically from corpus
- For each contributor: download from Arweave, index, run queries, measure metrics
- Rank contributors by composite score, update
OpenSkillratings pairwise - For each peer: connect over Tor, send queries, measure quality + latency
- Rank peers by composite score, update
OpenSkillratings pairwise - Publish results to Arweave