How the research
actually gets made.
Six mechanisms that make production telemetry safe to contribute and credible to study.
Read-only capture
One command installs capture hooks for Claude Code, Cursor, Codex, and Antigravity. Hooks only read finished-session files, never write to your environment, and are built fail-open so they can never break your tools. Revoke anytime.
Encrypted, segregated raw store
Raw traces are sealed with XChaCha20-Poly1305 under per-project keys the moment they arrive, in an isolated database the application layer has zero access to. Encryption keys live only in the redaction worker.
Default-DENY redaction
Two automated passes on our servers: deterministic secret detection (API keys, JWTs, private keys, connection strings, high-entropy tokens) plus a Presidio + spaCy ML pass for PII. Nothing is eligible for research without an explicit clean verdict — anything uncertain is quarantined, never auto-promoted. Redaction logs record categories and confidence, never values.
Provenance & attestation
Every trace carries a deterministic ID, three content fingerprints, and provider-response attestation (response IDs + cost-math checks) captured at ingest — so contributions are verifiable and duplicates can't be double-counted.
Transparent economics
Cashback is a published per-model-tier rate on output tokens — 10% Fable down to a 1% floor. Every payout runs through an append-only audit ledger.
Research in the open
PolymarketBench runs live with real money: every bet's reasoning, full agent transcript, and trace link is published. Corpus studies will ship the same way.