The best thing our engine does is refuse to answer
Why we deleted a feature so TRONORIGIN would stop naming a controller when it has no evidence — and how a benchmark caught it guessing on 13 out of 13 accounts.
Practical write-ups from real projects — AI, crypto engineering, and the things we learned building them.
Why we deleted a feature so TRONORIGIN would stop naming a controller when it has no evidence — and how a benchmark caught it guessing on 13 out of 13 accounts.
How we built TRONORIGIN's origin-tracing engine to explain every score — signal by signal — and why an auditable heuristic beats a black box for forensic work.
The SSE newline-framing bug that silently corrupts streamed tokens, the fix, plus HMAC-signed history, atomic Durable-Object rate limiting, and why the assistant hydrates lazily.
Notes from the workshop — what we're building at KeyPair Infra and what we'll write about here.
Practical write-ups from real projects — AI, crypto engineering, and the things we learned building them. No spam.