Byte-exact KV-cache reuse, verified
Restoring a model's saved state has to produce the same answer as recomputing it. We check this the hard way: logits byte-for-byte identical, SHA-256 equal across model scales and GPUs.
Read the benchmarksEngineering notes
Short, honest write-ups of what we build and how we measure it. Every claim links to the method behind it. More notes are on the way.
Byte-exact KV-cache reuse, verified
Restoring a model's saved state has to produce the same answer as recomputing it. We check this the hard way: logits byte-for-byte identical, SHA-256 equal across model scales and GPUs.
Read the benchmarksTesting the encryption that protects saved state
Every protection is paired with a sabotage test that must fail. Every NIST CAVP vector reproduces, corruption is always detected, and 15 of 15 deliberate breaks of the encryption code were caught.
Read the security reportMeasuring energy, not guessing it
The GPU-energy figures come from measured runs with their conditions written down, not vendor math. Here is what we measured, on what hardware, and where the limits are.
See the methodsThree engines, one memory
Merlin deduplicates, Taliesin grafts saved state back into a live request with zero tokens recomputed, and Blaise keeps documents as byte-exact text. How they fit together.
How a request flowsStay in touch
We don't run a mailing list yet. For now, the fastest way to follow the work is the GitHub repository or a quick message.