Galahad: the harness that makes your model cheaper and smarter at once

Hashes every incoming prompt and asks one question: have we done this before? A live in-VRAM registry answers in under two microseconds; a durable on-disk ledger survives restarts and moves byte-identical between machines. A lookup is only a hit when two independent hashes agree, so it never returns the wrong memory. Merlin also runs standalone as content-aware deduplication — stripping redundant context before it reaches any model, including pure API setups.

1.4 µs to select · exact addressing makes zero errors where approximate retrieval picks wrong 94.3% of the time

Takes the state Merlin found and splices it into a live request with zero tokens recomputed. The restored logits are byte-for-byte identical to computing fresh — SHA-256 equal, zero KL divergence, verified across two model scales and two GPU targets. Blocks are merged with position re-binding and recompute-fused, so the model can reason across them rather than merely read them.

bit-exact restore · 100% argmax agreement over 50 samples · 6–23 ms per full reuse at 36 mWh