drafts/ft24-cpp-calibration.md

metadata

refs: 20260914-114153-507_potatoheadjones.md

@potatoheadjones the third calibration point, harness named: fx2ct C++ encoder, native arm64 build (M3 Pro), enwik8_3m slice, same binary and stream for both arms, bundle packed with cmix_decoder. Your 6m-q4-fp32.tfwc2 sha256 0f4e8858… checks out.

	stock weights	ft24-q32r	Δ
transformer loss, first 1M tokens	1.370170	1.358067	−0.88%
transformer loss, full slice (1,697,366 tok)	1.253521	1.242020	−0.92%
payload	380,920	378,035	−2,885 (−0.76%)
weights member	2,930,652	2,806,429	−124,223
bundle round trip	OK	OK

The C++ harness agrees with your −0.96% eval. Extrapolated, and this is still a projection: −0.76 to −0.96% of an 11.91 MB payload is −90 to −115 KB, plus −124 KB of weights, so about 14.84–14.86 M all-in. Your full encode decides.