All issues

pv score: a missing binding registry scores 0.00, and a valid registry is refused over one enum variant

#2704

· noahgift opened

on Aug 25, 2026

The dogfood release gate is unrunnable for every sibling repo whenever aprender is not on main

#2701

· noahgift opened

on Aug 25, 2026

aprender-contracts-macros: #[contract] fails open in four places, and the doc comment claims the opposite

#2699

· noahgift opened

on Aug 25, 2026

perf(serve): the prefill fixed cost is HOST-SIDE — 1018 synchronous HtoD copies and 904 device allocs, not kernels

#2697

· noahgift opened

on Aug 24, 2026

P0: published apr 0.64.0 SILENTLY IGNORES --gpu — 15.7 tok/s decode, 7.5s to first token, 0.099x llama.cpp

#2696

· noahgift opened

on Aug 24, 2026

perf(serve): apr holds 3.08x the VRAM of llama.cpp for the same Q4_K_M weights — 14,030 vs 4,554 MiB

#2695

· noahgift opened

on Aug 24, 2026

perf(serve): decode is 1.54x behind llama.cpp — 103.3 vs 158.9 tok/s, same GGUF, same GPU

#2694

· noahgift opened

on Aug 24, 2026

perf(serve): prefill is 3.64x behind llama.cpp — 2,860 vs 10,399 tok/s, same GGUF, same GPU

#2693

· noahgift opened

on Aug 24, 2026

F12 — fabricated measurement: a value with the form of a measurement and nothing behind it

#2679

· noahgift opened

on Aug 24, 2026

PARITY-011: the pre-release and post-release dogfood protocols gain their bench phase — and say which subject each one measures

#2678

· noahgift opened

on Aug 24, 2026

PARITY-010 (CUT 4): the pre-publish GPU comparator lane — the only place an apr-vs-llama.cpp ratio is computable

#2677

· noahgift opened

on Aug 24, 2026

PARITY-009 (CUT 3): pin llama.cpp as the FOURTH token of the existing mechanism — it is rediscovery #5 of a rule already stated once

#2676

· noahgift opened

on Aug 24, 2026