PERFORMANCE THAT HAS TO SHOW ITS WORK
A speed claim without a baseline and a repeat rule is just a fast-looking sentence.
00 baseline
BASELINE v12.2.1^{}
WARMUPS 3
MEASURED RUNS 9
STATISTIC MEDIAN
RELATIVE MAD <= 5 percent
10 acceptance gate
The core geometric mean must be at least 1.25x. No established case may fall below 0.90x without explicit review.
20 6502 authority corpus
The tracked NMOS 6502 corpus covers 151 documented legal opcodes. The historical million-instruction workload must preserve exact machine state and reach at least 1.20x speedup under the release contract.
30 speed does not buy semantic credit
Error timing, bounds behavior, borrows, allocation effects, cleanup and Drop remain part of the proof. O1 is not allowed to make the program faster by becoming a different program.
BENCHMARK RULE. Measure the machine you actually ran. Do not benchmark your optimism.