nx_perf_verdict.nx
buildroot/runtime/nx_perf_verdict.nx
about
nx_perf_verdict.nx -- substrate primitive for honest per-axis
head-to-head verdicts.
Per honest-hard-perf-verdict cardinal 2026-05-14:
"every head-to-head bench emits sealed-enum per-axis verdict and
every LOSE row names a specific concrete improvement target.
UNMEASURABLE is OK; 'straddles or beats' / 'competitive with' are
REFUSED weasel words".
Sealed enum (do not add weasel intermediates):
NX_VERDICT_UNMEASURABLE = 0 no comparable data yet (honest about gap)
NX_VERDICT_LOSE = 1 we are worse on this axis
NX_VERDICT_TIE = 2 parity
NX_VERDICT_WIN = 3 we are better
Standard axes (use these names so scoreboards aggregate cleanly):
throughput, accuracy, memory_at_parity, serialized_size,
determinism, patent_clean, license_clean, portability
genealogy_id: substrate_perf_verdict_2026_05_14
lineage_id: honest_hard_feedback
dependencies 3 imports · 1 importers
imports: nx_syscalls.nxnx_runtime.nxnx_tier.nx
imported by: nx_bench_hll_vs_datasketches.nx
structs
| none |
consts
| 33 | const NX_VERDICT_UNMEASURABLE: nx_int = 0 |
| 34 | const NX_VERDICT_LOSE: nx_int = 1 |
| 35 | const NX_VERDICT_TIE: nx_int = 2 |
| 36 | const NX_VERDICT_WIN: nx_int = 3 |
| 76 | const NX_VERDICT_WIN_FLOOR_PCT: nx_int = 101 |
| 77 | const NX_VERDICT_LOSE_FLOOR_PCT: nx_int = 100 |
functions
| 38 | func nx_verdict_label(v: nx_int) -> *u8 |
| 50 | func nx_verdict_emit(axis: *u8, verdict: nx_int, called by 3: mainnx_verdict_rationx_verdict_ratio_lower_better calls 4: printnx_verdict_labelprint_i64println |
| 83 | func nx_verdict_ratio(axis: *u8, nishi_value: nx_int, incumbent_value: nx_int, |
| 105 | func nx_verdict_ratio_lower_better(axis: *u8, nishi_value: nx_int, incumbent_value: nx_int, |
| 141 | func nx_verdict_emit_qualitative(axis: *u8, verdict: nx_int, |
| 161 | func nx_verdict_qualitative_simple(axis: *u8, nishi_score: nx_int, incumbent_score: nx_int, calls 1: nx_verdict_emit_qualitative |