nx_coder_fidelity_gate.nx
buildroot/runtime/nx_coder_fidelity_gate.nx
about
nx_coder_fidelity_gate.nx -- L3b: MEASURED coder-weights bring-up (float ref vs 100%-integer
tower, same coder GGUF, greedy). The load-bearing L3b claim is NOT "bit-faithful to float"
(greedy argmax flips on near-ties and compounds -- expected) but: the coder weights LOAD and
GENERATE COHERENT code through the integer tower, the config is sound, and the integer tower is
deterministic by construction. Float-match is NOT the forge ground truth -- the compiler is.
Both runners wrote knowledge/forge/coder_gen_{f32,nf}.txt (2 lines: "GEN p1 ...text=[..]" /p2).
PASS on: both PRODUCED coherent output AND p2 text byte-exact (proves arch+config+tokenizer+
dequant all correct -- if eps/rope differed, p2 would also diverge) AND mutation negctl caught.
REPORT (not gated): p1 text exact? + p1 text longest-common-prefix (the honest divergence depth).
license_tier: ORIGINAL expect_exit: 0
dependencies 2 imports · 0 importers
imports: nx_forge_ctx.nxnx_gate_verdict.nx
imported by: nobody (leaf or entry point)
call flow from main pre-order; caps 40 nodes / depth 6 declared; ↻ = already shown
structs
| none |
consts
| none |
functions
| 16 | func cf_text_after(buf: *u8, n: i64, startoff: i64, out: *u8, endoff: *i64) -> i64 |
| 51 | func cf_eq(a: *u8, an: i64, b: *u8, bn: i64) -> i64 called by 1: main |
| 63 | func cf_lcp(a: *u8, an: i64, b: *u8, bn: i64) -> i64 called by 1: main |
| 80 | func main(argc: i64, argv: *i64) -> i64 |