nx_forge_rag_ab.nx
buildroot/runtime/nx_forge_rag_ab.nx
about
nx_forge_rag_ab.nx -- THE RAG VALUE A/B (2026-07-17): base-pack vs rag-augmented-pack on the
held-out fbench tasks, same engine, same judge, greedy pass@1 (mirrors nx_forge_curves cv_task:
fe_run maxc=1 maxnew=400 nsamples=0). Because the 0.5B floors at 0 on fbench pass@1 (A1 measured),
each leg ALSO classifies the failure from the banked loop ctx (/tmp/forge_loop_ctx.txt) --
1=GREEN 2=COMPILE-FAIL 3=RUN-FAIL 4=OUTPUT-MISS -9=unknown -- the softer signal with real
discrimination on a weak model. Rows -> sovereign store knowledge/store/rag_ab keyed
"<leg>:<task>" = [rc, class, us] (latest-wins; re-runnable suite command per growth-loop law).
Pre-step (documented): regenerate rag packs via _offc/nx_forge_rag.elf augment <taskf> <mini> ... 2 <ragpack>.
usage: nx_forge_rag_ab <port> <chat>
license_tier: ORIGINAL expect_exit: 0
dependencies 2 imports · 0 importers
imports: nx_forge_engine.nxnx_sov_ledger.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
| 14 | const AB_CTXCAP: i64 = 65536 |
functions
| 16 | func ab_read(path: *u8, buf: *u8, cap: i64) -> i64 |
| 36 | func ab_find(hay: *u8, hn: i64, needle: *u8) -> i64 called by 1: ab_classify |
| 52 | func ab_classify(rc: i64) -> i64 |
| 72 | func ab_row(key: *u8, rc: i64, cls: i64) -> i64 |
| 82 | func ab_leg(key: *u8, organ: *u8, taskf: *u8, expectf: *u8, packf: *u8, port: i64, chat: i64) -> i64 |
| 101 | func main(argc: i64, argv: *i64) -> i64 |