RINHA 2026 // YOLO MODE // PURE ASM API

YOLO SIGNAL
UNDER LOAD.

A public proof dossier for the assembly fraud scorer. The page now follows the C and .NET family: lane-labeled evidence first, runtime path second, CI caveats always visible.

Proof stays separated by lane: assembly runtime, corpus replay, GitHub Actions, Pages deployment, and reference-site lineage.

TCP arrives at the shared ASM load balancer, client descriptors move over Unix sockets, and `_start` writes response bytes back.

[asm] pure x86-64[corpus] 54100 clean[fdpass] stream sockets
PROOF LEDGER

ASM RUNTIME, CORPUS REPLAY, CI DEPLOYMENT.

Every number below is lane-labeled. GitHub Actions evidence, local replay targets, and official hardware are not collapsed into one decorative counter.

[ASM-RUNTIME]

IMPLEMENTATION LANE

Pure assembly API

ENTRY
_start
BUILD
as + ld
CHECK
purity pass

Repository-owned runtime implementation stays assembly-only. Tests block non-assembly implementation drift.

OPEN PURITY CHECK
[CORPUS-CLEAN]

LOCAL REPLAY LANE

54100 cases

FP
0
FN
0
HTTP
0

Public corpus replay target is zero false positives, zero false negatives, zero HTTP errors, and zero score mismatches.

INSPECT REPLAY
[CI-CLEAN]

GITHUB ACTIONS LANE

Main CI lane

SMOKE
pass
PURITY
pass
PAGES
live

GitHub Actions evidence confirms main branch build, smoke test, purity check, release build, CodeQL, and Pages deployment.

INSPECT RUN

OFFICIAL / REFERENCE SOURCE

The visual system is intentionally a variant of the C and .NET proof dossiers. Challenge facts remain separate from YOLO CI evidence.

LOCAL YOLO EVIDENCE

main currently publishes a Pages artifact, CI smoke and purity checks, an amd64 build, and CodeQL evidence for the assembly candidate.

RUNTIME TRACE

THE FAST PATH IS SHORT. THE CAVEAT IS VISIBLE.

01

Ingress

k6 / judge reaches :9999

The public boundary is the shared YOLO assembly load balancer.

02

FD pass

SCM_RIGHTS over Unix sockets

Accepted client descriptors move to api1.sock and api2.sock without a second public TCP listener.

03

Parse

manual HTTP + fraud fields

Assembly code reads only the bytes the workload needs for scoring and response selection.

04

Score

corpus table + branch-light path

The replay gate keeps shortcut decisions tied to the public corpus and score expectations.

05

Respond

prebuilt HTTP bytes out

Workers write prepared response bytes back on the inherited client descriptor.

A/B TEST SURFACE

THREE LANES TO TEST YOLO WITHOUT LOSING THE REFERENCE.

A

Evidence Terminal

Default lane. Lead with assembly runtime, corpus replay, CI status, and caveats above the fold.

REVIEW EVIDENCE LANE
B

Benchmark Ledger

Make provenance the lead object: source, run, compose, replay, workflow, and Pages deployment.

REVIEW LEDGER LANE
C

Runtime Trace

Lead with architecture: load balancer, fd pass, parser, scorer, and bytes written back.

REVIEW RUNTIME LANE

CI BENCHMARK STREAM

Candidate metrics are local evidence from GitHub Actions. Official hardware and reference-site proof remain separate lanes.

MAINCI LANE
0SMOKE FAILURES
54100CORPUS TARGET
LIVEPAGES DEPLOY