RINHA 2026 // .NET 10 NATIVEAOT

FRAUD SIGNAL
UNDER LOAD.

A public proof dossier for a raw HTTP/1 fraud scorer: official upstream history, current CI candidate evidence, and the runtime path that keeps the hot loop small.

[official] issue #5716 [candidate] 0% failures [calibrated] 0.66ms p99
VERIFY UPSTREAM RESULT OPEN LATEST CI RUN TRACE RUNTIME PATH
> lane.official    issue #5716 // p99 0.99ms // failures 0%
> lane.candidate  run 26666573496 // p99 0.67ms // failures 0%
> lane.calibrated api=0.425 proxy=0.15 // p99 0.66ms
> runtime        native_aot // raw_http // uds // prebuilt_bytes
> topology      yolo_lb:9999 -> api[2] over unix sockets
> target        official resubmission with 0% failures
> caveat        ci numbers are evidence, not official hardware
>

PROOF LEDGER

OFFICIAL HISTORY, CURRENT CANDIDATE, CALIBRATED PROJECTION.

Every number below is lane-labeled. Official Rinha hardware remains separate from GitHub Actions and CPU-quota calibration.

[stale official]upstream issue

Issue #5716

p99
0.99ms
failures
0%
score
6,000.00

Closed May 20, 2026, 9:50 PM UTC by the upstream runner. Kept as accepted history, not current candidate state.

open issue
[clean]GitHub Actions candidate

Run 26666573496

p99
0.67ms
failures
0%
score
6,000.00

Main candidate at dc021e25a9a4. CI evidence only, but this is the current clean public artifact.

inspect run
[calibrated-ci]CPU quota projection

api 0.425 / proxy 0.15

p99
0.66ms
failures
0%
score
6,000.00

Projection lane for the same commit and workflow. It predicts pressure, it does not replace official hardware.

compare reports
official source

Latest closed completed upstream issue authored by jonathanperis, synced through gh from zanfranceschi/rinha-de-backend-2026.

Issue #5716 Result comment Ranking
local evidence

326 archived report rows across 157 candidate, 129 calibrated, and 40 experiment entries.

Report archive Latest CI run Source

RUNTIME TRACE

THE HOT PATH IS A SHORT PATH.

  1. 01 ingressk6 / judge reaches :9999

    The public boundary is the yolo load balancer, not ASP.NET Core.

  2. 02 transportfd-pass fanout to two APIs

    The LB sends accepted client fds over Unix control sockets; APIs serve the client socket directly.

  3. 03 parseminimal field extraction

    The app scans enough JSON to build the 14-dimensional fraud vector.

  4. 04 scoreIVF first pass, bounded repair

    The default clean lane uses rounded int16 IVF plus repair; bucket/hybrid modes stay available for experiments.

  5. 05 respondprebuilt bytes out

    The decision returns a small HTTP response without framework ceremony.

A/B TEST SURFACE

THREE LANES TO TEST WITHOUT CHANGING THE BRAND.

A

Evidence Terminal

Default lane. Compare official, candidate, and calibrated proof above the fold.

review evidence lane
B

Benchmark Ledger

Make provenance the lead object: issue, run, image, p99, failures, score.

review report ledger

CI BENCHMARK STREAM

Candidate metrics are local evidence from GitHub Actions. Official hardware is shown separately in the proof ledger.

0.67ms CI p99
0% CI failures
6,000.00 CI score
0.23ms best clean p99
View report history Latest CI run p99 score 3,000.00 // detection 3,000.00