Transport stays out of the classifier.
The yolo LB accepts on port 9999 and passes accepted descriptors to the two C API processes. Fraud payloads stay in the API path.
RINHA 2026 // PURE C // OFFICIAL ISSUE #4545
0% FAILURES.
FD PASSED.
FRAUD SCORED.
Pure C fraud scoring under the official 1 CPU / 350 MB envelope. The page now leads with verifiable result evidence, then opens the hot path.
- official.p99
- 1.44ms
- failures
- 0%
- score
- 5,843.05
- budget
- 1 CPU / 350 MB
official source
Verified upstream result, not a decorative counter.
Latest closed completed upstream issue authored by jonathanperis whose result JSON matches https://github.com/jonathanperis/rinha4-back-end-c. Synced from zanfranceschi/rinha-de-backend-2026 and closed May 15, 2026, 2:12 PM UTC.
PROOF LEDGER
Official proof first, CI signal second.
Numbers stay split by provenance so the page does not mix public Rinha evidence with local candidate tracking.
- official.p99
- 1.44ms Rinha runner, issue #4545
- official.failures
- 0% 0 FP, 0 FN, 0 HTTP errors
- official.score
- 5,843.05 p99 2,843.05 plus detection 3,000
- resource.envelope
- 1 CPU / 350 MB LB plus two API containers
SYSTEM CUTAWAY
TCP in. FD passed. C writes back.
k6 / judge
|
v
rinha4-lb-yolo-mode :9999
| SCM_RIGHTS over SOCK_SEQPACKET
+--> api1.sock -> C epoll loop -> client fd
|
+--> api2.sock -> C epoll loop -> client fdThe API parses only what the workload needs.
Manual HTTP and JSON extraction produce the 14-dimensional fraud vector without framework overhead.
The index answers with prebuilt response bytes.
A mmaped int16 IVF/block index returns the fraud label count, then the server writes one prepared HTTP response.
Candidate telemetry with report links attached.
CI candidate metrics are local evidence for regression tracking. The latest candidate uses the canonical fd-pass compose layout and keeps the candidate pointer separate from experiments.
0.36ms latest CI p99
0% latest CI failures
6,000 latest CI score
50/60 clean candidates archived
3 experiments separated