# Artio: original shared-harness physical DAC qualification

## Outcome

Every report flow remains **S5 diag** on both rigs. The original full-size campaign is complete, with zero admitted cells and no unchanged timing rerun pending.
The original 10,000/s offer, 9,800/s minimum, 100 µs scheduling-delay p99 bound
and original run-to-run percentile gates remain unchanged. All four twenty-sample
functional cases pass on each host. No complete matched two-host Artio comparison
is admitted by this campaign.

Each host finishes sixteen pairs: four functional and three report repeats per
flow. Report repeats use 1,000 warmup and 10,000 measured cycles. Both the initial
physical reader and independent admission replay every one of the 120,080 primary
and 120,080 scheduler-delay values per host, including the functional runs. The
values below are median run percentiles and aggregate achieved rate. Diagnostic
numbers preserve failed gates and are not report-grade comparison cells.

| Host | Flow | p50 µs | p99 µs | Achieved cycles/s | Admission or blockers |
| --- | --- | ---: | ---: | ---: | --- |
| HP | nos_er | 70.637 | 227.313 | 8,590.79 | achieved_rate_below_98pct; scheduler_delay_above_interval |
| HP | md | 79.587 | 240.619 | 8,723.62 | achieved_rate_below_98pct; scheduler_delay_above_interval |
| HP | md_nos_er | 128.067 | 286.560 | 6,996.26 | achieved_rate_below_98pct; scheduler_delay_above_interval |
| HP | tick_to_trade | 56.569 | 162.278 | 8,579.07 | achieved_rate_below_98pct; full_percentile_stability; scheduler_delay_above_interval |
| SCAN | nos_er | 111.329 | 329.657 | 7,056.30 | achieved_rate_below_98pct; full_percentile_stability; scheduler_delay_above_interval |
| SCAN | md | 131.607 | 312.271 | 6,497.13 | achieved_rate_below_98pct; full_percentile_stability; scheduler_delay_above_interval |
| SCAN | md_nos_er | 216.137 | 424.375 | 4,225.79 | achieved_rate_below_98pct; scheduler_delay_above_interval |
| SCAN | tick_to_trade | 92.885 | 247.552 | 7,163.93 | achieved_rate_below_98pct; full_percentile_stability; scheduler_delay_above_interval |

## Exact producer, phases and runtime

The [qualified normal generation](RECORD_2026-10-25_artio_normal_runtime.md)
is reused byte-for-byte. Its 28 payloads contain the original 26 SDK/runtime
files, the fixed normal-user launcher and deterministic stored producer JAR.
All 119 class hashes and payload bindings match across rigs. The SDK remains
Artio 0.182 at declared source `5d1bb3b2deed9c5865e8584d23bb37c48fde048a`.
No Java or SDK rebuild, original helper change or new sudo registration occurs.

The unchanged original `BenchTask`/`BenchRunner` controls measured phases and
actual-start pacing. Achieved rate uses the actual `measuredDurationNanos`;
neither the pacing epoch nor warmup/run duration substitutes for that timer.
Request/reply begins before fresh request construction and ends after final typed
extraction/checks. Tick begins at accepted prepared-buffer publication and ends
at D callback entry before extraction, in the same process nanoTime domain.
All phase sinks, marker resets, SDK sequences and administration counts pass.
Only the twelve-cycle directional preflight is copied as FIX text; measured
applications are not formatted or copied into the bounded transport capture.

HP uses the Solarflare DAC namespaces with server mask 8–12 and client 13–17;
SCAN uses mlx5 with masks 10–13 and 14–17. The library owner uses the first CPU,
Framer the second, media-driver the third and other runtime threads the last.
Every one of the seven actual thread placements per role is checked. Normal
UID/GID, namespace inode, JDK hash, immutable class sources, fixed-entry logs,
whole bilateral transport bytes and advancing physical transmit counters pass.
Before/after endpoint host facts agree. Both runs hold `isol-no-shared`, HP at
high priority and SCAN normal, after slot/process/memory preflight.

## Independent admission and limits

The new cold reader reuses only immutable independent expected FIX body strings
for twelve-marker reference classes. Dynamic message identifiers, every observed
frame, typed owner, transport byte count, phase and raw clock are checked anew.
The running original physical capsules retain their original checker unchanged;
no producer or timing pair is rerun to accelerate the later unmeasured replay.

Per host, seven altered physical metadata cases and 64 corruptions of actual
copied functional output are refused: counts, false promotion, phase resets,
raw primary/delay values, end clocks, owner sinks, FIX prefix, UID, namespace,
CPU, executable, JDK hash, actual library-thread CPU and class source. Originals
are never edited; temporary per-case copies are removed at exit. Every admitted
cell additionally refuses a well-formed false empirical-CDF point. Both physical
and admission services finish successfully, exit zero and have no remaining
main process. Independent bounded collection confirms the same decisions.

The writer admits only S5-pass flows with the exact three full original runs,
original rates and generation bindings. Any admitted cell retains a pooled
empirical CDF bounded to 2,000 points per kind, with median-of-runs summaries.
Failed cells never replace good keys. Raw series and full clocks stay in the
producing `~/libhft-bench-runs/artio-physical-20261025-v1` folders and remain
subject only to the result-aware pruner. Website delivery exports compact
summaries with zero full per-test CDFs; no live OVH upload is attempted here.

The focused source suite passes 565 tests with zero selected skips. The initial
writer-only run retains a filename-length failure; the next full suite retains
one capture-count failure with 561 other tests passing. The corrected writer
uses a safe flavour name and exports only the first ten of twelve independently
verified preflight cycles. Native and exported counts/scope stay explicit, and
no result-format restriction or running capture changes. The first
physical staging attempt failed locally before any host/capsule/run creation
because of a missing `normal` component in the parent-capsule name. The one-line
source correction is retained; no native or physical pair was repeated.

This completes this exact Artio source campaign. Java/.NET matched minima,
wider roster/every-flavour qualification, QuickFIX-family anchors, llfix store/
recovery interoperability and public archive/release gates remain separate.
No source-specific diagnostic campaign completes the full plan.

## Exact pins and receipts

| Item | Commit or SHA-256 |
| --- | --- |
| Qualified normal runtime source | `aa40e2cf559f8402ab0864676fa70ee1bc32cfc0` |
| Physical controller source | `29bfb6646d5f2e2ef8f26ec19184eefa2df55cc0` |
| Independent admission source | `33aa9436bc9de4441a9c07afdd610122273a2837` |
| Both immutable installed generations | `db8c84ad5c3616185a38741b8dc2f582766ad098660c8fdad1b0cdf474b23d9b` |
| Both producer JARs | `ae0034e3f9ffc0e5763fc097013a66786e5cd8c006fc8c1d225d92a8eee9cee8` |
| HP physical input manifest | `2b9c9741ed21361b42d42c654a9de0531da74dba2979d2b50023223b21f9cfec` |
| HP completed receipt | `53136d13cfa4cc15dc41b3472843c3d9180296c00ce5b8dfe566a1b6f0fcf3f1` |
| HP independent original-value replay | `80a9166f9b6e7e37cca4d8c35ffbdfdab31cefe5ae4392f8729662dff9f5d44f` |
| HP admission receipt | `442ff0aa3f475ab0cfd6c96e8582a8600e664789c5ef6af80c53cbd621d80f0f` |
| HP collected review | `eb11f11eafed0d8a8b67d8287dfa886f4ed72f1be14b5c00803e1778578e2835` |
| SCAN physical input manifest | `ce3065c767796844aa8ccdba3a3b7197b3333297403468ee9167c7a894fb3c90` |
| SCAN completed receipt | `26f8093f6d6e44dfbf06fac541d55b317a42c6fa6daa739e95cb12debaf96945` |
| SCAN independent original-value replay | `14b4bf371ef815b6972b746367427ce440b505d192987900b8ee6e401d8f80f7` |
| SCAN admission receipt | `20042332eed458cf5074167690ce6f32440f3f8c0efdd47e4f8b6e3c85a5f923` |
| SCAN collected review | `38373adc9924715d8bdf0e9a28fcbcc3c498a8854d11110cd70b2f10d67d25af` |