# Artio shared-harness phased Session qualification — 2026-10-25

Both hosts pass independent replay of 22 original-SDK pairs. Each preserves
2,720 application frames, 272 administration frames and 696 exact values in
each of the primary and scheduler-delay series. Eighteen CLI refusals and
36 corruptions of actual output pass per host. All 89 existing SDK/class files
stay unchanged. All 77 newly compiled producer/common-harness files match
between hosts. The current Artio and generation source suite passes 547 tests
without selected skips. This is functional phase qualification; it exports
zero result cells and grants no installed-runtime or physical timing admission.

## Exact sources and successful replay

The production/capture source is
`b3380f0b5d77be52bce33c3002939ea13c62166f`. The corrected inventory reader and
separate replay worker are
`5da089a5118b90dd4b2c6d9d3e9e1e602fe67281`.
The new public producer is `ArtioTimedSession.java`; it implements the unchanged
shared `BenchTask` interface and executes the original shared `BenchRunner`.
The separately compiled 73 common-harness classes use read-only source inputs
and the original SDK's HdrHistogram 2.2.2 JAR. No libhft engine implementation
is linked. The producer adds four class files, including the Java 8 compiler's
synthetic nested-constructor access token. Their combined 77-file relative-path
and SHA-256 map has canonical digest `ff223823df8ec24a34f80733d6415ffb2306bc3663c33961861002351b788df9`.

All 26 original SDK/runtime files and 63 qualified parent classes remain
byte-exact before and after on both rigs. The latter retain the original numeric,
Session, clock, bounded transport, finite cadence and pressure qualifications.
Actual class-load logs bind every exercised generated business encoder/decoder,
Session, GatewayPublication, original dictionary/decimal/scheduler/TCP supplier,
normal bounded runtime and prepared-offer helper. The measured owner loads the
actual shared runner, configuration, context, results and JSON writer. The
pressure runtime, its scheduler/channel controls and minimal generated dictionary
are forbidden in this normal producer's class-load evidence.

## Phases, owner clocks and original administration

Every pair starts with twelve canonical preflight cycles and original bilateral
Logon/TestRequest/Heartbeat. Warmup and measured markers reset at their own
boundaries, while SDK sent/received sequence numbers continue. Separate complete
typed owner sinks/fold counts are checked for every preflight, warmup and measured
group, including explicit empty groups with zero warmup.

Eight pairs cover all four flows with zero/five warmup cycles, two inner runs of
24 measured tasks and no/measured administration. Ten pairs put real two-ms
scheduler, construction, serialization, extraction and completion probes around
NOS/ER and tick tasks. Four changed-book pairs exercise both clamp boundaries,
repeated runs, and both ordinary and busy-spin transport configurations for
mixed and tick flows. All use the original 10,000/s offered profile.

The shared runner includes the first pacing interval and bases each subsequent
measured deadline on the previous actual task start. Warmup follows its original
absolute-epoch loop and is discarded. The separately qualified finite
absolute-epoch diagnostic remains unchanged; it supplies no measured pacing
policy here. Every measured primary and delay value is compared to both the
original shared raw snapshot and the complete owner-clock rows, with no dropped
or clipped values. Request/reply spans include construction through complete
typed extraction. Tick starts at the accepted prepared-buffer attempt and stops
at the D callback before extraction. Diagnostic probes confirm those boundaries.

Each repeated run can insert a real TestRequest/automatic Heartbeat outside the
primary task but within its measured window. A separate original finishing echo
happens after the shared harness completes, followed by complete original Logout.
The shared JSON writer retains actual measured duration/throughput, probe counts
and configuration. Its measurement timer starts after telemetry setup; the
pacing epoch cannot reconstruct that timer. Independent checks bound duration
by actual task clocks and verify throughput arithmetic.

TCP copying closes after the twelve-cycle preflight prefix in each direction.
The independent reader checks the full prefix's original frames, all phase
owners, continuous SDK sequences and complete original connection byte totals,
including phase administration and shutdown. Timing flags stay false throughout.

## Retained failures and correction without producer reruns

The initial source-test command omitted the required explicit benchmark root;
the corrected full suite passes. Original V1 host attempts stopped after their
first pair because an independent duration assertion reconstructed measurement
start from the earlier pacing epoch. Their captures remain retained. The
corrected timer check has an explicit regression test.

Both V2 host workers then complete all 22 pairs and all eighteen CLI controls,
but their final inventory guard allowed only three source-declared classes.
The original Java 8 compilation also emits `ArtioTimedSession$1.class`. Both
original V2 worker statuses remain failed; neither receipt nor capture is edited.
The corrected reader requires exactly those four producer classes and refuses
extra or missing ones. Actual `javap` confirms the access token is synthetic,
major version 52, with zero fields and methods.

Separate sealed replay jobs copy the exact receipted bytes to fresh review
roots, replay the complete inventory/class/CPU/phase/raw gates, and run all
36 actual-output refusal controls. They execute zero producer pairs. Both
replay services finish successfully with status zero and no remaining process.
Independent local collection/replay agrees with both native cold proofs.

## Compact receipt bindings

| Host | Original native receipt SHA-256 | Input manifest SHA-256 |
|---|---|---|
| HP | `1e8a97547d6b3e391a1895ec923859bf4aee37effd9a9add25928a1d96fe8c5f` | `86dfa45587682655bc3c04554b05dd61053c31263c282eceb4d8266b5cb689ca` |
| SCAN | `ac1a32b834c4411313c5a0bfef418fb640b1068ea11eccca7fba48a85ed3d4b0` | `3e9f54599320edb8eef1a83953de7f37f90c7f3b81b6566d424746e332a6a6f6` |

| Host | Successful replay receipt SHA-256 | Collected review index SHA-256 |
|---|---|---|
| HP | `4bba86ca440ef6017c0f7e38d91c6375c2e1f6beb053db2b7a72d2d084565180` | `f440971156f59280da7789592ee30198f4f94b461b2697d1ff56aa8b2e0f42b9` |
| SCAN | `b8310823e18b8b676bdd17aa45f6e56a8ed1d3d0bf15885a934ad5478b06f64e` | `c545b2312a8e514a59c2b273730775ed1ce0c2dc9d3a336a4f08c1eb678fe60d` |

Original runtime/captures remain under
`~/build/codex-fullbench/libhft/artio-phases-qualification-20261025-v2`.
Fresh replay capture roots are
`artio-phases-replayed-qualification-20261025-v1/capture`; independently
collected roots are `artio-phases-replayed-review-20261025-v1/{hp,scan}`
under the same build root. The published retained
`artio_phases_replay_20261025.py` stages, replays and collects once; it never
reruns an unchanged producer. Earlier failed V1/V2 evidence remains distinct.

Normal immutable generation packaging is implemented separately, with source
refusal checks, using the existing generic normal-user helpers. Relocated and
installed DAC runtime qualification, fresh quiet physical collection and all
remaining full-plan admission/release gates remain open. Existing results and
older Artio grades are unchanged.