# hp tuned collection — 2026-09-20

## Outcome

The full hp tuned matrix completed as **collection-only evidence**. It is not a
promotion or public-result update. All 84 cells completed with consistent recipe
provenance; 48 were sustained and 36 were partial-load results. No cell failed.

The collection used the cell-v4 schedule: three 30-second measured windows after
five seconds of warm-up. It initially retained only local evidence. On 2026-09-20,
the owner explicitly requested that exact, recipe-audited matrix update the hp v4
source table and its generated HTML and handbook. That presentation does not change
the profile-free deployment baseline or remove either profile's `REJECTED` guard.

## Sitting

- Local evidence directory:
  `/home/yann/libhft-bench-runs/open-v4-hp-cascadelake-gcc14-20260919-tuned`
- Sweep state: `complete`, `collection_only=true`, 84 results and 84 recipe files.
- Source commit recorded by every recipe: `162ca9d2`.
- Profile recorded by every recipe: `hp-cascadelake-gcc14-20260919`.
- Profile fingerprint recorded by every recipe: `20105595833209`.
- Native marker recorded by every recipe:
  `tuned_build=1;checksum=FX_CHECKSUM_VPADDB_MASKTAIL;find_byte=FX_FIND_BYTE_AVX2;memcpy_tiny=FX_MEMCPY_TINY_DESCEND;memcpy_medium=FX_MEMCPY_MEDIUM_HEADTAIL`.

Status by topology:

| Topology | Sustained | Partial | Failed |
| --- | ---: | ---: | ---: |
| Single | 26 | 0 | 0 |
| Mono | 9 | 17 | 0 |
| Pool | 8 | 12 | 0 |
| Shard | 5 | 7 | 0 |
| **Total** | **48** | **36** | **0** |

## Why this does not approve the profile

The 2026-09-19 repeat emitted a byte-identical prescription to the candidate
rejected in the hp CPU-upgrade record:

- `libhft_tuned.h` SHA-256:
  `4bb98adbc2c761500a90d6858b03e85639fe1a7a8aa377e34518179be9c5fb54`
- `libhft_tuned.mk` SHA-256:
  `24823212597f7f13144a3d72361df9280ad75328f1b0613d42b7b03d510143a9`

The earlier real-FIX in-context gate rejected that exact prescription with a
2.73% median component regression and 83 material regressions. See
[the CPU-upgrade record](RECORD_2026-09-18_hp-cpu-upgrade.md) and
[its tuner evidence](evidence/hp-cpu-upgrade-20260918/TUNER.md).

The repeated profile is therefore marked `REJECTED`, along with the original
candidate. The hp machine configuration declares no `TUNED_PROFILE`; normal
native builds reject either profile unless an explicit diagnostic override is
given. The installed benchmark artifacts were rebuilt and restaged profile-free
after this collection, and their native markers report `tuned_build=0`.

## Future use

The collection answers the explicit question of how the tuned prescription
behaves across the complete matrix. It must remain separated from normal hp
baseline results. A future tune may be considered only if its generated header
or makefile changes materially, after which it requires a new in-context gate
before staging or public comparison.