# HP clock baseline before a future turbo comparison

2026-10-11. User deferred enabling turbo and requested continuation on HP only.
This diagnostic changes no firmware or OS setting and launches no FIX peer.
The retained Stage 2 ranking, physical llfix results and all old runtime
source generations keep their original clocks and grades.

## Exact source and actual execution

Source `dc5b362f47480f53898aa88699e5fad59d48e581` is preserved on
`s/codex-llfix-reference-hp-clock-source-20261011`. The five-file capsule SHA256
is `6b033a505d6a40dfd209a38f37e9d64f3904057eb5d4f05b5bda888e0ae9eca3`.
Four new private source/profile/test files and the unchanged reference lease
helper are bound; no shared clock, build, toolenv, layout or measurement API changes.
[Profile](../../bench/llfix_reference/diagnostics/clock_controls/PROFILE.md).

HP's actual GCC 11.5 C++17/O3 build passes with warnings treated as errors.
Five focused tests, including thirteen malformed native vectors, pass with zero
skips. A real high-priority isolated lease (shared work permitted) captures
48 complete windows: CPUs 16/17 sequentially, twelve sleep and twelve busy
100 ms intervals each. The observer stays on CPUs 0-7. Both actual native
controls refuse a wrong affinity and an invalid CPU argument with exit 1 and
the expected diagnostic; no timed data is fabricated for those controls.

The cold controller audit and separate integer/Fraction recomputation both
pass under a high-priority shared lease. They verify exact native count/order,
CPU/TSC_AUX, prepared binary/compiler/source and output hashes, bracket bounds,
stable sleep/busy clock scale and unchanged before/after policy.

## Results

| CPU | Window | Samples | Median TSC GHz against RAW | Median upstream time / calibrated time |
| --- | --- | ---: | ---: | ---: |
| 16 | sleep | 12 | 3.092249712 | 0.997499769 |
| 16 | busy | 12 | 3.092248897 | 0.997499657 |
| 17 | sleep | 12 | 3.092249726 | 0.997660724 |
| 17 | busy | 12 | 3.092248872 | 0.997499636 |

`CLOCK_MONOTONIC_RAW` calibration is approximately **3.09225 GHz**, agreeing
between CPUs and sleep/busy modes. Largest per-mode relative frequency spread
is 0.00004295 percent; largest bracket uncertainty is 0.00025986 percent. Both
remain below the unchanged 0.1-percent spread and 0.01-percent bracket gates.
CPUID leaf 16 reports base 3100 MHz / maximum 4000 MHz. Leaf 15 has no crystal
frequency, so a complete independent CPUID TSC-rate check is explicitly absent.
The captured kernel clocksource is **tsc**: this establishes Linux RAW nanosecond
conversion independently of llfix's CPU-MHz divisor, not an external oscillator
or independent physical clock measurement.

The first `/proc/cpuinfo` MHz during this diagnostic ranges from 3092.734 to
3102.996. Each observation retains llfix's integer truncation. The median sampled
ratio is about 0.9975: an upstream-style divisor would typically understate elapsed
time by approximately **0.25 percent in this context**. That magnitude cannot
explain the retained approximately 12-percent physical median difference by itself.
These sampled MHz values are not the historical run's original scale, and no
historical sample, result or vendor number is rescaled or regraded.

## Current policy and limits

Before/after policy is byte-for-byte equal within the capture. `no_turbo=1`,
`turbo_pct=0`, min/max performance percentages 100/100, performance governors,
and 3100000 kHz ceilings remain as observed. The boot journal states
`Turbo disabled by BIOS or unavailable on processor`. No sysfs write, root
command, firmware change, reboot, SCAN connection or other-session job action
was performed. Kernel `scaling_cur_freq` readbacks are retained but are not
APERF/MPERF or sustained core-MHz measurements.

This is clock provenance, not quiet latency, a vendor reproduction, NIC/Onload
qualification, SDK admission or a fresh benchmark ranking. Shared work was
permitted, and the busy loop is not the llfix workload. After the operator enables
turbo later, capture a new clock baseline and record actual effective core MHz
alongside a separately queued identical-workload latency comparison. TSC MHz and
the firmware maximum cannot stand in for the effective core clock.

## Retained evidence

[Selected exact outputs and operations](evidence/hp-clock-baseline-20261011/selected-files.json)
contain all native integers, capture/build/control/cold-audit receipts, source
bindings and original kernel/frequency readback. The whole original generation
also retains the binary and five frozen source files on HP and locally:

- `~/archive/libhft/codex-llfix-reference-20261011/hp-clock-baseline1.tar.gz`
- 38 sealed files; 66,140 compressed bytes.
- Archive SHA256: `ca1b94bec893abbe40cf30cd6bfb184ac753cdb0cfeb3359c5c6c3c1d3a67247`.
- Seal SHA256: `71e49a08aa85009443d2fe4f2865b0f195a66f2ea6f17c5944f8b56f10302439`.
- Both independent streaming replicas verify every file hash, mode, size and membership.

Disposable owned source/build/operations are removed only after those verifications
and the exact seal/member/visible-consumer guard. Cleanup receipts are retained
with the original archive and selected metadata. No other session's path is removed.
The original cleanup guard refused the archive's root directory before any
deletion. Its script and failure log remain intact; a separate continuation
accepts that directory and verifies all 38 members before successful cleanup.