Status: standing operational procedure, adopted 2026-09-12 by ruling 4 of receipts/R7.2.1D-CLOSED.md. It is a precondition and a preparation, not a gate and not evidence. It rules nothing about any result. Applies to: every official performance battery in the ComputeDriven lane from R7.2.1B onward. Battery records cite this file instead of restating it.
Three consecutive attempts at the R7.2.1D-2C battery on this host were refused by their own instruments before a single outcome run was made:
| attempt | host as started | baseline level vs fixed reference | baseline RSD | status |
|---|---|---|---|---|
| 1 | load 3.9–4.1, all-core idle 0.81–0.84 (passed an observational wait) | −34 % | 5.7 % / 6.3 % | NOT EXECUTED |
| 2 | load 6.3, all-core idle 0.76; WebKit, two BEAM nodes, a QEMU guest, a browser | −11.8 % / −17.2 % | 19.2 % / 23.2 % | NOT EXECUTED |
| 3 | rebooted, one terminal, nothing else running | +0.35 % / +1.23 % | 0.24 % / 0.37 % | EXECUTED |
The lesson is in the first row, not the third. An acceptable load average with 81–84 % all-core idle coexisted with a third of the compiled home's capacity gone. Load, idle and PSI are delayed, include work that may not contend, and are blind to contention outside the pinned cores — the four-core witness missed a 13–33 % capacity swing in 2B, and the observational wait passed while capacity was depressed by 34 % in 2C attempt 1. What saw all of it was the sentinel.
Before an official battery:
Reboot the machine.
Open one terminal only. No second Claude or Codex session, no browser, no
VM, no BEAM service, no substantial development workload — including the things that are easy to forget because they are not "running a program": an open preview tab holding a render loop, a paused guest, a language server.
Start the battery from that terminal, with the scratch and target directories
off /tmp and no development knob set.
That is the whole procedure. It takes the host from shared to dedicated; it does not certify that the host is fast enough, and it never substitutes for a measurement.
These are unchanged by this document and remain authoritative:
Baseline level — each policy's baseline mean within its band of a **fixed
absolute reference derived from prior calibration receipts on the same binary**, never centred on the run's own baseline. This is the guard that catches a consistently depressed host, which RSD alone cannot: a host held at two-thirds of capacity can be perfectly stable there.
Baseline RSD — the spread of the baseline sentinels against its cap.
Clipping to the cap is not a pass.
Both-policy sentinel pairs bracketing every block — either policy outside
its tolerance refuses the whole shared block.
Isolation re-applied per block, with each moved process's own mask recorded
and restored, and the restored set verified.
The core witness and the structural witnesses stay active for every block.
Either baseline guard failing means NOT EXECUTED before any outcome run, and the lane stops rather than searching for a favourable window.
It is not a gate. No record may report "rebooted" as a reason a result is
trustworthy.
It is not evidence. The four numbers that made attempt 3 believable are the
two baseline levels and the two baseline RSDs, followed by 21 in-tolerance sentinel pairs — every one of them a measurement taken after the reboot.
It is not sufficient. A dedicated host still drifts: attempt 3's RR-FAIR
sentinel fell 1.7 % monotonically over twelve blocks on an idle machine while CLOCK-AGE stayed flat. Within tolerance, no block refused, cause unestablished — and the reason it is known at all is that the sentinels kept measuring.
The load-average flag is uninformative during a battery. The battery's own
drivers, source and home lift the 1-minute load from ~1.2 to 2.3–2.8, so a "greater than 2× the block-0 start" flag fires on most blocks of a healthy run. It is a diagnostic and never gates; a future record may baseline it after isolation rather than before.
Diagnostics are recorded, never promoted. All-core idle, steal, PSI,
MemAvailable and load are kept in every receipt so a later reader can see the host, and no verdict may rest on them.