BACQM Systems Research Group · Instrumentation

Measured, not assumed

Every figure on this site comes from instrumentation at the kernel–hardware interface, under real work. Differences are taken against a baseline that is the same kernel with nothing of ours loaded (E1); shapes are taken on the deployed system itself (E2–E4). The probe never adds load to flatter itself: what the system does while it is observed is what it does.

E1A/B campaigns against the baseline

Protocol. CONTROL is the unmodified kernel with its default fair scheduler. Each campaign alternates CONTROL and BACQM on the same node, n≥3 runs per arm, and reports the median. A campaign counts only if the thermal state stays within 5 °C across its runs; two campaigns failed that test and are excluded.

QuantileBACQM against CONTROLBasis
p99 · p999 · max−90 % to −99 %8+ A/B campaigns, n≥3 each
p50+10.5 % to +13.8 %same campaigns

The median rises a little and the tail collapses. That is a declared Pareto choice of the scheduler's architecture, not a tuning artefact: the tail is where a saturated system fails.

E2Passive telemetry, weekly

Probe. Every 30 minutes, under normal use and with no synthetic load, a 5-second probe measures the wakeup delay of a periodic 1 ms thread. 2,582 samples with the scheduler active, 15 Aug – 14 Oct 2026.

Wakeup delay, weekly medians

µs · passive_weekly_medians.csv
p50p99p999
Table view
ISO weeksamplesp50p99p999maxmissed
IntervalSamplesp50p99p999maxMissed cycles
15 Aug – 26 Sep1,79183.3 µs119.8 µs317.0 µs513.9 µs0.080 %
27 Sep – 14 Oct79187.3 µs259.9 µs415.0 µs559.6 µs0.013 %
The step of 26 September · non-invasive measurement

Until 26 September a system thread woke every 1 ms, the same period as the probe, and that favoured the probe. Since then it sleeps when there is no work and no longer perturbs the system. An A/B on 15 October with the same module confirms it: with the 1 ms pulse the probe reads p99 = 159 µs, without it 280 µs. We do not perturb the system to improve a measurement, so the second interval is the one that stands; missed cycles fell from 0.080 % to 0.013 % across the same step. These figures describe the probe, not a workload.

E3Saturation of the fast domain

Workload. A real-time graphics pipeline whose resident set exceeds the fast domain, run as a user plays it: seven runs on 23–24 Oct 2026, 425,860 frames, 133 minutes. It is chosen because it drives s=‖σ‖𝒱/C𝒱 to the boundary (maximum 0.9984) while the accelerator runs near full duty: the regime in which (1.1) predicts a discontinuity.

Observable. The frame period τf, recorded per frame by an in-process overlay, in units of the presentation period T0=1/60 s. Occupancy, accelerator duty and host load are recorded on the same frame and enter only as ratios. Arm. Single arm, BACQM active: this campaign tests the shape of qp(s) near s=1, not a difference against CONTROL (that is E1). Kernel log over the seven runs: no panic, no out-of-memory action, no guard termination.

Frame period per occupancy stratum

τ/T₀ · saturation_strata.csv
q50q95q99
Table view
Stratum of sframesq50q95q99q99.9τ > 3 T₀host load

The median does not move: 1.01–1.04 T0 in every stratum. For s≥0.5 the p99 stays at or below 4.85 T0, and the top stratum s≥0.98 has a lighter tail (p99 3.55) than the one beneath it (4.84). The only heavy tail is at s<0.5, where the frames are asset loading and the fast domain is nearly empty. This is Observation 4 of the formal note.

One run, 15-second windows

saturation_windows_15s.csv
s = σ/C𝒱accelerator dutyhost loadq50q99

Table view
t (min)framessdutyhostq50 / T₀q99 / T₀

The shaded band is s≥0.95. When occupancy and duty rise together the median frame period rises with the duty (Observation 5: 1.32 T0 on the 31,222 frames with s≥0.95 and duty ≥ 0.90), while the host load stays at 0.20: the pipeline is bound by the accelerator's compute, not by the boundary of its memory. Within single 15-second windows the p99 can leave the scale (▲): these are the rare long frames that the p99.9 column counts, finite and isolated, never a sustained stall.

Run (local start)framesminutesmax sq50 / T₀q99 / T₀q99.9 / T₀host | duty ≥ 0.9

E4The transport ledger

Cumulative counters of the transport operator 𝒯 of §4, read at the kernel interface on 24 Oct 2026. Counts are in granules, the unit the operator reads from the running system; requests alternate direction.

Directionrequestedcompletedrejected
𝒱 → ℛ (evacuation)6306300
ℛ → 𝒱 (admission)6306264
Total1,2601,2564

Requested mass equals completed plus rejected mass to the byte; the net transport Δ=630−626=4 granules is exactly the rejected set, as Proposition 3 requires, and every rejection was an admission into a fast domain at its boundary. Completion 0.9968, imbalance 0.32 % of the transported mass.

E5Extreme-entropy asymptotic allocation

Workload. An unconstrained symbolic proof-resolution pipeline executing deep recursive dependency search over an extensive axiomatic corpus, operating continuously for >60 minutes (2 Nov 2026). The active resident set exceeds 20 % of total host memory (>6.5 GiB in a single task tree), accompanied by concurrent asynchronous dirty-page writeback workers.

Observable. Run-queue scheduling delay (p50,p90,p99,p99.9) sampled at 60-second intervals via non-invasive eBPF scheduler probes (sched_ext with the nava operator manifold), memory pressure stall duration (PSImem), and directional page transport flux (pswpin,pswpout).

Sustained symbolic saturation: run-queue latency and paging flow

60 s sampling · kernel telemetry
Epoch (UTC) Load 1m p50 p90 p99 p99.9 pswpout (pages) pswpin (pages) PSI mem Attractor Zone
03:48:45 (Basal) 0.25 4 µs 18 µs 42 µs 2,621 µs 0 70 0.00 ms POTENCIA_PLENA
03:58:45 (Allocation shock) 6.08 2 µs 17 µs 85 µs 421 µs 41,772 998 320.07 ms POTENCIA_PLENA
04:01:45 (Entropic evacuation) 2.17 1 µs 25 µs 147 µs 2,003 µs 124,137 6,417 704.27 ms EIGENSTATE
04:03:45 (Memory peak) 7.31 3 µs 78 µs 539 µs 1,694 µs 5,363 733 1,069.68 ms EIGENSTATE
04:10:45 (Compute peak) 9.25 12 µs 920 µs 2,694 µs 4,011 µs 11 1,620 477.06 ms EIGENSTATE
04:20:45 (Bifurcation settling) 5.54 21 µs 1,402 µs 4,608 µs 7,168 µs 320 748 495.86 ms EIGENSTATE
04:24:45 (Asymptotic quiescence) 4.50 4 µs 720 µs 4,018 µs 7,680 µs 0 5,773 0.06 ms CONVERGIENDO
04:36:45 (Equilibrium) 5.10 4 µs 812 µs 7,680 µs 10,240 µs 0 536 0.02 ms EIGENSTATE
04:46:46 (Stabilized) 1.60 4 µs 20 µs 7,680 µs 13,312 µs 0 294 0.03 ms POTENCIA_PLENA
04:52:46 (Steady state) 2.33 2 µs 15 µs 521 µs 9,216 µs 0 7,445 0.30 ms CONVERGIENDO

The 7-collector bundle partition. Paged allocations are uniformly distributed across k=7 equal-precedence compressed collectors (π=100, cyclic symmetry ℤ7):

Collector Bound Raw Volume Compressed Footprint Compression Ratio Precedence
ℛ₀4.0 GiB194.8 MiB72.2 MiB2.70 : 1100
ℛ₁4.0 GiB192.0 MiB54.0 MiB3.55 : 1100
ℛ₂4.0 GiB183.7 MiB52.8 MiB3.48 : 1100
ℛ₃4.0 GiB179.7 MiB68.3 MiB2.63 : 1100
ℛ₄4.0 GiB190.0 MiB69.7 MiB2.72 : 1100
ℛ₅4.0 GiB192.4 MiB70.7 MiB2.72 : 1100
ℛ₆4.0 GiB194.1 MiB73.0 MiB2.66 : 1100
Aggregate28.0 GiB1,326.7 MiB460.7 MiB2.88 : 1—
Dynamic invariance under extreme saturation · non-heuristic regulation

1. Microsecond median under load 9.25: Conventional schedulers experience priority inversion and tail explosion under memory-induced writeback stalls. Here, the scheduler's median dispatch delay remains locked between 1 µs and 12 µs throughout the entire session.

2. Asymptotic zero-evacuation: Following an initial burst of 1.24×105 pages evacuated into the 7-way bundle, outward swap drops to strictly zero (pswpout=0) for over 30 minutes, preventing any swap-thrashing stall.

3. Zero-OOM safety: The invariant memory guard recorded zero liquidations (kills=0, danger=0), and the uncertainty bound remained certified on every tick.

E6Substrate

The implementation lives in kernel space and names no device, no vendor and no size: every bound it uses is read from the running system. One source, two instruction-set architectures.

ArchitectureState
x86-64continuous operation; every measurement on this page
ARM64modules loaded and verified, 21 Sep 2026
Multi-node fabricpilot stage: the law is the same, the measurement is next

Kernel source is private. Methodology, data and the mathematics are open. For a pilot, open an issue.