Stability
Functional-correctness base Β· runtime gates Β· self-correction Β· adjudication discipline
Β§1Functional-correctness base
| Architecture | Result | Scale |
|---|---|---|
| riscv64 (TCG, 4 CPU) | ALL PASSED | 1584 (2026-11-20) |
| x86_64 (KVM) | ALL PASSED | 1263 (2026-11-01) |
| aarch64 (TCG, virt) | green | 1119 (last recorded; grew since, unarchived) |
- 76 test files across memory / process+scheduler / fs / ipc / signal / network / driver / syscall.
- All assertions run inside the QEMU guest (a kernel crate cannot host host-side
#[cfg(test)]β a no_std + own-panic-path constraint). - CI smoke gate (smoke.yml): the riscv64 suite + proptest on every push; the GUI carries a separate 23-check pixel/serial gate.
- Integration cluster: multi-process unattended runs + probe matrix (integration-test-report.md).
- Formal verification standing (157 Kani harnesses / 1104 proptests / 4 SPIN models / Miri) β see the Security section Β§4.
Β§2Runtime and desktop session gates
- The full GNOME session 2,100-second gate (M1βM5 milestones, zero panic) reproduced across three independent boots (2026-10-27).
- The desktop session held 45 minutes with zero panics (the end-to-end GDM chain, 2026-11-01); boot-to-login ~4.5 s (4 vCPU MTTCG).
- The 22.04 GNOME desktop stability-verified 5/5 (2026-11-16, main @ 1553b75ac, 188 commits past the previous campaign): the campaign's full success criteria (shell up with guest-true timestamps, the desktop frame by screendump, the pointer + Super-key pixel-diff interactivity gate) green on 5/5 boots β shell up 5 s, desktop frame 9.1 s, Super-key interactive 5/5, zero panics/OOM/SIGSEGV; evidence: the gnome-regress work lane's run/GNOME-REGRESS.md.
- The GUI verifier (
test/ubuntu-gui/verify.py): 23 pixel/serial checks per boot; typed terminal conversations pixel-identical across two independent boots (convcheck). - OH side: bootanimation β SystemUI lock screen β the 135+ service set, continuously exercised by the cluster's randomized-testing instances (see Stress).
- riscv64 boot-parameter matrix 16/16: smp{2,4,6,8} Γ mem{1,2,4,8}G β all sixteen cells boot clean to a shell, zero panics and zero hangs (before the matrix was rolled out, several cells panicked or hung).
Β§3Self-correction (census + heal)
- Per-task per-field tick observation + phantom preempt-count adjudication (200 consecutive ticks; legitimate brackets are microsecond-scale) + a 20 ms starvation bound + a softirq dispatch-latency gate.
- Measured: 64.5k/52.1k heals totaling 605 ms, model-vs-reality MISS=0; IF=0 share 25.7% β 20.4%; longest closed cli segment 594β599 ms one-shot β 6β9 ms; WAKE-STRANDED 233 β 31.
- Every qualified quiet boot reads the same census exactly (TERM-DESCHED 8, WAKE-REFUSED 32) β stability itself is a gated standing metric.
Β§4Adjudication discipline
- Flake isolation: new failures isolated-rerun 2Γ before adjudication (host-load timeout/precision families: accept02, getrusage04, kill02, genfmod, rename/statfs); the perf gate follows the same rule (rerun once before calling a regression).
- NOT-OURS ledger: the same binaries run on the host Linux kernel; identical failures are test/environment issues.
- Instrument misdiagnoses archived: PTGUARD/FREEGUARD false positives (mapcount=β1 is the PAGE_MAPCOUNT_BIAS sentinel), the first PT_LEDGER false double-free + a real frame leak, a GRQ race theory that was wrong β negative results with receipts stay in the record.
- The panic surface is designed: every fired panic instance retired with a regression test (the UTF-8 printer family, negative-memo null deref, mmprof slot namespace, and more) β see the Security section Β§6.