01 / DISTROPROOF → OMNISCIENT
A kernel tool proven on 33 targets.
Omniscient’s installed releases ran on 17 distro images on their own kernels, 5 variants, 3 arm64 images and 8 rootless containers. distroproof found 27 defects: 13 in the tool, 14 in its tests. 158 source runs that Omniscient called usable were refused by the kernel at start in the first run, and none after the fixes. The latest run passed all 25 targets (33 cells), and every refusal reason matched envprobe’s reading of the same guest.
$ distroproof up --no-tui --jobs 4 --json target/matrix/distroproof.json
$ cargo test --test portability # the judges, over every cell
02 / OMNISCIENT → DISTROPROOF
The tested tool improved the tester.
The campaign went both ways. distroproof gained 3 images (Gentoo, Debian sid, Fedora 43), 2 boot fixes and 3 [run] options, including file_caps and collect. envprobe gained 2 facts, the memlock limits and the function tracer configuration, because two Omniscient failures needed them to be explained.
[run]
file_caps = { "/opt/omn/bin/omn" = ["bpf", "perfmon", "net_raw"] }
collect = ["cross-tools.json", "witnesses.json"]
03 / DISTROPROOF → XPAC
A package manager tested on 16 images.
xpac follows Omniscient’s installed-release matrix. Its GNU and musl releases are installed on 16 distro images, plus root and lockdown variants. Each guest sends back a cell.json for the host to check. Real installs and removals run only in new disposable overlays.
$ distroproof up --no-tui --jobs 4 --json dist/matrix/inspection.json
$ python3 scripts/distroproof/check.py dist/matrix/inspection.json
04 / XPAC → OMNISCIENT
xpac installs the reference tools.
Omniscient checks its numbers against perf, bpftrace, bcc, strace, sysstat and valgrind. In each chain guest, xpac installs these reference tools through the distro’s own package manager: APT, DNF, Pacman, apk or Portage. Each installation is recorded in xpac-history.json.
05 / OMNISCIENT → VMANYWARE
A real server, measured live.
The chain starts VMAnyware’s server, saves a definition, gets a 403 back for a write from a browser origin, and checks that a second server refuses the locked database. Omniscient attaches to the running server and records mem.rss every 100 ms, without new rights.
$ omn mem.rss --pid $SERVER -m poll@100ms -e timeline,summary -d 1s --format json
06 / ATFQ ↔ PROCFS · MCP
An explanation held to the machine.
In each guest, atfq probe mem.summary must match /proc/meminfo byte for byte. The chain also starts atfq mcp and checks that its five tools are there and still declare readOnlyHint.
$ atfq probe mem.summary --json
$ atfq mcp # lookup, search, toc, run_probe, list_probes
07 / OMNISCIENT → ATFQ, XPAC, ENVPROBE
The read-only promises, held to the kernel.
Omniscient runs each tool and follows every process it starts, from the kernel. atfq’s nine hubs, xpac’s inventory and plan against Pacman, and envprobe --json each created, renamed, removed and synced no file. None opened a TCP connection, and none started a program other than its own. atfq opened no IPv4 or IPv6 socket at all: only netlink, to read the interfaces. All eleven checks passed on 21 October 2026.
$ python3 scripts/devutils/promises.py
atfq PASS no file changed 1464 opens, 8084 reads
atfq PASS no IPv4/IPv6 socket netlink only
xpac PASS only its own programs xpac, pacman, pacman-conf
envprobe PASS no TCP connection 19 ms of cpu
08 / OMNISCIENT → THIS SITE
The front end, measured before and after.
The same 28-second visit of the Design page, three times on the version before the front-end work and three times on the version after, while Omniscient counted the browser’s cpu time and the graphics card’s energy. It refused to trace the cpu time on this machine and said why (tracefs needs CAP_SYS_ADMIN), so the run counts it instead. The results are just below.
$ omn -f bench/site.yml --format json --program-stderr visit.err > visit.json
