← MRM/Projects/Small tools
Tooling / Utilities Six shipped · in daily use Rust · one Node CLI

Small tools

Six utilities I built because the thing I needed to measure wasn't measurable.

None of these started as products. Each one exists because a question came up in the middle of a job — is this site actually accessible, where did my RAM go, what is eating my mouse clicks — and the honest answer was that I couldn't tell. They are small, they are finished, and every one of them ships with the thing that would catch it lying. Two have public results you can check yourself: the auditor's fixes are live on torsova.com and lazarus-systems.com, and scrub has its own site at scrub.theggnetwork.com.

ada-audit — a clean result, and the control that earns it
$ node audit.mjs https://lazarus-systems.com --crawl 50 --real-ua --settle 2500 ✔ 3 pages · 0 violations · 0 errors · exit 0 $ node audit.mjs ./pre-fix-copy --settle 2500 known-positive control ✘ 3 pages · 0 clean · 31 serious · exit 3 the control fires on exactly the rules the fixes address ⇒ the scan reached the pages, and the fixes — not the measurement — changed the result

Representative output — not a live instance.

The tools
●ada-audit — crawls a whole site in real Chromium and runs axe-core against WCAG 2.0 and 2.1 level A and AA, then writes a report a non-developer can act on: every issue, who it affects, how to fix it, and every page and selector it appears on. Its oracle: the same binary scanning a copy of the site with the fixes reverted, which flags the 40 serious violations the clean run doesn't. Used on torsova.com and lazarus-systems.com — 29 of 29 pages, zero violations, re-scanned against production rather than local files.
●ghostbox — finds the invisible window that is swallowing your mouse clicks, whatever created it, and hides it reversibly with a hotkey. It requires two independent facts before calling anything a ghost, because either one alone produces false positives. Its oracle: it can generate a genuinely invisible window on demand — that known-positive is detected at 100% mismatch across 144 sampled points, while a real desktop reads 0 to 6%.
●memhog — answers "where did my RAM go", including the kernel memory no per-process view can show. Its discipline: it reports three numbers rather than one — what the processes explain, what the OS says is in use, and the gap — then resolves the gap into nonpaged pool, paged pool and system cache, which belong to no process at all. On this machine that's how a 4.29 GB nonpaged pool turned out to be a driver, not an app.
●scrub — finds and removes encoded watermarks and metadata from your own AI-assisted files: invisible characters, lookalike letters, exotic whitespace, image and document metadata. Its oracle: it re-scans the bytes it just wrote and reports what survived; when it cannot verify removal it says so and offers the harder pass, instead of reporting success.
●desk-probe — captures any window on the real screen to a PNG, so I can look at what is actually in front of someone instead of asking them to describe it. Its oracle: a self-test that captures, saves, re-decodes and checks the dimensions, so a silently empty capture fails loudly rather than returning a black rectangle.
●imgredact — blacks out and crops regions of a screenshot so evidence can be attached to a report without carrying live session values with it. Its guard rails: it refuses to write when there is nothing to redact, and refuses if every box fell outside the image. The failure it exists to prevent is silently emitting an unredacted copy.
The failures they were built to catch
●A blocked scan reads as a perfect score. A firewall returned 403 to headless Chrome, the auditor scanned the challenge page — almost no content, so almost no violations — and printed "Clean". It now refuses 403s, rate limits and interstitials as errors, and the exit code counts them. A page that was never audited is not a page that passed.
●An animated page gives a different answer every run. The same page returned 49, then 47, then 29 contrast failures on three consecutive scans, because the check was measuring text mid-fade. With a settle step it returned 4, 4 and 4 — and 4 was the truth. I don't quote an unsettled number to anyone.
●The clever detector missed the known-positive. ghostbox's first idea — count the pixels a window never painted — failed, because Windows fills an unpainted window with opaque black when you ask it to render itself. It looked painted. That false negative is why the tool now also compares the window against what is really on screen.
●A tidy table that was about 1000× wrong. Windows reports process memory as text like "16,728 K"; splitting on the comma keeps the last three digits and produces a plausible, confident, wrong report. In the same tool, grouping by program name made one virtual machine look like three.
●Redaction boxes sized to the text leave the tops of the digits. Two redactions failed silently on first use — one landed a line high, one clipped the ascenders — and both looked complete at a glance while the numbers were still reconstructable. Hence padding by default, and a rule: verify by cropping each redacted region and reading it back at native resolution, because scaling is what hides the surviving pixels.
●A tool that won't tell you what it can't do is worse than no tool. scrub cannot remove statistical watermarks that live in word choice rather than in characters, so it says that plainly instead of handing back a file marked clean.
Proof
29 / 29
pages verified clean in production, not on local files
0
serious accessibility violations removed — proven by a reverted-fix control
49 → 4
same page, unsettled vs settled — only one of those is reportable
87×
ghostbox scan after profiling: 21.8s → 0.22s
100%
detection on ghostbox's own generated ghost; clean desktop 0–6%
0
finished tools, each with its own oracle
Rust (five of six)Node · puppeteer · axe-corewindows-sysclippy -D warningsno panics, no unwrapslocal-only, no telemetry
ℹ
Honest scope on the accessibility auditor: automated checking catches roughly 30–50% of WCAG — contrast, labels, landmarks, roles. It cannot judge whether alt text is meaningful or whether the focus order makes sense, so a full defensible pass still wants a human with a screen reader. The tool is the fast, repeatable half, and it says which half it is.
🔒
These run locally and nothing leaves the machine without an explicit click. Source is private — this is a capability showcase, not a distribution.

Need a tool
that tells the truth?

The tools are small. The standard isn't: each one ships with the check that would catch it lying, because a green result you can't falsify is worth nothing. Tell me what you need measured.

Start a project →