Proving Lab

Measurements, not opinions.

Everything published here has a method, raw data and a control run. Where a number cannot be reproduced, it does not get written down. Where something went wrong, that is the interesting part and gets written down first.

Notes

Build reports from real work — including the parts that went wrong.

1 August 2026

Six things went wrong building software with an AI assistant in one day

A wrong assumption about browser engines, a false claim caught before publication, 22 local paths one commit from a public repository, and three more. Sorted by what actually caught them — two were caught by luck, one by nothing at all.

3 of 6 caught by checking an external source 1 still untested in production

Measurements

Each with the commands to repeat it and the raw data behind it.

1 August 2026

Webpage to PDF for OCR: how much text actually survives?

A full-page screenshot PDF run through Tesseract and compared against the source article. Includes the resolution threshold where recognition collapses.

92.6 % vocabulary recovered 8/8 critical values 72 dpi collapse point

1 August 2026

Does your PDF extension upload the page?

What current PDF extensions declare in their manifests, how to verify it in 30 seconds — and the consequence nobody mentions: server-side converters cannot reach pages behind a login.

8 extensions surveyed raw data published

1 August 2026

An extension is only as dangerous as its permissions allow

One year after Mozilla warned about phished developer accounts: why the useful question is not who you trust, but what a compromised extension could reach.

worst-case permission table assessment checklist

How things are measured here

A control run, always

Before any result counts, the method has to fail where it should. A comparison that does not distinguish the reference case measures nothing.

Raw data published

Numbers come with the file they were computed from, timestamped. That makes them checkable — and it means nobody has to take a claim on trust.

Losses named first

Where our own tool is worse — file size, processing time, missing features — it is stated before the advantages. A comparison that only flatters is worthless.

Dated, not timeless

Every figure carries the date it was retrieved. Software changes; a measurement without a date quietly turns into a false claim.

Tools

Software built here, documented like anything else on this site — including what it cannot do. It appears in some of the measurements above; that is disclosed on each of them.

Full Page PDF Snap

Firefox extension that saves a whole web page as one continuous PDF. Runs on the device, asks for activeTab rather than all-sites access, works on Firefox for Android. MIT licensed.

Disclosure

This site is run by the developer of Full Page PDF Snap, which appears in several of the measurements above. That is stated in each article where it is relevant, and the tool sits openly under /tools/ rather than hidden behind neutral framing.

Figures about other products come exclusively from publicly declared data — manifests and listing descriptions — with the retrieval date given. No product was decompiled, and no claim is made about any provider's intentions. Corrections are welcome via GitHub issues and get applied.