The weakest figures, on the same page as the argument

Zero revenue. Zero customers. Zero completed transactions across both businesses. One of them has live payment links, a complete delivery chain on disk, and a gate that refuses to let anything take money without a producer, a renderer and a checker present — and nobody has ever used it.

There is no metric key for revenue or customers, and that is deliberate rather than evasive: zero is countable, so it is stated in prose, but no counter exists because there is no source system to count from. The moment there is one, it becomes a figure like every other figure here.

What this means for reading the rest of the site: everything here demonstrates engineering and process. Nothing here demonstrates that anyone wants to buy the result. Those are different claims and only the first one is supported.

The first business went through all 7 stations, producing 6527 lines of research, product and technical documents, and 58 evidence records each verbatim-matched against a hash-pinned source.

The second went through 0. It was built directly. That is measured by checking whether seven filenames exist in its repository — they do not — rather than by recollection.

The uncomfortable implication: the second business is finished, tested and gated without any of that documentation existing. So either the document pipeline is not load-bearing, or the second build is worse in ways nothing has surfaced yet. I do not currently know which, and a portfolio page describing a “seven-station process” without stating that would be overselling a method on a sample size of one.

Mutation testing breaks the code on purpose and asks which breakages the gates notice. The first business scores 7%: 243 mutants applied, 17 killed, 226 survivors.

The second scores 29.7%. 1442 mutants applied, 428 killed, and 1014 survived — that many deliberately broken versions of its code got past the 7 gates a mutant is actually shown, out of the 20 that business runs.

That sentence used to read “passed every gate it has”, and it was false. A mutant faces a fast subset — the checkers that run in seconds — never the visual baselines, Lighthouse, the contrast rasteriser or the payment reconciliation. The mutation tool prints that limit on every run, in those words: “survived a subset and survived everything are different claims, and conflating them is the defect this file exists to find.” The disclaimer lives in a terminal nobody keeps, the report on disk did not record which gates ran, and this page made exactly the conflation the tool warns about — overstating the apparatus, on the page whose entire job is not to.

Found by an audit of the process on 20 August 2026. The size of the oracle is now a counted figure like every other number here, so the claim cannot outrun it again.

A surviving mutant is still not a hypothetical. It is broken code that everything it was shown accepted, which makes those survivors a map of where that business is unprotected. The number is published rather than omitted because the alternative is a site that shows one mutation score and not the other.

The defect log breaks down as 12 found by reading carefully, 4 by an adversarial audit, 1 by a gate, and 1 by a mutation run.

Every gate on the line was written after something got past everything that existed at the time. That is the honest reading of the whole apparatus: it is a record of failures already made, not evidence of foresight. It says the next novel defect will probably also be found by a person, hours late.

While building this site, three separate sources gave three different answers about the same two figures.

• Gate count, business #1. Its own generated METRICS.md said 14. Its runner declares 20. The generated file was stale — regenerated from a repository that had since grown two more gates.
• Test count, business #2. The brief commissioning this site said 356. A static count of it() and test() declarations across its spec files gives 235. Its own METRICS.md reports a third number again, because Playwright multiplies declarations by project and parameter.

No figure on this site is transcribed from any of those sources. Each is re-derived here by opening files in that checkout, and every one prints the file and pattern it came from. The test figure is labelled “test declarations” rather than “tests” precisely because the larger number cannot be re-derived from the repository — and publishing a number you cannot reproduce is the exact defect this line exists to prevent.

Until 20 August 2026 this site had none either, and its siteUrl was null rather than a plausible-looking URL, so nothing in its sitemap or metadata pointed at an origin that did not exist. That was the right call and it still cost something: with the value null, metadataBase fell back to localhost, so the built pages signed every preview image with a hostname nobody serves and any link to this site would have rendered with no preview card at all. Nothing failed. The build succeeded, the pages looked correct, and the only symptom was an absence somewhere else — which is the shape of most of the entries on this page.

The second business is the same: finished on disk, no deployment, no Stripe product, no payment link, and a service registry that reports status “building” so that every gate agrees with every page. Only 1 of 2 businesses is actually live.

Both businesses run a gate that asks the only question a founder actually has: can a stranger find it, understand it, buy it, and receive it? It runs last, and it is the one that decides whether the work is finished rather than merely correct.

This repository does not have it. Two of those four verbs have no referent here — there is nothing to buy and nothing to receive — so the gate was removed rather than narrowed. Narrowing a definition of done until it matches whatever currently passes is not a definition of done; it is a green check standing in for one.

The consequence is worth being plain about: of the 17 gates defending this site, every one measures whether something is correct, and none measures whether it is done. That judgement is currently mine, unaided, which is precisely the arrangement that produced most of the defect log.

Four other gates that Wryko Records runs are absent here too, each named with its reason and its replacement at the top of scripts/verify-all.mjs — because dropping a check quietly is the same class of act as a check that passes when it cannot run.

The honest boundary of the evidence on this site:

• No team. Everything here was built solo. Nothing demonstrates code review, mentoring, handover, or working inside someone else's conventions.
• No scale. No production traffic, no load, no database under contention, no incident at 3am. The gates measure correctness, never behaviour under pressure.
• No maintenance history. These repositories are young. Nothing here shows what the code looks like after a year of other people changing it, which is the condition most professional software actually lives in.
• No external validation. No user research with strangers, no usability testing, no third-party security review, no accessibility audit by anyone but the tooling described here.