Proof · verification indexreproducible

Don’t take my word — check it

Behind every number in the case is code, a git command or an artifact. This is a verification map for the sceptic: run the command and see for yourself.

7
repositories
145
commits
10
days in git
200+
AI agents

доказаноevery number is checked by a read-only commandto the table →

Checkable anchors

What’s inside the system — in numbers

The internal metrics are pulled from the real repository — not estimates, but facts from the code and the artifacts. The make-up first; below, in the table, is how to recompute each number yourself.

  • Flagship PWA — 42,510 lines of TS/TSX: 111 components across 13 domains, 21 REST routes, 25 database tables in one schema.
  • Quality — 66 test cases across 26 files; CI runs types → style → tests → build, stopping on the first error.
  • Book — 261 pages, 43 chapters, PDF + EPUB 3 (epubcheck PASS, 0 messages), 3 editions.
  • Performance — a static prerender of every route: interactivity is added by 6 pinpoint client inserts on top of the static page, not a full re-render of the whole page.
  • Pace — 145 commits in 10 days across 7 repositories: 1 root + 6 modules.
A year in git

Behind the case’s 10 days — a year of live work

The flagship was built in 10 days — but it wasn’t a one-off sprint. Over the past year on the public GitHub profile — 3 512 contributions, activity on 281 of 370 days, and a streak of 130 days in a row. Each green square is a real trace you can see on the profile.

доказаноa snapshot of the public profile github.com/temadev — open it and checkopen the profile →

Claim → how to check

Every row is a read-only command or an artifact

The commands only read, they change nothing. Run them from the repository root. No internal paths, secrets or network addresses — only what anyone can reproduce.

ClaimHow to check
145 commitsgit rev-list --count HEAD per repository
10 days in gitgit log --format=%ad --date=short — first and last commit
7 repositories (1 + 6)cat .gitmodules — root + 6 submodules
42,510 lines of TS/TSXfind src -name '*.ts*' -exec cat {} + | wc -l
111 componentsfind src -name '*.tsx' | wc -l
21 REST routesfind src -name route.ts | wc -l
25 database tablesgrep -c 'sqliteTable\|pgTable' schema.ts
66 test casesgrep -rho 'it(' --include='*.test.*' | wc -l
261 book pagespdfinfo book.pdf | grep Pages
EPUB is validepubcheck book.epub → PASS, 0 messages
Static prerender of routesnext build → each route marked ○ (Static)
200+ agents / wf-idgrep -rhoE 'wf_[a-z0-9-]+' qa/ | sort -u

The commands are generic. The full index, with the expected numbers per repository, I walk through on a call — together with read-only access.

AI orchestration

I design AI pipelines — not just use AI

The strongest proof isn’t the volume of code, but how it was made. Each deepening stage passes an adversarial barrier: draft → three parallel veto lenses → revision by the verdicts → a block verdict of “to production” or “back for rework”. 200+ agents across 9 Workflow cycles; each cycle leaves a real wf_* run id.

Lens G1

Accuracy

Facts, sources, internal consistency. Whatever fails the check goes back for revision.

Lens G2

Safety

Block-specific risks: ethics, wording, the absence of harmful claims.

Lens G3

Voice and structure

One brand voice as a system prompt, the integrity of the structure and the transitions.

4th layer

“Break it”

An independent audit by a fresh agent with a mandate to find what the three lenses missed.

Each cycle’s output is machine-readable findings (findings.json) and a block verdict of “to production” / “back for rework”. Book audit: 62 agents → 77 findings.
This is an operating system, not a one-off prompt. One stream of workers checks the work of another by the veto lenses; I make the final call on the verdict.
Honesty boundary

What’s proven, what’s an estimate — and what I deliberately don’t claim

Three kinds of numbers, with different statuses: proven metrics from the repository, external market reference points (a calculation, not a fact), and business outcomes I deliberately don’t claim yet. Each group’s colour comes from the trust code above.

доказано

Internal metrics

  • Lines of code, components, routes, tables, tests — pulled from the repository.
  • Commits and time in git — by a command across each of the 7 repositories.
  • Book pages, EPUB validation, the number of editions — from the assembled artifacts.
  • The agent count and wf-id across 9 Workflow cycles — from the audit reports.
оценка

External reference points

  • Competitors’ market statistics (renewals, churn) — the sources weren’t kept.
  • Penalty amounts and external price ranges — a reference point, not a fact.
не заявляю

Business outcomes

  • Revenue, customers, sales — the first cohort has only just started; I don’t pass it off as fact — I claim only what’s provable.
  • Reviews and before/after cases — I publish only with consent and only when I can confirm it.
Data protection under 152-FZ (Russia’s personal-data law) — consent to collection, an immutable log, two-way erasure (app ↔ bot) — really works in the code and is checked against the schema. Production is shipped; I don’t pass revenue and customers off as fact — I claim only what’s provable. This is a matter of honesty, not a gap in the skill.

Next

The numbers add up — see how they came to be

The evidence base holds together. Three doors — each answers a different question.

Check it in person

On a call I’ll open read-only repository access and walk through the index — with the expected number behind each claim.