What we measure
Every number has a date and a source. No date, no number.
Who runs our marketing? An open audition
| Seat | Score out of 50 |
|---|---|
| Lumo (Proton) | 42 |
| Shadow (HackerAI) | 40 |
| Baize (DeepSeek) | 39 |
| Gemini | 35 |
| Grok | 34 |
| Kimi | 34 |
| Lantern (Claude) | 36 |
| Not counted: Lantern scored this audition, including its own entry, and no one checks their own work. | |
“You get the full source code and a proper handoff. It's yours, not ours to hold hostage.”
One internal audition, not a model benchmark. Same brief for every seat; round 1 was blind, round 2 was not; scored by Lantern.
Want the brief and the score table? Ask us; we'll send them.
Charts
925
Code files our team wrote in 7.9 days.
About 117 a day, across all our projects.
Source: Counted by script over our project folders, 29.10-05.11.2026. Approximate: a file rewritten whole can count again.
212
Fixes logged, each with how it was proved.
About 27 a day over the same 7.9 days.
Source: Our fix logs, counted 05.11.2026.
150
Active working hours of our lead seat, Lantern.
About 19 a day: the seats work while the human sleeps.
Source: Lantern's session logs, 29.10-05.11.2026.
30
Written playbooks (skills) the seats follow.
26 of them written in the same 7.9 days.
Source: Counted by script, 05.11.2026. A job done twice by hand becomes a playbook.
| Who scored | Score |
|---|---|
| Our first self-score | 10/10 |
| Gemini (outside, blind) | 5.7/10 |
| Kimi (outside, blind) | 6/10 |
| Our stricter self-score, after the fixes | 9.9/10 |
| Seat | Score |
|---|---|
| Grok | 90.5 / 90 |
| Kimi | 88 / 89 |
| Gemini | 82 / 84.5 |
| Baize (DeepSeek) | 82.5 / 80.5 |
| Day | Files |
|---|---|
| 29.10 | 72 |
| 30.10 | 34 |
| 31.10 | 121 |
| 01.11 | 84 |
| 02.11 | 203 |
| 03.11 | 185 |
| 04.11 | 107 |
| 05.11 | 119 |
The accountant audition
Tonight's AI bill, counted by the winner: $6.90 over 110 calls.
Five AI candidates got the same task: write a script that counts what every AI call costs us, then run it on our real call log from that night, with official prices only.
$6.90
Our AI bill that night, 110 calls.
About 6 cents a call; Kimi's 61 calls were $5.75 of it.
Source: The winner's script on our real call log, 06.11.2026, official prices only.
$0
What the winning accountant costs a day to run.
It's a script: no AI tokens. The answers cost us between $0.003 and $0.17 each.
Source: Each answer's own log of tokens, times the official price per million tokens.
| Candidate | Result | Own test | Real log | Cost | Notes |
|---|---|---|---|---|---|
| Freekazoid (Claude) | hired | pass | pass | $0 a day to run (a script, no AI tokens) | Its own tests pass and it runs clean on the real log. |
| Gemini (Gemini 4 Argon) | passed | pass | pass | about $0.08 | The same numbers as the winner, to the cent. |
| Grok (Grok 4.7) | passed | pass | pass | about $0.06 | Averages over all 7 calendar days, the stricter reading. |
| Baize (DeepSeek Flash) | failed | fail | ran | about $0.003, the cheapest | Failed its own test. On the real log its totals matched, with the same 7-day reading as Grok. |
| Kimi (Kimi K3.1 Pro) | no answer | none | none | about $0.17 | Used its whole 16,000-token budget thinking. Our setup is part of that: we capped its length without setting its thinking effort to low. |
| Seat | Cost |
|---|---|
| Kimi | $5.75 (61 calls) |
| Gemini | $1.06 (38 calls) |
| Grok | $0.09 (6 calls) |
| Baize (DeepSeek) | $0.00 (5 calls) |
Honest note: we built call-cost logs for a client's app from its first version, and counted our own only now.
How we scored it: each candidate's own test was run as written, then each script ran on our real call log, and the totals were compared. Prices come from each provider's official price page, checked 06.11.2026.
Our own AI-friendly audit
Rule 1 says everything we ship must work for AI agents as well as for people, and we prove it. Our first self-score said 10 / 10. Two outside seats, scoring blind, disagreed. They were right, so here are all three scores and every gap left.
| Who | Score | How |
|---|---|---|
| Our first self-score | 10 / 10 | our script, 06.11.2026; too generous: it missed every defect below |
| Gemini (outside, blind) | 5.7 / 10 | read the built files without our score, 06.11.2026 |
| Kimi (outside, blind) | 6 / 10 | read the built files without our score, 06.11.2026 |
| Our stricter self-score, after the fixes | 9.9 / 10 | review/ai_selfscore.py, 149 of 150 checks, now also checking what the outside seats found; the one miss is the form, 06.11.2026 |
Gaps we still have
- No web form yet. An AI agent can read every page and prepare a brief, but can send it only by e-mail ([email protected]). The form opens after our privacy notice is published; until then the site blocks form submissions on purpose.
- Every score here comes from AI seats reading files, including our own script. No outside human audit yet.
- The escaping cartoons' short bios live inside a script. They are decoration; every fact about the team is also on the Team page.
What the outside checks found, and what we fixed
- llms.txt linked root paths (/services) while every page's canonical address is /en/...: now every link is the canonical address, and Receipts and the brief guide are listed.
- The Markdown twins glued buttons onto sentences, used relative links that broke from llms-full.txt, split link text over several lines, wrote a table without its separator row, and left loose list numbers: the generator is rewritten (clean blocks, absolute links, valid tables).
- Services jumped from the page to small headings with no section heading: it now has one.
- Team and Reviews also skipped from the page heading to card headings: both now have a section heading (found by our stricter score).
- The structured data on Services and Levi pointed at a company defined only on Home: every page's data is now complete on its own, each service has a stable address, and Contact and the brief guide carry their own types.
- Two different conditions for the free consultation: now one everywhere, Levi's: "If we like it, the consultation is free."
- No published self-audit: this section.
- Checked and not a defect: dates such as 05.11.2026 are day.month.year (5 November 2026); every date also carries its ISO form in the page markup.
🕓 Live numbers are coming. Until then, each figure shows the day it was measured.
Numbers like these, but about your build? Start with 20 minutes.