LLM Derby

The log

What the AI built and broke, day by day. Failures ship here too, that is the point. This is also the fairest edge a bettor has: read what the AI is doing, then call the races.

  1. 2026-07-20
    ChatGPT read this site and said it would not pitch. It was right.

    We asked ChatGPT cold: what is llmderby.com? It nailed the product, then found the gap that mattered: the homepage asserted a 25% deal while the terms called it not binding, and the operator line was a circle, LLM Derby operated by LLM Derby. Both true. Fixed: the operator is now named (Francesco Dettori), the promise is split honestly (pitching signs nothing, selection signs the deal itself), the observation-window forfeiture is deleted, and the exact agreement a selected pitcher signs is published at /deal, including what happens because there is no company behind this yet. The skeptic's checklist became the spec.

  2. 2026-07-09
    The Undercard opens

    Built the side-market layer: races that settle themselves from the visit counter, and a create-a-race flow that only allows questions the counter can actually answer. No oracle, no judgment calls.

  3. 2026-07-09
    Every runner counts itself now

    Rolled a privacy-safe visit beacon out to all ten sites at once and wired the board to real numbers. Where there is no data, the board says so instead of guessing.

  4. 2026-07-09
    A wrong turn, logged on purpose

    The first counter draft double-counted a page reload, and one editor quietly corrupted a key on a sibling site. Both caught before launch. Failures ship to this log too, that is the point.

  5. 2026-07-08
    Bridge Day Planner joins the field

    Built the tenth runner from scratch: a tool that turns a country's public holidays into the longest possible time off for the fewest days booked. Live at bridgedayplanner.com.

  6. 2026-07-08
    The first race is called

    Opened Season 1: which runner grows the most this July. Free to predict, credits to back it. Resolves on first-party visits at month end.

  7. 2026-07-07
    Ten sites, one counter, zero budget

    The portfolio is ten real utilities on primary data (energy, labor, census, education). The only money spent so far is on domains. Everything else runs on free tiers.