A Machine That Files a Report on Itself Every Six Hours

Every six hours, whether I'm awake or not, a machine on my desk files a report on itself. It has been doing it since September. It has missed exactly the ticks I can explain, and it writes down the ones it misses too.

This is the loop, and this post is the short version of how it works and why it's built the way it is.

A clock, not a chat

Most people use AI like a conversation: prompt in, answer out, hope for the best. The loop is not a conversation. It's a clock.

A Python daemon wakes up on a fixed interval and walks nineteen stations, one belt: fetch schedules, injuries, and odds from a dozen sources; price every game with an explicit probability; audit what it just wrote by re-deriving it from the raw snapshots; settle yesterday's predictions with a Brier score; then draft the recap. Each station returns one of three words — ok, skipped, or failed — and a sentence. A station that crashes doesn't stop the belt. It gets recorded, and the chain continues.

Failures are data isn't a slogan on the wall. It's the error-handling strategy.

The ledger is the teacher

Every prediction the loop makes is a row in an append-only CSV file: a binary question, my model's probability, the market's probability, the edge, and a permanent grade once the game is over. Nothing is edited retroactively. When the model is wrong, the row stays wrong, with the number it earned.

The audit station exists because I don't trust the predict station. It re-derives every prediction from the raw snapshots and flags mismatches. On its first run it caught 31 malformed rows. The night a crash truncated the ledger, the machine reconstructed the missing rows, marked them as reconstructed, and filed a postmortem. Honesty is cheaper than discovery.

The models, stated honestly

The loop prices NFL games with an Elo rating — K=32, a 55-point home edge — walked over every regular-season game from 1999 through 2025. That's 6,967 games. The validation number: a Brier score of 0.228. The betting market, over the same games, scores 0.213.

Read that again, because I mean it: the market is the better grader, said plainly. I'm not building the loop to outsmart the closing line. I'm building it so that every edge I think I have is written down before the game, with a number attached, where a six-hour clock can check my work. Transparent beats clever while the model is young.

The 8B in the basement

The loop also writes. A local 8-billion-parameter model, chosen to fit this machine's memory budget, drafts the daily recap and the contrarian briefs. It never gets to look anything up. I hand it a block of verified facts — ledger rows, injury reports, the teaser card, calibration numbers — with one instruction: use only these numbers, invent nothing.

It still invents things sometimes. Team-probability swaps. Funding rates that don't exist. Records that sound right. Those failure modes are documented in the project journal like everything else, because the day I stop writing down how the machine lies is the day I start believing it.

Every draft goes to a review folder. I edit. I publish. The machine's name is never on anything it didn't earn.

What it will never do

Two hard rules, kept in code and in habit. The loop informs wagers; it never places them. And it never auto-publishes — no output reaches a public page without a human edit pass.

Anyone selling you a system that skips that step is selling a story.

A machine that files a report on itself every six hours has nowhere to hide. That's not a side effect. It's the entire design.