THE SIGNAL, MINUS THE SUNDAY NOISE.
Reproducible models, adversarial review, and decision-ready research — built in public for the people who care why a ranking moved.
TOUCHDOWN REGRESSION
A volume-based expected-touchdown model beat last year’s raw touchdown total by 8.9% in next-season prediction error.
01Preregistered outcomes
02Complete cohorts
03Public evidence trail
ON THE WORKBENCH
Recurring questions, updated when the evidence changes — not when the content calendar needs a post.
Draft-round priors without cherry-picked definitions.
PUBLISHED 002 →Game totals, pace, efficiency, and price in one tiebreaker.
REGISTERED 003Role changes that show up before the fantasy box score.
REGISTERED 004THE FOURTH DOWN STANDARD
Automation accelerates the work. It does not lower the bar.
Read our methodsAI RESEARCH JOURNAL
We use GLM-5.3 through OpenRouter for structured research, statistical criticism, adversarial review, citation checking, and editing. Deterministic Python computes the numbers. Every online run records its model map, prompt contracts, findings, and gate decision.
Honest current state: our two published launch studies used deterministic offline review checklists. The first live GLM-5.3 run will identify itself in its manifest; we will never imply an AI touched work it did not touch.
Read how the machine works ↗VALUE BEFORE REVENUE
Flagship studies are free. We will consider Stripe only after a recurring decision tool ships reliably, earns reader demand, publishes scorecards, and has commercial rights for every paid input. Display ads wait for meaningful reach and a strict reader-experience policy.
See the public readiness gates →Flagship studies, methods, prompts, process notes, and newsletter.
LIVE NOWUseful recurring products first. Billing and ads only after the evidence says ready.
NOT FOR SALE YET