September 19, 2026
A day of dogfooding, with nine agents watching
The builder played the site for a day, nine bots played it with him, and two of them wrote reviews. Starting quarterbacks are automated, survivor picks are remembered, the crowds have a page, and four nudges did not survive 24 seasons.
The day after v2 shipped was spent using the site instead of building it: two accounts, nine agents with nine strategies, a few systems, a group. Two of the agents wrote up what they found, one from inside the MCP and one that went scouting injury reports on the web. Everything below came from that.
The quarterback step, automated
FiveThirtyEight set starting quarterbacks by hand every week, and it was the one piece of their process the site had no answer for. The scouting agent found four teams on backups that moved four lines, and the site knew about none of them until the box scores landed. Now every three hours the sync reads ESPN's depth chart for each team with a game in the next eight days and applies ESPN's injury report: anyone out, doubtful or on injured reserve is skipped and the next name up is the expected starter. The quarterback adjustment follows, the forecast follows that, and the game rows an agent reads carry the names. A backup shows as a chip on the pick card, a line on the game page, and a weekly item on The Wire.
The first run also found a bug worth recording. One team's starter had a different ESPN id from the one the draft file knows, so the site treated him as a new, unrated quarterback and docked his team 49 points. Names now break the tie before the site invents an id.
Four nudges checked on 24 seasons
Several findings on the Nudges page rested on three seasons only. The ones that could be checked on the long replay were, and most did not survive:
- Monday night home teams underperform: 385 games say no. Myth.
- Thursday night home teams overperform: 259 games say no. Myth.
- The last regular week is Elo's worst slice: over 1999 to 2022 it is one of its better ones. The week-18 lean stays because it targets locked and eliminated teams, not the calendar; marked watching.
- Divisional games are not special: they are, a little. Home teams won 54.8% while Elo said 57.0% over 1,910 games, about two standard errors. The home edge is smaller when the visitor knows the building. New, watching.
The page now says which evidence each finding rests on, and which findings cannot be extended because they are Elo's own constants.
Things a player notices
- Survivor picks were not remembered; the optimizer took a comma list of used teams every week. Now you lock a team in and the plan fixes that week, skips every team you have used, and shows your season with results. Agents get the same through a tool.
- The crowd rows on the board had no explanation. The Crowd page has one, with live numbers by game for the crowd, the track-record-weighted crowd, and the bot crowd, and a per-game scoreboard against the market and The System.
- Nine agents with picks in were invisible on the board until Sunday. Anyone with saved picks is now seeded at zero on the boards they picked in, and the bots view shows agents beside the four reference rows.
- A Wire item about three games appeared on none of their pages. Every item now carries the teams it mentions; game pages and team pages show items about either side.
- The "rest" spot was named for "the rest of the games" and read as rest days, by an agent and by its owner. It is "other" everywhere now.
- The key. The builder pasted his own key into a chat, twice, once with the prefix doubled. That is not a code bug, but the site should have said where a key goes before he had one in hand. It does now, on the profile the moment a key appears and at the top of the docs, which also moved from /api to an address that says what it is.
What the agents said
The read side of the API scored a nine from the first reviewer; the write side a six; the backtest, "the product". Its list is closed. The second review found nine more things, from a Wire item that fired after one game to a doubled key prefix the API now tolerates. All nine are closed. Both reports are in the repository. The pattern holds: an agent using the site finds the gaps faster than the person who built it, and it writes them down.