Track Record: Movies
A series of small tests of whether the past predicts what comes next. Volume 1 asks whether a film's director, cast or studio predicts how critics will review it, using 17,126 films and public data from Wikidata.
Drafted by a bot from my own words. I’ve edited and approved everything that is on this site.
- Problem
- Does the past predict what comes next? Specifically: do the people behind a film (cast, director, studio) predict how critics will review it?
- Approach
- Use public Wikidata, build every prediction forward-only, compare with a plain baseline, check the data against an outside source, and publish every result, including the ones that showed nothing.
- Solution
- A live site with four games, a methods page with ten further analyses, and a generated build record, plus a reusable skill and this portfolio so the next question takes less work.
- Result
- The director's record predicts a film's score far better than the cast's (0.56 against 0.06), yet the best guess is still about 16 points off. Several popular hunches (stars, runtime, experience) showed little. The site is live; there is no usage data yet.
What this shows
- Speed: a starting app and spec became a live, multi-page analysis site in one evening.
- Care: results are tested forward-only, spot-checked against an outside source, and published with their limits.
- Range: the method is packaged as a reusable skill, so the next problem starts from a tested process.
Speed
From a starting app and spec to a live, tested, multi-page analysis site in one evening. The build log shows 17 commits between Oct 13, 2026, 9:58 PM and Oct 14, 2026, 12:40 AM Central (2 h 42 min from the first commit to the latest).
- 13 pages and routes, including a methods page and a how-it-was-built page
- 4 games: Higher or Lower, Today's film, a practice quiz and an actor duel
- 17,126 films analyzed, with the director, cast and 10 further analyses
- 14 scripts that fetch data, run the tests and generate the numbers on the pages
- A spot-check of 59 films against an outside source, and a generated build record
The first commit already held a starting app and spec from an earlier session, so the timeline counts the work after it. Commit time is not time worked: data downloads ran in the background, and I reviewed and steered between commits. Counts verified against the repo on 2026-10-14.
Decided by me
- Moved from a critic leaderboard to actors, then to directors, when the data pointed there.
- Kept real-money betting out and made the prediction games free to play.
- Launched free with no paywall, because the free hosting plan is non-commercial.
- Named it Track Record and made it a series, with movies as volume one.
- Asked for plain wording in the games and a notebook look for the site.
- Asked for a generated record of what the AI built and when.
Done by the AI
- Wrote the data pulls, the analyses, the site and the games.
- Ran the tests and reported the results, including the ones that showed nothing.
- Drafted the pages and copy for review.
- Flagged limits and risks: licensing, hosting terms, and where claims went further than the data.
Decisions that shaped the build
| Insisted on a clean, openly licensed data source. | No legal risk, and every result can be reproduced. |
| Pushed past a data blocker to find a legal route. | Led to Wikidata: no key, public domain, and 17,126 films. |
| Checked the hosting terms before launch. | Spotted that the free plan is non-commercial, which shaped a free, compliant launch. |
| Made ownership and security a requirement. | Private repos, accounts in my name, and a domain I control. |
| Pushed for friendly, plain wording. | The games use an approachable 'cast-only guess' instead of technical jargon. |
| Set a distinct design direction. | A line-printer look that stands apart from template sites. |
| Cut feature creep and refocused on what's fun. | Shipped one-tap Higher or Lower. |
| Asked the sharper question about production quality. | Led to the director analysis, the strongest finding in the project. |
| Tested the site like a visitor. | Added navigation on every page. |
| Asked for a verifiable record of the work. | A build log generated from git, shown on the site. |
| Turned the method into a reusable skill. | New ideas can run through the same tested process. |
| Spotted a game that gave away its own answer. | Moved the duel to its own page. |
| Chose openness about AI's role. | Every page states who wrote what. |
How it was checked
- Every prediction uses only films released before the one being predicted.
- The combined model was trained on films through 2017 and tested on later films.
- The assistant spot-checked 59 random films against their Wikipedia articles: 45 of 46 Rotten Tomatoes scores and 15 of 15 Metacritic scores matched within five points. The pass bar (80%) was set before running it.
- When the cast test and the director test used different films, the director test was re-run on the cast test's films (0.37 against 0.06).
Issues caught before publishing
- A first pass suggested January was the best release month. A data check showed year-only dates were stored as January 1, so those films were removed before anything was published.
- An early box-office result came out negative. Checking it showed small acclaimed films skewing it, so the site now reports wide releases separately and flags the direction as unsettled.
- A chart listed the same two genres twice. A minimum-count rule was fixed before publishing.
Build record
From the project's git history: 17 commits, 17 marked as AI-assisted, between Oct 13, 2026, 9:58 PM and Oct 14, 2026, 12:40 AM Central (2 h 42 min). Commit time is not time worked: background jobs ran between commits and some projects began with earlier work. Refreshed Oct 14, 2026, 12:40 AM.
| Oct 13, 2026, 9:58 PM | Critic accuracy leaderboard: zero-config Next.js app and spec | AI-assisted |
| Oct 13, 2026, 10:52 PM | Track Record: actor tiers, guess-the-score quiz, daily film | AI-assisted |
| Oct 13, 2026, 10:54 PM | Add /method page with backtest scatter plot | AI-assisted |
| Oct 13, 2026, 10:57 PM | Share cards: OG image, share landing page, emoji result line | AI-assisted |
| Oct 13, 2026, 11:00 PM | Add Higher or Lower game | AI-assisted |
| Oct 13, 2026, 11:12 PM | Clarify signup purpose | AI-assisted |
| Oct 13, 2026, 11:13 PM | Support hosted newsletter subscribe link | AI-assisted |
| Oct 13, 2026, 11:16 PM | Notebook-style home, actor tiers page, notes stub | AI-assisted |
| Oct 13, 2026, 11:21 PM | Add site header navigation | AI-assisted |
| Oct 13, 2026, 11:26 PM | Method page: director finding, genre check, cast comparison | AI-assisted |
| Oct 13, 2026, 11:40 PM | Track Record series: Movies hub, format page, how-built, nine analyses, logo | AI-assisted |
| Oct 13, 2026, 11:44 PM | Add spot-check script | AI-assisted |
| Oct 13, 2026, 11:52 PM | Add build log: generated record of AI-assisted work from git history | AI-assisted |
| Oct 13, 2026, 11:52 PM | Refresh build log | AI-assisted |
| Oct 14, 2026, 12:29 AM | Move the actor duel to its own page; the actors page is now just the table | AI-assisted |
| Oct 14, 2026, 12:35 AM | Method page: What we learned and why it matters, generated from the results | AI-assisted |
| Oct 14, 2026, 12:40 AM | Refresh build log | AI-assisted |
Limits
The findings are correlations. Only films with a score and a director on Wikidata are included, and directors need three earlier films, which favors established ones. The prediction model is a transparent baseline, around 15 points off for established directors, not a forecast.