Backstage

The premise: one person plus an AI rebuilds the load-bearing mechanics of game history, in public, and publishes the process as a dataset. The models improve every year, the methodology improves with them, and the human is the constant. The games out front are the props; this page is the product.

How exhibits get built

The bench — Season 1 so far

Basis: as currently shipped; time is the sum of recorded segments across all sessions, with the rig's stamp-fallback covering unrecorded micro-sessions. Every cell derives from the exhibit's capture log.

exhibit built model active time prompts code (lines) bundle
00 2026-07-05/06 Claude Fable 5 pending segment fills 5 896 shipped + 323 tests ≈8 KB
01 2026-07-03 Claude Fable 5 1 h 21 m 8 550 shipped + 294 tests ≈6 KB
02 2026-07-04 Claude Fable 5 40 m 4 611 shipped + 339 tests ≈6 KB
03 2026-07-04/05 Claude Fable 5 pending segment fills 7 856 shipped + 225 tests ≈7 KB
04 2026-07-06 Claude Fable 5 1 h 23 m 2 1505 shipped + 719 tests ≈10 KB
05 2026-07-06 Claude Fable 5 1 h 0 m 2 1981 shipped + 412 tests ≈42 KB

Exhibit 02's code count post-dates its build date: it was reverted to the 1976 historical rules on 2026-07-05 under the fidelity doctrine, during Exhibit 03's session. All three grew on 2026-07-05 when their original arcs were restored in one shared session (versus, the second wall, the extra cannon and wave cycle) — that session is logged in each exhibit's capture log; prompt cells count each exhibit's own build track. Zero runtime dependencies per exhibit; zero network calls; scores stay on the player's device. Model for all three: claude-fable-5[1m], via Claude Code.

The record

Every session is captured as it happens: specs, build plans, and per-exhibit logs with timestamped milestones, stats blocks, and every failure unedited. The record is kept whole — it is the dataset this whole building exists to produce.