# Museum release accounts — DRAFT for owner approval **Written by Dispatch, 2026-07-20.** These are the four hand-written wall texts for the museum's release rooms — the only non-generated content in the whole system. **Why these are written rather than derived.** The generator can count artifacts, plot growth and list files. It cannot say what a release was *for*. These are frozen prose: written once, about something finished, never updated. That is what makes them safe to hand-write — the thing they describe has stopped moving, so they cannot rot. **Owner: read and correct.** Anything you disagree with is wrong by definition; these are your releases. Once approved they're frozen and nobody edits them again. **A note on honesty.** For 2.5.0 and 2.6.0 I was not present and the archive holds one document each. Those accounts are drawn from the changelog and the code, and they say less because less was written down. That's the truthful position, and it matches the sparse-room treatment those rooms already get. --- ## 2.5.0 — Living Operating System *Shipped 26 June 2026 · 747 checks* The release where the terminal stopped being a single-device toy. Accounts and cloud saves arrived, along with the rule that still governs them: uploading never overwrites, and loading or deleting always asks first. Sign-in was optional from the start and remains so — the terminal has never required an account to work, and the offline path was never allowed to become second-class. Much of what shipped here is invisible and turned out to be load-bearing. Every save became checksummed and version-stamped, with three rolling backups taken before any load. Remote controls arrived that can pause cloud or AI features without shipping an update. Both still exist, and both mattered later: the backup ring became the recovery path when a phone evicts storage, and the remote pause is why a broken networked feature never needs a deploy to disable. The API key policy was set here too, and hasn't moved since — your key, your device, synced only if you explicitly ask. ## 2.6.0 — Content, Crafting & Trackers *Shipped 27 June 2026 · 1,078 checks* Shipped **one day** after 2.5.0. That is not a typo, and it is the single most revealing fact about this era of the project. Where 2.5.0 built the machinery, this filled it. Skill books and magazines, a crafting bench with live ingredient checks, traits, Lincoln memorabilia, collectibles wired to their map regions. The reference data roughly doubled — Fallout 3's location list reached ninety sites, New Vegas gained dozens of consumables, vendors and unique apparel — all sourced from the wiki, which was already the rule. The quieter theme is the app learning it was hosting two different games rather than one game with a skin. Panels began showing the data for whichever wasteland you were actually in, the game selection finally survived a reload, and faction reputation was corrected to canonical thresholds after moving in jumps of fifty that skipped every intermediate standing. The AI also got its first real constraint here: invented collectibles started being filtered against the game's own catalogue. The idea that the AI proposes and the app verifies begins in this release. ## 2.8.0 — The Physical Machine *Shipped 10 July 2026 · 2,951 checks* The release where the app became an object. Everything before this was a competent web interface. This is where it grew a casing, a bezel, a glass frame, nav keys with their own LEDs, and a module bay where components physically seat. The design language the rest of the project now reuses — including the museum you are reading this in — was authored here. Underneath the chrome, the more consequential change: a steady conversion of AI-driven features to native, deterministic, offline code. The combat and barter calculators, the command router, the level-up path, the world map, loot import. The app began the process of outgrowing the prompt it was born as, which is a strange thing for a project to do and is visible across the mockups in this room. It also carries the largest planning apparatus in the archive — three hundred artifacts, an audit chain that examined performance, accessibility, code quality, content, mobile behaviour, test strength and its own diegetic voice before a line was written. Those audits measure the *older* code they were run against; they live here because this is the release their work produced. ## 2.8.5 — Code Health & Integrity *In flight* Deliberately not a feature release. It began as cleanup — readability, test health, splitting oversized files — and became the round where the project turned its instruments on itself. Two external models were asked to review the codebase and the process, blind to each other, one specifically instructed to argue the whole apparatus was over-engineered and should be halved. What that produced was uncomfortable and useful. A cache-update guard that had been written, hooked, tested, and silently doing nothing for weeks. Tests that verified code *said* it escaped dangerous input rather than that it did. A rulebook so large that every session read a hundred kilobytes regardless of the task, in which contradictory instructions had gone unnoticed. Documentation confidently describing file sizes off by a factor of forty. The Fallout 3 device skin shipped in this round, and the terminal's data was re-sourced wholesale from the wiki behind guards. But the defining work was structural: a rulebook split into a short contract plus notes retrieved by what you touch, the project's first rule for *retiring* a rule, and the discovery — during ordinary play, in five minutes — that the AI could silently delete eleven different parts of a save. Every one of those was found by looking at the real thing rather than reading about it. That is the lesson of this release, and it is why this museum shows artifacts instead of summaries.