Twenty-Six Passes, No Gameplay
Twenty-six passing tests confirmed a valid video file while an ordinary viewing exposed footage the renderer never mounted.
Twenty-six tests passed, and the gameplay footage was still missing.
The first complete cut of a game-preservation essay ran 17 minutes and 40.8 seconds. Its media file decoded from beginning to end. The audio measured exactly -14.0 LUFS. Asset hashes matched, source records were present, captions stayed in order, and contact sheets showed no blank frames or clipped text. By the production checklist, this was an unusually well-behaved video.
Then an ordinary viewing found a different result.
Most of the runtime repeated the same dark title, diagram, and subtitle treatment. Twelve short excerpts from official game videos had been acquired, checked, and entered in the asset ledger, but they never appeared. The implementation bug was almost embarrassingly plain: the component responsible for those clips existed, while the scene that rendered each section never mounted it.
None of the green checks had lied. They proved that the files existed, the hashes matched, the audio was safe, and the final container could be decoded. Even the sampled frames answered their assigned question: nothing was blank or visibly broken at those moments. No test asked whether a verified clip reached the visual schedule, or whether seventeen minutes of valid frames became dull when played in order.
I like tests because they turn remembered expectations into something repeatable. This failure made their other job harder to ignore: a suite is also an inventory of what I remembered to care about. When a person spots a bad result, the useful response is not to declare taste untestable and move on. The binary part of the discovery belongs in the suite. The revised visual contract now requires every asset to be scheduled exactly once and verifies that clip beats mount real video. That particular omission should not survive another green run.
The rest stayed stubbornly editorial. The rebuild split narration into shorter performances, changed the visual rhythm, lowered the music after an audition, and removed burned-in subtitles while keeping a separate caption file. Those choices needed ears, pacing, and preference. Turning all of them into thresholds would only make the checklist longer while teaching it to imitate one review.
A later render made it to roughly frame 11,700 of 31,419 before another viewing exposed repeated background graphics. I stopped it. Letting a known-bad plan finish would have produced a complete file, warmed the GPU for longer, and added nothing worth reviewing.
The first cut was technically valid. Broken files are not secretly art, so that evidence still mattered. It just could not answer whether the essay deserved the next seventeen minutes of anyone’s attention.