Method
Five projects, one person, all of it between 26 July and 19 August 2026. The method is written down in the repositories, and none of it is fast. It rests on one rule: whatever built a thing does not get to say whether it is good.
The technique is the Gauntlet Loop, named and developed by Matt Shumer.
Umbra Nine was converted from Kart Royale’s codebase. Homie is still being built.
Two frontier models, two labs, comparable results. Halocline is the cleanest reading: 16,883 lines over 17–18 August, built with Grok 4.6, and it accumulated 230 art-direction reviews on the way. Nothing below is specific to a model.
The first prompt named a bar that could not be used — compare it side by side against the actual Mario Kart. The bar was rewritten as a table with numbered bands, and that substitution is the single change that made the loop work.
Almost nothing deserves 88+ on an early round. If you are inclined to give 85, look harder — you are probably missing something.
A critic gets the rubric and the artefact. Not the diff, not the plan, not the intent. On Kart Royale the critics scored rendered screenshots from a real headless build, which is what makes them an oracle rather than a second opinion.
The protocol package is the interface every part of Homie codes against, and nothing edits it as a side effect. Even so, the daemon wrote a field name the television did not read. The television kept its previous screen rather than blanking it, so a stale shelf sat there with nothing logged and both halves behaving exactly as designed. A probe now asserts across that seam using the television’s own reader.
A frame watchdog reported 100% black on frames that had presented perfectly; it was reading a discarded buffer. Every harness now injects a known fault and proves it detects one. A harness whose self-test did not execute is BLOCKED, not passing.
That includes an assumption written into the document that owns the topic. Two of Moonbase One’s design documents argued in prose against a measurement file, and both lost. Homie’s ground-truth file outranks its README on any question about the house, and says which parts of the README are now false.
A cleanup deleted a guard because its comment looked wrong. The comment was wrong. The guard was right for a different reason — multisampling on the buffer the occlusion pass reads as a texture — and 7.6% of frames came back part-black.
Drifting in Kart Royale is measurably faster than a clean lap, 2.03 seconds a lap against a 0.44% noise floor. But 83% of drift attempts never bank a mini-turbo tier, so the ladder that should make it a skill is still decoration. That sits at the top of the backlog rather than being smoothed over.
Halocline accumulated 230 art-direction review documents. This is not a shortcut, and there are defects the loop is structurally unable to see.
The instrument
Homie’s certificate check wrapped its call in a catch returning “no certificate”. The import was missing, so every call threw, and the daemon reported no certificate for one that was present, valid, and had 89 days left.
Occupancy
At four people the typed address was off screen. At thirty-two the QR code was 1637 points down a 1080-point screen. Every harness in the repository put at most one person in the room, and every screenshot anyone had taken was of an empty one.
Stills
Inverted steering, missing mobile controls, a pause menu that ended the race permanently, a crash at ten seconds on a phone. All of it survived three rounds of six reviewers scoring beautiful frames. A human playing found every one.
This site
The hero headline ran at 2.45:1 against a 3:1 floor, and 20 of 68 images served a different photograph to their JPEG fallback than to their AVIF. Both harnesses that would have caught them were written afterwards, which is the wrong order.
None of this needs a larger model or a better tool. It needs a standard written down before the work starts, something separate to judge against it, and the patience to keep going after the thing already looks finished.