The two sentences this product has to answer are short. An engineer is handed a repository they did not write, runs it locally, and sees the whole thing broken down visually. A learner asks what machine learning is and gets it drawn rather than described. Both went from measured failures to working on the same night, and neither for the reason I would have guessed.
The first was finished before anyone could open it
The view that answers the engineer’s sentence had been built and correct for some time. Ask the anatomy panel about the repository node and it returns ten cells partitioning the whole line count exactly, on both axes, with two packages holding 86.5 per cent of it — 130,881 lines and 107,845 against 275,837 at the scan I measured on 9 September, which anyone can divide for themselves. The current scan is larger; the shape has not moved.
Nothing could open it. Anatomy paints into a card’s own slot, keyed to the node that card is for, and the board draws services, datastores and topics. It never draws the repository. So the one node the whole-repository view is about was the one node with no card, and the view sat there rendering correctly for anybody who could reach it, which was nobody.
An entry point is the least glamorous thing in software and this is what it is for. Every part of the feature existed — the computation, the layout, the rendering, the tests — and the product still could not do the thing, because there was no surface on which a person could ask for it. Shipped and unreachable is a state that passes every test you would think to write, since each of them addresses the part rather than the path, and the only instrument that catches it is somebody trying to use the product (the board that drew one connector is the same fault one level down). Making it reachable was a repair. Landing on it after a scan would be a product decision, and that one is Max’s — it changes what every user meets before they have asked for anything.
The second needed the product to stop asking the model to lie
The learner’s sentence was measured at nought of twelve across three surfaces, and it came with something worse than a failure to draw: a fabricated repository fact, delivered as a lesson. The model said machine learning was defined in a specific module of the scanned codebase, naming a real file, which is not true of any codebase.
The cause was ours. The instruction belt tells every teaching turn to ground the concept in this repository and name the real node — unconditionally. For a subject the scan does not contain, that instruction cannot be obeyed honestly. The model was not ignoring the contract; it was following it into a fabrication, which is a different problem with a different fix. Detect that the lesson queue came back empty against an attached graph, and tell that turn the truth: the subject is not here, do not claim it is, teach it from general knowledge.
That removed the fabrication completely — nought of four across four runs. It did not produce a single drawing. Still nought of four.
Which is the finding worth keeping. The model does not call the drawing tool: five calls across 296 benchmark teaching turns, none in twelve runs of this question. A clause that is present, unconditional and unobeyed is this codebase’s oldest failure, and writing a louder clause has never once fixed it. So the product draws. One extra call on that turn shape, everything it returns validated, and a chart claiming a node it should not have voids rather than being stripped — because stripping keeps the model’s claim and hides the evidence that it made one. The caption is written by the product. Four of four drawn, none refused, none answered without a drawing, no fabricated facts.
What I take from having both land the same way
I keep expecting these to be reasoning problems and they keep being plumbing. One was a missing door in front of finished work. The other was an instruction that made honesty impossible, followed by a tool the model would not reach for no matter how the prompt was phrased.
The rule I would defend from both: provenance is not something the answer gets to decide. A caption saying where a picture came from is worth nothing if the thing being audited wrote it, and a diagram that names files is only trustworthy if something other than the model checked that those files exist. That is why the invalid chart voids instead of being quietly cleaned up, and why the caption is not the model’s to write.
Neither loop is finished. The engineer’s view opens now and is not yet what the app lands on; the learner gets a drawing on this question and I have measured one question. But the gap between built and reachable is where both of these were hiding, and it is not a gap any test I had was looking at.