A hard problem gets lift when every decision inside it is laid out as a series of alternatives — each with what it costs, whether it has ever worked, and the place it was read from. You can move up and down the parent and child chain, stop at any rung, and see the options for that rung alone. Where an answer cannot be derived from the record, the rung is marked rather than filled in.
One real descent. The alternatives narrow the candidates in cost order; when nothing in the material determines an answer, the output is a proven statement about what is not there, with the place that proves it. Nothing here is typed into the page — it is read from the run's own output file.
Every mechanism in the zoo is stated the same way: a signature that is allowed to refuse, the condition that must already hold, and how you would know an answer is right. Then the verdict, and the measurement that fixed it. This is the similarity-score mechanism — it earns a place as the thing that proposes candidates, and is barred from being the thing that decides.
Every number above is injected from the generated data file at load. The page source contains no figure of its own, and a check scans it to make sure — a page that states a number its data does not contain is staging its own success.
The zoo is built by reading the ladders declared in code and the alternate-path sections of each use case. Every node carries the file and line it came from, and the check refuses if the published file differs from a fresh build — which is what happens if someone edits the output instead of the source.
| A node whose file and line do not resolve on disk |
| A status that is not one of the derived kinds |
| A status with no statement of how it was derived |
| A rung marked as worked without a recorded count behind it |
| A figure appearing in the page source instead of in the data |
| A published data file that differs from a fresh build |
Regenerate and verify with