How well it actually works
Every quality number we have, including the ones that do us no favours. These are produced by commands anyone can re-run against a public corpus, they are refreshed by a job that runs nightly whether or not the result is flattering, and the page shows the defects our own measurements found in our own engine.
1. What we can measure, and what we find
The strongest check we have is not "did every construct map" — a construct can map perfectly to code that computes the wrong number. It is: run the original and the conversion on the same bars and compare the values they produce.
2. What we cannot measure
Not every script can be executed exactly. Where ours cannot — because a script depends on behaviour we do not yet reproduce with certainty — we refuse to measure it rather than publish a number that would be read as evidence. Those conversions are reported as Not verified, never as passing, and the output and its warnings are unaffected.
This is the larger share today, and shrinking it is the main line of work.
3. Construct mapping
The older, weaker measure: how many scripts convert with no unmapped construct. It is kept because it covers the whole corpus, not only the part we can execute — and because watching it fall when we tell the truth is a useful discipline.
4. What is not yet true
The limits that matter, stated before anyone has to discover them.
5. How these numbers are produced
Each figure comes from an automated run over a fixed, public corpus of community scripts, repeated nightly against the current engine — not from a hand-picked sample. A release cannot ship unless it is measurably no worse, script by script, than the one before it.
If a conversion of yours disagrees with what this page implies, we would like to see it: hello@stratlate.com.