← Lab — Found while building — 2026-08-20
Will it even read my file?
12 formats, detected automatically, and you see the parse — format, trace count, what was dropped — before a cent is spent.
224 tests, hostile files includedUpload and hope
See the parse, then decide to spend
What we did
Detection ranks 12 sniffers over a sample and never lets one format claim another's. Parsing streams line by line, so a 200MB upload doesn't take the tab down.
What came back
Malformed JSON reports how many lines were skipped. An HTML error page saved as .json is rejected with a preview. Parquet is caught by magic bytes, a single enormous line is clamped, mixed formats are classified per line, and a file that parses to zero traces gets a specific error.
What it changed
You get a parse preview — format, confidence, trace count, warnings — before any model call. If detection is unsure, it says so first.
What it doesn't show
Not one of the 12? Generic JSON usually still works and CSV covers most exports. Counts and timings are from the repo as of 2026-08-20.
Other notes
- Can it find failures nobody told it about?6/6
- On real traces, does it agree with human experts?87.2%
- Does its taxonomy match one researchers published?5
- What happens when the judge disagrees with the answer key?13 of 30
- What does a run report when a third of it never finished?183/300
- How much of your trace does the judge actually read?644,000
- Is the quoted evidence actually the evidence?2
Run it on your own traces.
Free, no account, parsed in your browser.