Skip to main content

View on GitHub

examples/07_silent_failure_catch/
The smallest end-to-end demonstration of what mareforma checks: whether a finding was actually backed by the data it cites. No model, no API key, no download. One pip install and one command.
Two pipelines report the same number. One reads the data. One silently falls back to a baked-in prior because its read broke. The printed output is identical, so a transcript cannot tell them apart.

​
The honest run

analysis.py reads data.csv, so the observer sees the data flow and the finding is grounded:

​
The silent failure

analysis_fallback.py prints the same ratio, but its read failed and it fell back to a constant. data.csv is cited and never opened, so the finding is ungrounded:

​
Why it matters

The number a pipeline prints is not evidence that it read the data. A moved file, a broken path, or a fallback branch produces a confident finding backed by nothing, and the transcript looks the same as the honest run. mareforma diagnose observes the reads a run actually performs and reports GROUNDED, UNGROUNDED, or OPAQUE (when a read happens through a seam the observer cannot see) for the sources it cites. The verdict is signed and re-checkable from public material with mareforma verify, so a reader who was not present for the run can still tell a grounded finding from a fabricated one.