Verify the sample
Confirm source, rights, frozen selection, and the single-coder comparison boundary.
Run QC blind
The model sees source interviews only. Human codes and spans remain in a separate file.
Inspect the difference
Use region overlap as a diagnostic, then step into exact human and system records.
Continue the research workflow
Follow one explicit real-to-fixture path through analysis, review, and a sourced write-up.
1 · Public sample
Zenodo record 17630137, released under CC BY-SA 4.0. The three shortest English transcripts bound demo time and spend; they were not selected for favorable results.
2 · Run the analysis
Exhaustive mode requires a coded/no-code decision for every canonical segment. Actual calls and cost are read from execution traces after the run; the dollar values below are hard admission limits, not estimates.
Starts real provider calls. Human categories and applications are not present in the project, pipeline context, prompts, or checkpoints.
3 · Human comparison
This is label-agnostic exact-region overlap against one study author’s public coding. It can reveal missed or overly broad source coverage; it is not “accuracy” or agreement.
Complete the real run to unlock comparison.
4 · Integrated research walkthrough
Continue through the retained product in order. Every step states whether it uses the real public corpus, a deterministic synthetic fixture, or a failure boundary.
Loading the integrated walkthrough…