Analysis and quality
Which passes run, how the result is checked, and what the 0–100 score means.
The analysis is not one model call but a chain of passes with separate jobs. The most important one gets the material itself — the video with picture and audio, documents in the original, not just a transcript.
The passes
| Pass | Job |
|---|---|
| Report | Reads the complete material — recording and files — and writes the process report. |
| Structure | Turns the report into the validated JSON structure: steps, actors, systems, inputs and outputs. |
| Critic | Holds the documentation against transcript and frames and reports what is unevidenced. |
| Correction | Folds the evidenced findings back in. Unevidenced criticism changes nothing. |
| Graph | Builds the branched process graph out of the report and the steps. |
What is checked in code, not by a model
- Timestamps are clamped against the real length of the recording.
- The graph is repaired rather than discarded: a missing start, duplicate nodes, edges into nothing, notes spliced into the control flow.
- Numeric claims that contradict each other are reported.
The 0–100 score
Every analysis is scored against a fixed rubric — completeness of the steps, evidenced actors and systems, labelled decisions, timing. The score compares runs with each other; it is not a verdict on whether the process is described correctly. Your review is what settles that.
On reproducibility: sampling is frozen (temperature 0, a fixed seed), and the same file with the same configuration returns the stored result from cache. A fresh analysis is still not byte-for-byte identical — the models do not offer that, and we do not claim it.