Control Assessment Analysis
Investigative view over the offline corpus
(app/Data/training/control_training_data.jsonl) — controls
under RCSA assessment mapped to their risks, issues, events and the human
rating + rationale. Use it to sanity-check that we actually understand
the self-assessment use case.
python data_ingestion/fetch_control_training_data.py --env p --with-history
(after az login), then reload.
Batch audit vs. use-case spec (RUS-006)
The latest assessment per control, checked against the requirements the coach is meant to enforce. An Effective control that carries a linked issue/event is not a spec breach (journey step 8) — but the assessor must have considered it. These checks flag where the current corpus falls short.
Drill down: Effective-with-case controls silent on their linked case (0)
| Control | Name | RCSA | Issues | Events | Linked cases |
|---|
Effectiveness rating
The assessment label we want the coach to reason toward.
Mapping coverage
How many controls carry each kind of context/label.
Rating vs. open cases
Do weaker controls actually attract more issues/events? A core hypothesis for the use case.
| Rating | Controls | % with an open case | Avg cases / control |
|---|
Rationale comment length
Signal quality — how rich is the human narrative we train on.
Assessment history depth
Prior assessments per control (how much longitudinal signal exists).