Reviewing a run
Watch it run while one is in flight, or Review the run afterwards, opens the run review screen: one version of the analysis, step by step.
The ladder
The left rail lists the nine steps in order, grouped into five phases — ingest, frame and retrieve, structure, score and decide, compose and commit. Each step shows what it is for:
| Step | Its job |
|---|---|
| Knowledge base | The same party, spelled four ways, resolved |
| Case summary | State the matter neutrally |
| Research checklist | The questions this matter raises |
| Similar cases | Analogous decisions |
| Legal research | Authority for the questions asked |
| Theory and damages | The theory, the damages, the defences to expect |
| Case assessment | Price the judgement already made |
| Memorandum | The prose an attorney reads first |
| Developments | What changed, and what to do about it |
Watching live
While a run is going the header shows Live, and a ticker across the top tracks elapsed time, steps finished, current case strength, and steps needing a look.
Each panel fills in the moment its step writes its row — you do not have to wait for the run to finish. Scroll away and the screen detaches from the live step rather than dragging you along; scroll back to follow again.
Required and skippable steps
Each step says which it is.
Required — if it fails, the run ends there.
Can be skipped — if it fails the run carries on without it, and the step records why. You will see the run carried on without it; nothing below was blocked.
Carried forward
A step marked carried forward from run 6 was reused: nothing it depends on had changed. That is what makes an incremental run cheap.
Not recorded
A step with Not recorded finished without writing a row — usually because the step did not exist when that run ran. It is shown anyway so the ladder reads the same on every run.
Inside the knowledge base step
The first step is worth opening on its own. It shows the resolved parties with their aliases (aka …), the fact and event counts, what is new since the last run, the chronology it built, and the evidence list every later stage cites into.
Stages cite an index into that list, never a quote they composed. The index resolves back to the fact, the document and the page — so the worst a model can do is cite the wrong one, which points at a real passage that visibly does not support the claim.
If the generative merge was unavailable, the step says so: parties were resolved by normalisation alone, and the knowledge base is usable and may be coarser than usual.
Audit history
Audit history on the Analysis tab records everything that happened to the analysis on this matter, including what the engine itself did — actor Analysis engine. Export CSV takes the lot.