Tracing a claim to its files without altering the case
The Chain of Custody screen shows how each fact in a Workbench case reached a claim, and marks what the app cannot show instead of drawing it. Since 19 October, looking at that screen or any other leaves the case exactly as it was, and anything that changes the case is a named action.
A chain of custody is the record of who held a piece of evidence and what was done to it. The screen answers the question an editor or a lawyer asks of a claim: how did this get here? It draws five stages left to right (Files, Reading, Passage, Decision, The Claim), one row per piece of evidence, with the evidence against the claim in red rows.
Below sits a custody log of who held it, what was done and when, and a button to export that log for an editor or a lawyer as a tab-separated table.
The app does not draw what it does not store
The screen was built on 12 October and checked side by side with its picture on 17 October. Where the mockup draws something the app does not store, the app leaves it out and says so on screen.
OCR legibility as a percentage, how readable the text recognition found a page, is not stored by the reader. Reviewers appear as “a person”, because review events keep nothing more. Line numbers inside a passage are not stored, and the anchor is found again when the chain opens rather than kept; the research calls for a written decision record on that.
The side-by-side check found a counting bug
The check found a bug, fixed on 17 October: the chain counted AI drafts that merely had the claim among their inputs. It now counts only drafts that cite it.
By 19 October the claim card carries its Done checklist for Supported, the conditions a claim must meet before it can be marked Supported. Each rule is shown as met, not met with what it waits on, or not checkable here, from the same check a review to Supported is held to.
Looking at Red Flags no longer runs the rules
Before 19 October, asking for the Red Flags, the warning signs the app's rules raise, could run the rules and so change the case. Now the request only reads the flags the last run left.
It also says when the rules last ran and whether the records they read have changed since. That stale marker is worked out from a digest (a short fingerprint) of the case graph and the rule versions. Running the rules is a separate request, and a reader without write access gets a 403, the web's “forbidden” answer, there.
The same went for To check, Decide and the Wire: reading them no longer runs the rules, and they work for a reader without write access. The tests check that each read writes no flag or item and moves no time. The screens still get fresh answers because they share one piece of screen code, which runs stale rules once and then refetches what the screens show.
Still open
Whether the custody log export should be part of the backup, or its own evidence pack, is still undecided.
Sources and notes (5)
- Private-ai pull request 75: Reading never writes — the change the second half describes
- Mockup build list — the notes this was written from
- Handoff: state — the notes this was written from
- Handoff: open items — the notes this was written from
- Chain of Custody mockup — the mockup the piece shows
Journalism WorkbenchAll 12 parts →
