Redaction Throughput Report

The Redaction throughput report shows how much severance work is happening across your requests and how much of it a machine proposed rather than a human drawing it directly. Use it to see where AI-assisted redaction proposals are piling up and how much a reviewer is actually agreeing with.

What This Report Shows

For the period selected at the top of the panel:

• Marks created, by origin — how much redaction was drawn manually versus proposed by AI Assist.
• Proposal backlog — the number of machine-proposed marks still awaiting a human decision, and the age of the oldest one sitting in that backlog.
• Accept and dismiss rates, per origin — how often a human agreed with what was proposed versus rejected it, split out so AI-proposed marks can be judged separately from manually drawn ones.
• Flag resolution time — how long a flagged mark takes to get resolved once it's raised.
• Pending-approval depth — how much redaction work is currently queued and waiting on approval.

How to Read It

• Use the period selector to set the reporting range; Edit the range inline and click Apply to refresh the figures, and use Compare where offered to see the previous period alongside the current one.
• Read the backlog and oldest-item age together. A growing backlog with an aging oldest item is where proposals are piling up unreviewed, not just where volume is high.
• Treat accept/dismiss rates per origin as a quality signal specifically for AI Assist's proposals — a low accept rate on machine-proposed marks is worth investigating even if overall throughput looks healthy.
• Use pending-approval depth to find where the human-review step is actually the bottleneck, rather than assuming it's redaction itself that's slow.
• Export to Excel to share the figures or analyze them further.

A Caveat on Accept/Dismiss

Accept and dismiss figures are inferred, not recorded: no explicit "accepted" stamp is kept on a mark, so the rate is read from each mark's current status rather than from a logged decision event. Treat it as a close reading of the outcome rather than a literal count of decisions made. Where a scan over a long period hits its row cap, the panel says so and the figure should be read as a floor — narrow the range for an exact count.