Part 6 of 12 · 2 min read
What the simulation measured
Items shown per view, approval-policy read rates, and the ask-margin trade-off.
How to read the view-items table: all rows describe the same simulated 82-event run; only the view mode differs.
| View of a simulated run (mean 82 events) | Items shown by default |
|---|---|
| Raw event log | 81.7 |
| Grouped timeline | 25.7 |
| Status line, attention items and changes | 1.4 |
| Unresolved approvals and unrecovered errors visible in the default view | in every simulated run (tested) |
What to notice: the fourth row isn't a count at all. It's the invariant that holds regardless of which of the first three rows is showing. That's the actual point of the design, not a side detail.
How to read the approval-policy table: all three rows face the same 20-action, 2-risky-action session under the same assumed fatigue model; only the policy differs.
| Approval policy (20 actions, 2 risky, assumed fatigue) | Prompts per session | Risky actions actually read |
|---|---|---|
| Ask about everything | 20.0 | 52% |
| Tiered, one prompt each | 6.4 | 79% |
| Tiered and batched | 3.0 | 86% |
What to notice: fewer prompts and a higher read rate move together here, not in opposite directions. Unlike the endpointing trade-off in voice interfaces, cutting prompt count is not a cost paid against safety. It's how the safety is achieved.
How to read the ask-margin table: each row is a different margin threshold for when to ask instead of act; the two right-hand columns move in opposite directions as the margin widens.
| Ask when margin between top two readings is below | Ask rate | Wrong action rate |
|---|---|---|
| never (margin 0.0) | 0% | 13.3% |
| 0.1 | 22% | 4.9% |
| 0.2 | 44% | 1.5% |
| 0.3 | 65% | 0.3% |
| 0.4 | 82% | 0.0% |
What to notice: wrong actions keep falling all the way to the last row, but the ask rate it costs keeps rising just as steadily. There's no margin where both numbers are small. The right choice depends on how expensive a wrong action actually is, not on an abstract preference for fewer questions.
{{svg:layers|Status, needs-you, timeline, and changes with controls.}} {{svg:results|Items shown, approval fatigue and the ask-or-act trade-off.}}