Decisions
Thandi Mokoena, you are the signaller on shift. Signalbox does not ask you to read diffs. It asks you the questions that need judgement, and it checks a small random sample of the rest.
7 decisions, 5 audit samples, chosen from 18,302 changes since yesterday. That is about 1 item per 1,500 changes. Everything else landed under policy.
Today's attention budget: 45 min · 31 min left
Open items: 6 decisions and 5 audit samples. Times are estimates per item. Deciding early frees budget for tomorrow.
Needs your judgement
Each question arrives with the options agents already explored, the evidence for each, and what happens next. Agents mark their recommendation. You choose.
Decided earlier today
Sampled audit
Like acceptance sampling in a factory. You inspect a few changes drawn at random, weighted by risk. The defect rate you find sets how much autonomy each fleet gets.
L3 fleets: 1 in 50. L4 fleets: 1 in 500.
High blast radius x6. New model version x4. Coverage under 60% x3.
0 of 5 inspected. About 1 min each.
Trust: what the audits say
A p-chart of the defect rate found in audits, per fleet, for the last 14 days. Points inside the limits are noise. A point above the upper control limit is a real change, and autonomy drops by one level without a meeting.
Defect rate found in audit
Autonomy by fleet
Updated 09:30 from 14 days of audits| Fleet | Autonomy | Defect rate, 14 d | Trend | Note |
|---|
Why this is different
- Attention is the budget. Signalbox gives each human a fixed number of minutes a day and spends them on two things only: decisions and audit samples. Everything else lands under policy.
- GitHub asks for the wrong thing. It sends a review request for each pull request and a notification for each event, sorted by recency. Reviewers approve diffs they cannot fully read, and the inbox grows with the agents.
- It breaks at 100k agents. 18,302 changes a day at 5 minutes each is about 1,500 hours of review. No team can do that, so review becomes a rubber stamp.
- Decisions arrive pre-explored. You get a question, two or three options that agents already built and tested, the evidence for each, and what happens next. You supply taste, policy and judgement.
- Audit sets autonomy. A random, risk-weighted sample gives a measured defect rate. The control chart separates noise from a real shift, so trust rises and falls with evidence, fleet by fleet.