The 47-minute recovery.
Meet Mara, the operator of a small software portfolio, on the morning one quiet regression reached three different systems. Follow the evidence and learn Apphelm by doing.
- 47 min signal to decision
- 3 proofs before recommendation
- 48 hr recovery window
Northstar House
A fictional studio running InvoiceKit, VocaMax, Grape, One Client and Portfolioix with a six-person team.
Mara
Product lead, reluctant dashboard collector, and the person everyone asks when a graph looks strange.
Proof before prose
No recommendation earns attention until its metric, session, release or monitor evidence can be opened.
- Chapter 01
A quiet alarm
Mara’s Monday began with five products and one sentence: InvoiceKit activation is 18% below baseline. No wall of charts. No red siren. The portfolio home put the exception above the healthy noise and showed the value exposed.
What this teachesPortfolio shows where attention is worth spending.
Open the portfolio - Chapter 02
Proof, not a hunch
The Action Inbox ranked the finding first. Mara opened it and found three independent strands: an accepted-event drop, rising delivery latency, and a release aligned to the first bad minute.
What this teachesEvery finding stays attached to inspectable evidence.
Inspect the finding - Chapter 03
The receipts behind the line
Analytics narrowed the change to one product and one event type. Mara opened the newest delivery receipts, checked their timestamps and privacy modes, and compared the same window with the previous period.
What this teachesTrends stay connected to the production records that produced them.
Open analytics - Chapter 04
The second signal
Reliability showed signup API p95 climbing from 620 ms to 1.8 seconds after the same release. Apphelm called it a likely cause—not a confirmed one—and kept desktop Chrome’s stable activation as counter-evidence.
What this teachesCorrelation is classified honestly; counter-evidence remains visible.
Open reliability - Chapter 05
Beam follows the trail
Mara asked Beam one plain question. Beam queried approved metrics, sessions and release context. Its answer came back with 91% confidence, the exact queries it ran, and a modest warning: two changes overlapped, so the safest fix was reversible.
What this teachesAI sits beside evidence instead of floating above it.
Ask Beam - Chapter 06
A measured decision
Mara took the evidence into the team’s system of work and started a 48-hour measurement window. Nothing changed production automatically. The team rolled back the API release and watched activation return without declaring victory early.
What this teachesApphelm supports a decision; people confirm consequential changes in the system that owns the work.
Review the evidence
11:20 · two days later
The line came back. The lesson stayed.
Activation recovered to baseline. More importantly, Northstar House kept the evidence trail, the decision, and the measurement window together. The next incident would start with shared context instead of a meeting.