Domain Guide¶
Guide Maps¶
graph LR
raw["service_incidents.csv"] --> fields["Operational columns"]
fields --> target["Escalated outcome"]
target --> pipeline["DVC pipeline"]
pipeline --> publish["publish/v1/"]
flowchart LR
question["What is this repository modeling?"] --> domain["Read the incident domain contract"]
domain --> columns["Name the features, target, and review surfaces"]
columns --> route["Only then inspect the pipeline and publish boundary"]
Use this guide when the capstone feels structurally clear but the modeled problem still feels abstract. The goal is to make the incident-escalation story concrete before you reason about DVC state and review routes.
What the raw dataset represents¶
The repository models service incidents that may or may not escalate into a broader operational response. Each row describes one incident with a small feature set that a team could plausibly review before deciding whether closer intervention is likely.
Column meanings¶
| Column | Meaning | Why it matters |
|---|---|---|
incident_id |
stable incident identifier | keeps split logic and review examples tied to a real record |
team |
owning service or platform team | lets prediction review stay anchored in operational ownership |
backlog_days |
age of the incident in days | captures unresolved operational pressure |
reopened_count |
number of reopen events | captures churn and instability |
integration_touchpoints |
number of connected systems involved | captures coordination breadth |
customer_tier |
customer criticality bucket | captures business impact pressure |
weekend_handoff |
whether the incident crossed a weekend handoff | captures time-and-coordination friction |
severity_score |
numeric severity proxy | captures direct operational urgency |
escalated |
whether the incident escalated | is the target outcome the model predicts |
What the publish bundle is trying to help a reviewer answer¶
- what population the model saw and evaluated
- what threshold was used for the promoted decision policy
- which eval rows were predicted correctly or incorrectly
- whether the promoted metrics describe a problem a human can still reason about later
Best companion guides¶
- read STATE_LAYER_GUIDE.md when the domain is clear but the repository state layers are not
- read STAGE_CONTRACT_GUIDE.md when the next question is where each domain fact gets transformed
- read PUBLISH_CONTRACT.md when the next question is which domain evidence survives into
publish/v1/