Inheriting an Analytical Estate: Audit to Verification Memo.

Thabo Molefe runs operations at a South African wholesale plant nursery. His analyst left in January, and the dashboards she built now produce numbers he cannot verify — so he needs an independent audit of everything she left behind.

The discipline skills, every one of them turned around. You don't author metric definitions — you read someone else's and find "active customer" implemented three different ways across the estate. You don't build dashboards — you check which of the 14 are actually used, and which are trustworthy. You don't design an experiment — you replicate one from a one-page write-up and measure what happened in the months since rollout. You don't write analysis — you re-run a broken notebook and decide whether its conclusion was ever supported. The deliverable is a verification memo: what works, what's broken, what cannot be verified, and a fix/rebuild/retire plan per artifact.

The AI-direction lesson is the one the whole track has been building toward. AI verifies what you ask it to verify, one artifact at a time. It will not, on its own, notice that a metric definition and the SQL behind a dashboard silently disagree, or that a hardcoded filter froze a business rule that has since changed. Coherence across the chain — definition to query to dashboard to the decision a stakeholder makes — is the work AI cannot do for you. You design the coverage, you hold the chain, and you decide which judgments are not AI's to make at all.

Your Role

You're the external auditor. You're delivering a verification memo with a confidence level on every finding, a remediation plan, and the durable infrastructure Greenfields never had — because the previous analyst left no version control, no documentation, and nothing for the next person to inherit.

What's different is the direction of the work. Every project so far asked "how do I build this?" This one asks "how did someone else build this, and what went wrong?" No colleague steps in unprompted here — the inherited system is the investigation, and the analyst who built it is gone. You can still open the chat for a senior auditor's sanity check if you're stuck, but the judgment in the memo is yours.

What's New

Last time you built Bwalya's Streamlit energy-audit calculator from an empty workspace — tool selection, configuration-driven methodology, validation against 85 installations, and handoff documentation for a non-developer.

This time the workspace is not empty. It's full of someone else's undocumented work, and your job is to figure out what to trust. You've never inherited an estate like this, never audited metric definitions against SQL you didn't write, never replicated an experiment from a write-up, never written a memo that grades its own confidence and says plainly what it could not verify.

The hard part is restraint and judgment under uncertainty. The instinct will be to fix everything you find. The more valuable deliverable is the memo — and one of its most important sections is "what could not be verified." Knowing what you don't know, and saying so to Thabo, is the professional skill this project tests.

Tools

• Established and carried forward: Claude Code, Codex CLI (cross-AI review of every Metabase query and the portability test), DuckDB with the read-only MCP server, Python (pandas, numpy), Jupyter, Git and GitHub, and the four-skill library.
• Metabase as an inherited environment — new; not a build target. You read the previous analyst's 14 dashboards, the queries behind them, and the usage logs.
• Metabase REST API reference and the metric-definitions doc as a Markdown export — the shipped materials package the REST API surface as a reference document (reference/metabase-rest-api-reference.md) plus pre-extracted dashboard SQL under inherited/dashboards/, and the Google Doc as inherited/metric-definitions.gdoc.md. You read these directly rather than running a live API call; the units name what a live call would look like and what to verify against the reference.
• inheritance-audit-review and data-quality-monitor skills — new; you author both in the final unit, encoding the audit method and the going-forward monitoring for the next analyst.
• CLAUDE.md, AGENTS.md, and infrastructure/connectivity.md — established patterns, authored fresh here because Greenfields had none.

Materials

• The forwarded email chain (first-contact-email-chain.md) — Thabo's terse forward of the warehouse manager's complaint. The entry artifact; first contact is in chat.
• The inherited estate under inherited/ — the metric-definitions Google Doc as Markdown (inherited/metric-definitions.gdoc.md), one SQL file per Metabase query (inherited/dashboards/), the 60-day usage logs, the Q3 margin notebook, the free-shipping experiment write-up and results, the silently-failing ETL logs, five years of dispatch snapshots (with a visible two-month gap), and the analytical database (inherited/data/greenfields.duckdb).
• Reference primers at reference/ — the six-stage audit method, the verification-memo structure, coherence breaks across the chain, dashboard adoption-and-fitness, experiment replication with post-rollout impact measurement, metric-implementation comparison, the Metabase API reference, and Greenfields domain context.
• Starter layout — an empty audit/ directory with skeletons for every deliverable and the two new skills, a scaffolded pre-commit config, and the carried-forward .claude/ infrastructure.

Download the materials zip (https://learntodriveai.dev/materials/analytics/p-21/materials.zip) and unzip it to ~/dev/analytics/p-21.