What you're building. An inheritance review for the Data and Performance Manager at a public bus authority in Beirut. A consultant's notebook arrived with the engagement; the consultant has gone, and the deputy thinks something is wrong. You produce three things: a verification map naming each problem, a corrected analysis, and a plain-language brief for the ministry.
What you're practicing. Reading someone else's notebook diagnostically. Naming methodological failures in plain words. Deciding per finding whether the work can be salvaged, must be rebuilt, or cannot be recovered. Re-running corrections under the right discipline. Writing honestly about an overclaim without sounding like a hatchet job.
What you're learning about the loop. Verification has the same shape whether you built the work or inherited it -- but the directing register changes. AI summarizes an inherited analysis as if it were valid; it reads a 97% accuracy figure at face value the same way it reads its own output at face value. It does not diagnose at the methodology level on its own -- not random splits on temporal data, not unfair comparisons, not wrong question types, not causal language used for a descriptive finding. You direct AI to verify rather than summarize, and you hold the methodological coherence the consultant did not.
Your Role
You are the senior data scientist hired for the inheritance review. The frame is set at the start: read the consultant's analysis, diagnose what is wrong, distinguish salvage from redo, execute the corrections, produce a brief Rima can hand to her ministry.
The familiar terrain is everything you have built across the prediction and synthesis projects -- leakage discipline, comparison validity, question-type decomposition, honest-evidence communication. What is new is the seat. You are the investigator, not the builder, and there is no prior reasoning chain of yours to lean on. The methodological judgment comes from reading the work cold.
AI is the tool that runs the verifications you design. It does not decide the salvage strategy, and its first drafts soften bad news in ways that bury severity. You catch those defaults and reframe.
What's New
Last project, you produced a layered communication package for a Senegalese microfinance nonprofit -- three parallel analyses synthesized into one coherent picture, designed for inheritance. The hard part was holding three question types coherent across one deliverable.
This project inverts the engagement object. You inherit a finished but broken artefact instead of building from a brief. The first contact is a forwarded internal email; the deputy has already named operational symptoms in his own register. The deliverables triple: a verification map, a corrected notebook, and a plain-language brief for a non-technical ministry audience.
The hard part is the salvage-vs-redo judgment per finding. Some sections are repairable in place; some must be rebuilt under a different framing because the consultant answered the wrong kind of question; some headline results do not survive any reasonable correction. AI defaults to fixing everything in place -- you decide what gets repaired, what gets rebuilt, and what cannot be recovered.
Tools
Carry-forward. No new packages.
- Python 3.11+ in the conda "ds" environment, Jupyter, pandas, scikit-learn, statsmodels, matplotlib / seaborn, plotly
- MLflow -- you read the consultant's existing runs and log a new
bta-inheritance-reviewexperiment for the corrections - DuckDB, DuckDB MCP server, Jupyter MCP server, MLflow MCP server
- Claude Code with the assumption-checking skill and pre-evaluation hook active (you do not modify these). The structural absence of these in the inherited project is itself a finding
- Codex CLI -- the cross-model review on the load-bearing leakage diagnosis is where this earns its keep
- AGENTS.md, Git / GitHub
The data and the inherited project bundle are new -- ~18 months of route-level operational records from a recently-relaunched 7-line bus system in Beirut, in Lebanese Pound (LBP).
Materials
The project lives at ~/dev/data-science/p-17. The materials zip downloads from https://learntodriveai.dev/materials/data-science/p-17/materials.zip.
What is provided:
- A fresh project memory file (
CLAUDE.md) scoping the inheritance review -- your engagement, not the consultant's - The deputy's memo (
deputy-memo.md) flagging the operational symptoms in plain language - A data dictionary describing the three inherited data exports honestly
- Templates for the three deliverables: a verification map, a salvage-vs-redo judgment, and the plain-language brief for the ministry
- The full inherited project bundle under
inherited/-- the consultant's Jupyter notebook, brief report, three data exports, an under-curatedCLAUDE.md, and an MLflow runs directory present but methodology-version-untagged
No diagnostic findings, no corrected analysis, no salvage verdicts, no draft brief. The diagnosis, the corrections, and the brief are yours to produce.