Verification Checklist for AI-Assisted Empirical Work

Run before any result, citation, or claim is treated as final. Each item is pass or fail; a skipped item is a fail. The agent satisfies this checklist before a human reads the output. · Project: Date:

A. Citations

#	Check	P/F
1	Every citation was fetched from its source, never accepted from generation: the DOI or URL resolves	☐
2	Author names and title match the fetched source exactly	☐
3	Journal, year, volume, and pages match the fetched source	☐
4	The claim attributed to each paper appears in that paper	☐

Unchecked citations accumulate hallucination debt: one fabricated reference becomes the foundation for claims built on top of it.

B. Code and estimators: the planted-truth routine

#	Check	P/F
5	The estimand is written down from the paper or canonical implementation before reading the generated code	☐
6	The low-visibility choices are named in the spec: comparison set, weighting, transformation window, indexing	☐
7	A controlled data-generating process exists with a planted true value for the target parameter	☐
8	The implementation recovers the planted value across many simulated draws, within a stated tolerance	☐
9	The simulated data break the easy symmetries (unequal group sizes, unbalanced timing, missing cells)	☐
10	The code was read line by line against the source on every weight, aggregation rule, and index	☐

The kit's validate-estimator skill walks an agent through items 5–10. Some errors appear only in the reading, never in the simulation output.

C. Empirical claims

#	Check	P/F
11	Every number in prose traces to a computed value in a saved output file	☐
12	Every factual claim about external data names its source, and the source says what the claim says	☐
13	Anything published or shared externally received the highest scrutiny	☐

End-of-session capture (five minutes, before closing)

What changed: files touched, functions added, results produced.

Why it changed: reasons behind non-obvious choices; constraints and trade-offs.

What comes next: concrete next tasks, written while still clear.

What failed: approaches rejected, each with its reason, so nobody re-explores a dead end.

Week-One Glossary

Ten terms from the first week of using Claude Code on an empirical project. Keep beside the keyboard; retire when they feel obvious.

commit: A saved snapshot of the project at a point in time, recorded by git with a short message saying what changed. A commit right after each working result makes any past state of the analysis recoverable.
diff: The line-by-line difference between two versions of a file: what was added, what was removed. Reading the diff is how AI-written changes get reviewed before they are trusted.
PR (pull request): A proposed batch of commits packaged for review before it merges into the main line of the project. A coauthor can read exactly what changed in the analysis and approve it before it becomes the record.
.gitignore: A text file listing what git must never track: data files, credentials, caches, generated output. It keeps restricted-use microdata and API keys out of a repository that may one day ship publicly.
DuckDB: A database engine that runs inside a script or notebook, no server to install, and queries files directly with SQL. It aggregates a multi-gigabyte extract on a laptop without loading the whole file into memory.
parquet: A compressed, column-oriented file format for tabular data. Column types survive between sessions, so FIPS codes saved as strings stay strings, and large panels load fast.
FIPS: Federal Information Processing Standards codes: numeric identifiers for states (2 digits), counties (5), census tracts (11). The merge keys of US regional data; they carry leading zeros, so storing them as integers silently corrupts every merge downstream.
context window: The bounded amount of text a model can consider at once; every file read and every exchange consumes part of it. The window is a budget: reading ten files costs more than deeply analyzing one.
subagent: A helper the agent creates with its own separate context window; it executes a focused task and reports back a summary. Like delegating retrieval to an RA: dead ends stay in the helper's workspace, only the summary lands in the main session.
compaction: The automatic summarization Claude Code performs when a session nears its context limit. Summaries lose nuance: the decision survives, the three reasons an alternative was rejected may not. Decisions belong in notes files that persist.