This research agenda is the public work plan for Too Early To Say, Victoria Cholette's independent applied-economics lab. AI agents draft code, search records, and run candidate analyses. Victoria sets the economic question, identification strategy, checks, and interpretation.
The agenda is for economists, policy teams, and research partners who need evidence about where AI saves work, where it creates review costs, and which institutional checks hold up. A tool can finish a task quickly, while checking the result can erase the gain.
The research program tests whether AI shifts the cost of policy research from execution to specification, verification, and judgment.
Status reflects the evidence currently available. Validating work remains provisional. Next work has an approved question or design. Queued work follows the core studies. Parked work waits for an experiment or named policy case.
Evidence on time saved, errors introduced, and review work required.
In Difference-in-Differences in Python, a planted effect of 1.60 tests whether the estimator answers the intended economic question.
Agents: implement and rerun candidates. Researcher: defines the estimand, answer key, and failure rule.
Policy findings with traceable data, methods, and limits.
In The Food Desert Myth, SNAP participation varies 4.7-fold across neighborhoods with equal store access.
Agents: assemble and check source files. Researcher: defines access, defends the comparison, and interprets the finding.
Reusable checks for clean-looking failures.
In Instrumental Variables in Python, an F-statistic of 2,051 coexists with a causally invalid estimate.
Agents: run estimators and diagnostics. Researcher: defines the data-generating process and decides which claim the evidence supports.
The release standardizes the existing estimator cases into a clean-session runner, saved outputs, documented failure conditions, and a citation record.
The Food Desert Myth estimate is being traced through the article, code, data, findings page, metadata, and distribution copy.
The current analysis produces a provisional spending-gap estimate. The value stays out of the public findings while source reconciliation, model checks, and interpretation review continue.
Thirty analyses will answer the same policy question under a frozen brief. The study measures convergence, error, review time, and dependence on expert specification.
The published workflow essay offers a provisional estimate. This study times verification, rework, and accepted output under a frozen task protocol.
The published IV case shows that a strong first stage cannot validate exclusion. The next case turns to genuinely weak instruments and robust inference.
When an agent constructs a treatment, outcome, or covariate, how does classification error move the policy estimate?
How can a policy claim retain its source, units, population, and caveat from analysis through public communication?
Which agent-written checks survive review by an economist who did not design the workflow?
Generic workflow posts, model leaderboards, and standalone tutorials remain parked until an experiment or named policy case supports each claim. Completeness remains the coverage goal, but publication requires a checkable contribution.
Every agenda item extends a published case. Each foundation links a result, a research decision, and the limit the next study must address.
Grounds the AI and evidence-production program with a planted estimand and a comparison failure.
4.7×Grounds the policy-applications program with a result that changes when access is measured beyond proximity.
2.62Grounds the measurement-infrastructure program with a clean diagnostic and a causally invalid answer.
4Grounds the flagship harness with four verification layers and an explicit researcher approval step.
A contribution clears the agenda only when the claim is tied to a named case, the test can fail, and the evidence package says what a reader can inspect.
Economists, policy teams, and research partners can contribute a domain case, rerun a protocol, adjudicate an error, or challenge a publication gate.