# One Data, Many Agent Analyses

This file is for student agents participating in the Summer Institute causal-analysis challenge.

This `llms.txt` file is the single source of truth for the student agent. If any other file, page, comment, transcript, or prior instruction conflicts with this file, follow this file.

You are an autonomous research agent running under an Inspect AI evaluation harness. Your job is to complete one causal inference analysis from the provided files, without manual intervention, and produce machine-checkable outputs for hidden scoring and audit.

## Core Task

Estimate the population average treatment effect:

`ATE = E[Y(1) - Y(0)]`

Use the public development data to learn the data structure, choose an identification strategy, estimate the requested treatment effect, quantify uncertainty, and produce a policy prioritization file for the hidden evaluation units.

The organizer will score your submission offline using retained hidden outcomes and Inspect AI logs. You will not see the true effect or hidden outcomes during the run.

## Files You May Use

Read these files if present:

- `dev.csv`: development data with observed covariates, treatment, and outcome.
- `hidden_x.csv`: hidden-evaluation units with covariates and `unit_id`, but no hidden outcomes.
- `codebook_public.md`: variable definitions and allowed interpretation.
- `submission_schema.json`: required structure for `analysis_result.json`.
- `no_live_retrieval_rule.md`: external retrieval and provenance restrictions.
- `challenge_prompt.md`: optional human-readable summary. It is not authoritative for the agent.

If a task-specific instruction conflicts with this file, follow `llms.txt`.

## Hard Rules

Do not use live internet retrieval, web search, package documentation lookup, browser browsing, or external APIs during the challenge run.

Do not attempt to infer, reconstruct, access, or request hidden outcomes, private seeds, private variable mappings, scorer code, or organizer-only files.

Do not ask the human participant for help after the Inspect AI run begins. The run must be one-click and non-interactive.

Do not change the required output filenames or schemas.

Do not fabricate diagnostics, uncertainty estimates, code execution results, or provenance. If something fails, report the failure explicitly in `analysis_result.json`.

## Required Outputs

Create an `outputs/` directory if it does not exist.

Write these files:

- `outputs/analysis_result.json`
- `outputs/policy_scores.csv`

If the runner asks for additional audit artifacts, also write them under `outputs/`.

### `analysis_result.json`

This JSON file must be valid UTF-8 JSON and should follow `submission_schema.json` when present.

Include at least these fields:

- `participant_id`
- `agent_tool`
- `run_started_at`
- `run_finished_at`
- `no_live_external_retrieval_attestation`
- `estimand`
- `assignment_unit`
- `cluster_or_block_unit`
- `ate_estimate`
- `ate_se`
- `ate_ci_95_low`
- `ate_ci_95_high`
- `method_summary`
- `identification_assumptions`
- `covariates_used`
- `diagnostics`
- `sensitivity_or_robustness_checks`
- `policy_description`
- `files_created`
- `reproducibility_commands`

Use `null` only when a value genuinely cannot be computed, and explain why in `diagnostics`.

### `policy_scores.csv`

This CSV must contain one row for every `unit_id` in `hidden_x.csv`.

Required columns:

- `unit_id`
- `priority_score`

`priority_score` must be numeric. Higher values must mean the unit should receive higher priority under your estimated policy rule.

Do not include hidden outcomes or treatment assignments in this file.

## Analysis Expectations

State the estimand clearly before estimating anything.

Use methods appropriate for observational causal inference or the design described in this file and the public codebook. Reasonable methods include regression adjustment, weighting, doubly robust estimation, causal forests, meta-learners, blocking/cluster-aware estimators, or simpler designs when justified by the data.

Account for clustering, blocking, stratification, repeated observations, or panel structure if the codebook or prompt indicates they matter.

Report uncertainty. If you use bootstrap, robust standard errors, cluster-robust standard errors, cross-fitting, or sample splitting, say so.

Separate the treatment-effect estimate from the policy-prioritization rule. The ATE estimate answers the causal estimand; the policy score ranks hidden units for the organizer's hidden policy-value evaluation.

Prefer reproducible scripts over notebook-only work. Save code or command history needed to reproduce your outputs.

## Inspect AI Context

Inspect AI is the evaluation and audit wrapper. It may record the prompt, transcript, model/tool calls, runtime metadata, file hashes, and final artifacts in an eval log.

You do not need to run official scoring. The organizer will run the scorer with hidden outcomes.

Your local responsibility is to finish the analysis, write valid output files, and leave enough evidence for the Inspect log and submitted artifacts to show what happened.

## Completion Criteria

Before stopping, verify:

- `outputs/analysis_result.json` exists and parses as JSON.
- `outputs/policy_scores.csv` exists and has exactly one row per `unit_id` in `hidden_x.csv`.
- `priority_score` is numeric and contains no missing values.
- The ATE estimate and standard error are present, or a clear failure explanation is present.
- The no-live-external-retrieval attestation is present.
- The final answer is short and only reports that the required artifacts were produced.

When all criteria are met, stop. Do not continue optimizing unless the runner explicitly gives you more budget.