Data Table Analysis

Data Table Analysis Skill

Generate accurate, source-grounded tables and computed quantitative summaries using Python/pandas. This skill produces text artifacts that can be read and included in the final report.

Required Execution Standard

To ensure the calculation is reproducible and useful, you MUST:

1. Structure Inputs: Convert facts from research notes or the user request into explicit rows before running pandas.
2. Preserve Provenance: Keep source URLs, filing names, or note references in the input table when available.
3. Normalize Units: Convert currencies, magnitudes, periods, and date labels into consistent fields before comparing values.
4. Compute Deterministically: Call the execute tool to run Python/pandas for arithmetic, rankings, growth rates, aggregates, and formatting. Do not hand-compute these values in prose.
5. Return Text Outputs: Include the markdown, CSV, or JSON output in your returned ResearchNotes (e.g. a ResearchFinding's evidence and/or narrative_notes). Do not call write_file; run_research_batch persists your returned notes.
6. Report Caveats: Include assumptions, missing values, restatements, estimated figures, or non-comparable metrics in the output notes.

Data honesty

The table is the trustworthy, gap-aware deliverable that any downstream chart depends on, so it must be honest about what is and isn't known:

1. Per-cell status: treat each value as reported, estimate, or not disclosed. Leave undisclosed cells explicitly empty (e.g. —); never fabricate or infer a number to fill a gap, and never carry a prior period forward to hide one.
2. One metric definition: compare like with like. If sources use different definitions (e.g. "cash paid for property and equipment" vs "capital expenditures including finance leases"), keep them in separate rows/columns or pick one and label it - do not silently blend definitions into a single series.
3. Surface coverage: in the notes, state how many cells are reported vs estimated vs undisclosed, so the reader (and any chart built from this table) can judge how much weight it bears.

Execution Flow

1. Gather candidate facts from researcher outputs, user-provided data, or source excerpts.

2. Create a normalized input table with one row per comparable observation. Prefer explicit CSV or JSON records embedded in the Python script. If the source rows are in /shared/..., call read_file first and embed the returned content in the script, or write a sandbox-local input file under your sandbox working directory (sandbox_workdir; e.g. /sandbox on OpenShell or /workspace on Modal). Sandbox code cannot open /shared/... directly.

3. Call the execute tool with a Python command or script that:

   • imports pandas,
   • builds a DataFrame from the normalized rows,
   • validates data types,
   • standardizes units and period labels,
   • computes the requested metrics,
   • prints markdown, CSV, JSON, and data-quality notes as text.
   • uses your sandbox working directory (sandbox_workdir) for any sandbox-local input or output files, and writes any script file at the job-unique path your instructions specify (the <job_id>_<name>.py form) so a shared sandbox never reuses a stale leftover from another job.
   • does not read from or write to /shared/... inside the sandbox process.

4. Inspect the execute output. If the code fails, fix the code and call execute again. Do not continue with hand-computed fallback tables unless the sandbox or pandas is unavailable.

5. Return the final outputs from the successful execute run in your ResearchNotes — put the markdown table, CSV, or JSON into a ResearchFinding's evidence and/or narrative_notes. Do not call write_file/edit_file; run_research_batch persists your returned notes under /shared/ automatically.

6. In the response or report, cite the original sources for the input figures. Computed columns should be clearly labeled as calculations.

Required Tool Use: For tasks that request calculated tables, growth rates, rankings, summary statistics, normalization, CSV, or JSON, this skill requires at least one execute call that runs Python/pandas before writing the final artifacts.