validate-log

name: validate-log description: Scan a Stata log file for errors, warnings, suspicious patterns, and silent failures. Cross-check claimed numerical results against log contents. disable-model-invocation: true argument-hint: "[path/to/log file] [optional: claim to verify]" allowed-tools: ["Bash", "Read", "Grep", "Task"]

Validate a Stata Log

Two modes:

Mode A — log scan: pass only a log path. The skill greps for known error/warning patterns and reports them.

Mode B — claim verification: pass a log path + a numerical claim. The skill delegates to the log-validator agent, which confirms (or denies) the claim against the log.

Mode A: Log Scan

Steps

1. Resolve log path from $ARGUMENTS (or list logs/*.log if not given and ask).

2. Run the standard error grep:

   grep -nE "^r\(\d+\);" logs/<file>.log              # Stata error code lines
   grep -nE "invalid syntax|not found|no observations|conformability error" logs/<file>.log
   grep -nE "no variables defined|file ... already exists" logs/<file>.log
   

3. Check the merge tab (silent dropped rows are a common bug):

   grep -A 5 "_merge" logs/<file>.log
   

   Flag any merge where _merge == 1 (master only) or _merge == 2 (using only) is non-zero AND the do-file does not handle it.

4. Check sample-size funnel (required by econometric-best-practices):

   grep -E "Sample N|after restriction" logs/<file>.log
   

   The log should show N at each restriction step.

5. Check the tail for the most-recent estimation output, sanity-checking magnitude / sign / SE plausibility.

6. Report findings as a structured list:

   • Errors: count + context for each
   • Suspicious patterns: count + context
   • Sample funnel: present / missing
   • Last estimation: spec + headline coef + SE

Mode B: Claim Verification

If $ARGUMENTS includes a claim (any sentence with a number after the log path):

1. Delegate to the log-validator agent:

   Claim: <quoted claim>
   Log: <log path>
   
   Verify per the log-verification-protocol rule.
   

2. Report the agent's verdict (VERIFIED / MISMATCH / UNVERIFIED) verbatim.

Examples

• /validate-log logs/03_analysis_main_regression.log → Mode A: scans for errors and reports.

• /validate-log logs/03_analysis_main_regression.log "ATT = -1.632 (SE 0.584)" → Mode B: confirms the claim is in the log.

• /validate-log (no args) → lists logs/*.log and asks which to scan.

Troubleshooting

• No log files — either the do-file hasn't run, or it doesn't open a log. Fix per stata-coding-conventions.
• Log exists but is .smcl — convert with translate <name>.smcl <name>.log in Stata, or update the do-file to use log using ..., text.
• Grep finds errors but the do-file ran "successfully" — capture likely swallowed an error. Check do-file for capture blocks.

Notes

• This skill is the operational backbone of log-verification-protocol. Use it before any commit that adds numerical claims.
• The log-validator agent (Mode B) refuses to validate fabricated claims — that's the design.