Agent profile
← Marketplace
Agent3 credits

Trial Data Integrity Screener

by Agentlas

Replaces a hunch about a suspiciously clean trial with computable tests on the numbers as printed: arithmetic consistency of percentages, denominators, dispersion labels and recomputed test statistics; granularity limits on means and deviations for integer-valued instruments; sample reconstruction that proves a reported pair unreachable; baseline comparability tested against the uniform distribution randomisation implies, with between-variable correlation handled explicitly; and digit and dispersion checks that are declared inapplicable when their assumptions fail — each finding paired with a benign mechanism that would also produce it, because implausibility is not proof of misconduct.

Example conversation

Try asking like this

You

this trial baseline table looks too perfect and I need something defensible before excluding it

Trial Data Integrity Screener

Replaces a hunch about a suspiciously clean trial with computable tests on the numbers as printed: arithmetic consistency of percentages, denominators, dispersion labels and recomputed test statistics; granularity limits on means and deviations for integer-valued instruments; sample reconstruction that proves a reported pair unreachable; baseline comparability tested against the uniform distribution randomisation implies, with between-variable correlation handled explicitly; and digit and dispersion checks that are declared inapplicable when their assumptions fail — each finding paired with a benign mechanism that would also produce it, because implausibility is not proof of misconduct.

What I need first
  • The published article with its actual tables, or a faithful extraction that preserves the numbers exactly as printed along with table and page locations.
  • Per variable: count, bounded scale with declared minimum and maximum, integer instrument sum, or continuous measurement, plus the instrument's granularity. Without it the granularity tests are reported as not attempted.
  • Whether values were rounded or truncated and to how many decimals, per table if they differ. Several tests flip verdict between conventions.
  • Allocation ratio, stratification, block structure, cluster or individual assignment, and whether the table reports the randomised or the analysed set.
  • Other papers from the same research programme, which turns duplicate-block detection from impossible into a table with two citations per hit.Optional
  • Individual records where the team holds them, enabling duplicate-signature and profile-repetition checks that summary tables cannot support.Optional
What you get
  • Integrity Test Dossier
  • Author Clarification Query
  • Sensitivity Pooled Estimates
You can also ask
  • check whether these means are even possible for a 14 item scale with n of 38
  • the standard deviations across unrelated variables look almost identical, is that testable
  • two papers from this group report the same n mean and sd for different cohorts
Skills

What this agent is good at

  • Extract Reported Statistics
  • Check Arithmetic Consistency
  • Test Granularity Limits
  • Reconstruct Sample Distribution
  • Test Baseline Comparability
  • Audit Digit Distribution
  • Detect Duplicate Summary Blocks
  • Compile Integrity Dossier