ElenchosElenchos

Result auditing for independent researchers

Audit your results before you trust them.

Every result states what was checked, what was not, and why it matters.

See the checks

Works with Claude and Codex

Point your AI agent at a folder. It runs the audit.

Elenchos ships its own MCP server. Connect it once, then ask Claude or Codex to audit a backtest in plain language — it lays out the inputs, runs the nine checks locally, and reads you the verdict.

Connecting Elenchos to an AI agent via MCP inside an editor, shown with Claude Code, then asking it to audit a backtest

Shown with Claude Code

Inside the software

The desktop app, tab by tab.

Click the tabs. Every audit family states what it needs, what it found, and what it refuses to conclude without the missing input.

Elenchos

Pick up an audit, or load a new result set.

Load a result set
Previously audited analysesSee all →
model_a
18 Aug 2026 · dataset A · 2,894 obs
NOT VALIDATED
model_b
11 Aug 2026 · dataset B · 1,240 obs
NEEDS REVIEW
model_c
2 Aug 2026 · dataset C · 3,010 obs
VALIDATED
model_d
27 Jul 2026 · dataset D · 4,520 obs
NEEDS REVIEW
model_e
14 Jul 2026 · dataset E · 2,118 obs
NOT VALIDATED

Nine independent checks

Eight target the analysis. One targets the audit itself.

OVERFITTING

Combinatorial audit

Ranks every variant across recombined splits, so a winner cannot be the product of one lucky partition.

Check 01 of 09
OVERFITTING

Multiple testing

Judges the surviving variant against the full number of attempts that produced it.

Check 02 of 09
OVERFITTING

Sequential testing

Detects monitoring and stopping choices that quietly distort the statistics.

Check 03 of 09
DATA LEAKAGE

Data leakage

Finds future information used in a simulated past decision.

Check 04 of 09
DATA LEAKAGE

Researcher degrees of freedom

Tracks the small choices that can manufacture a result on one dataset.

Check 05 of 09
STATISTICAL RELIABILITY

Regime change

Tests whether the sample holds one data-generating regime, or several.

Check 06 of 09
STATISTICAL RELIABILITY

Audit self-check

Makes each detector prove it still catches the faults it claims to catch.

Check 07 of 09
ASSUMPTIONS & REPRODUCIBILITY

Declared vs applied parameters

Checks that the parameters you declared match the ones the pipeline actually used.

Check 08 of 09
ASSUMPTIONS & REPRODUCIBILITY

Implementation risk

Surfaces silent conventions that make two correct implementations disagree.

Check 09 of 09

How it runs

From raw run to a finding you can rerun.

Load the run
Preflight
Challenge
Inspect
Correct and rerun

Supply the run log, the output series and the assumptions the analysis used.

A named, traceable audit case.

Check schemas, dates, parameters and missing evidence before statistical tests begin.

Ready, blocked or incomplete. Never silently accepted.

Run the eight audit families. Apply the detector self-test gate to every positive result.

A register of passed, failed and unknown checks.

Open a finding with its source, evidence, impact, remaining unknown and required correction.

A finding that explains what to change and why.

Fix the input or configuration, then repeat the same audit against the revised run.

A before-and-after evidence trail.

Pricing

Unlimited audits on every plan. Pay monthly or once.

Starter

The two core overfitting checks plus the audit self-check, with unlimited audits.

Unlimited audits
  • Combinatorial audit
  • Multiple testing
  • Audit self-check
$20/mo

Pro

The whole overfitting family, regime-change detection and the audit self-check.

Unlimited audits
  • Combinatorial audit
  • Multiple testing
  • Sequential testing
  • Regime change
  • Audit self-check
$40/mo

Max

All nine checks across the four families, with unlimited audits.

Unlimited audits
  • Every check*
$70/mo

Lifetime

The same nine checks and unlimited audits, paid once and kept for good.

Unlimited audits
  • Every check*
$990

Open the sign-in and sign-up flow.

Legal notices
Contact