---
title: "Undominated · Elo drift audit: review, setup & limitations · Undominated.ai"
canonical: https://undominated.ai/skills/undominated-elo-drift-audit/
description: "Audit Bradley-Terry win-rate confidence interval bounds, active pair sampling bias, and Elo deflation/inflation drift across benchmark dataset versions."
---

# Undominated · Elo drift audit: review, setup & limitations · Undominated.ai

> Audit Bradley-Terry win-rate confidence interval bounds, active pair sampling bias, and Elo deflation/inflation drift across benchmark dataset versions.

[← Explore all skills](/skills/)

EVIDENCE AND EVALUATION / Undominated.ai

# Undominated · Elo drift audit

Audit Bradley-Terry win-rate confidence interval bounds, active pair sampling bias, and Elo deflation/inflation drift across benchmark dataset versions.

 [See setup guidance ↓](#setup)[Original source ↗](https://undominated.ai/resources/skills/undominated-elo-drift-audit/definition.md)

SOURCE REVIEW

 Reviewed 2026-10-14
 Evidence 3 linked sources
 Publisher Undominated.ai
 Licence [MIT ↗](https://undominated.ai/resources/skills/undominated-elo-drift-audit/LICENSE.txt)
 Revision 4900cbb36b81
 [Read what was—and wasn’t—checked ↓](#review)

“Audit Bradley-Terry win-rate confidence interval bounds, active pair sampling bias, and Elo deflation/inflation drift across benchmark dataset versions.”

 [Undominated.ai · upstream description ↗](https://undominated.ai/resources/skills/undominated-elo-drift-audit/definition.md) Our analysis follows below.

01 / THE REASONING

## Why this made the selection.

 - Pairs a bounded workflow with an offline Python checker and explicitly synthetic example inputs.
- Created for recurring evidence and comparison failures: missing facts remain unknown and every conclusion keeps its scope.
- It names a concrete failure mode: Flag drift claims that are merely sampling noise ($|\Delta \text{Elo}| \le \text{CI}_{\text{combined}}$).

### A good fit for

 - Distinguish genuine model capability improvements from statistical sampling noise and rating inflation on arena leaderboards.
- Audit Bradley-Terry win-rate confidence interval bounds, active pair sampling bias, and Elo deflation/inflation drift across benchmark dataset versions.

### Weigh up before choosing

 - Computes statistical significance over supplied benchmark rating snapshots; does not compute Elo from raw battle votes.
- Fixture checks establish the documented local behaviour, not performance on an unrestricted set of real tasks.
- Public GitHub source availability is not yet verified; the complete download contains the reviewed source.
- A new npm install path is not yet verified. Use the complete source download.
- `0`: Valid observations and drift analysis completed within bounds.

02 / THE REVIEW RECORD

## What we actually inspected.

Source review has boundaries. A clear record is more useful than a “safe” badge.

### Material inspected

 - skills/undominated-elo-drift-audit/SKILL.md
- skills/undominated-elo-drift-audit/scripts/check.py
- skills/undominated-elo-drift-audit/examples/synthetic.json

### Our findings

 - First-party resource authored by Undominated.ai; this is our own verification record, not an independent endorsement.
- The download preserves the complete source and supporting files. Its manifest hashes identify the exact bytes.
- One input rule the file states is: `maxAverageAnchorDrift`: optional decimal string (e.g. "15.0").
- One condition that blocks a pass is: Average anchor drift ({avg_anchor_drift}) exceeds maxAverageAnchorDrift ({max_drift}).

### Not established by this review

 - Model compliance with these instructions on arbitrary tasks
- Native integration in every agent host
- Public GitHub source identity
- Remote installation through the standard Skills CLI
- The new npm resource CLI installation

The review applies to the material and revision named here. A newer upstream release can change its behavior.

03 / PUT IT TO WORK

## Add a skill to your workflow.

[Upstream setup instructions ↗](https://undominated.ai/resources/skills/undominated-elo-drift-audit/definition.md)
 - Download the complete bundle from https://undominated.ai/resources/skills/undominated-elo-drift-audit/bundle.zip and extract it into a new directory.
- Place the extracted directory at .agents/skills/undominated-elo-drift-audit/ or your host's documented skills directory. Preserve SKILL.md, scripts, examples and licence files together.
- Run the documented synthetic example from the skill directory with Python 3.10 or later before using your own evidence.

### Before you start

 - An Agent Skills compatible host
- Python 3.10 or later for the optional checker

### Compatibility

Agent Skills compatible hosts · Python 3.10+

### Implementation

Python · Markdown

### Local, explicit inputs

 - read:user-selected-local-file

### Cost model

MIT source at no charge. Your agent host or model provider may charge for use; the included offline checks require no paid API.

THE COMPLETE REVIEWED DEFINITION

## Read it before you reuse it.

Original source bytes, with attribution. Review the host-specific setup notes above.

 Copy definition ↗ [Download complete resource + licence ↗](/resources/skills/undominated-elo-drift-audit/bundle.zip)[Raw Markdown ↗](/resources/skills/undominated-elo-drift-audit/definition.md)
 # Elo Drift Audit

Audit Bradley-Terry win-rate confidence interval bounds, active pair sampling bias, and Elo deflation/inflation drift across benchmark dataset versions.

## Workflow

1. Ingest model score observations across two benchmark releases ($T_1$ and $T_2$):
 - Model identifier
 - Elo rating at $T_1$ and $T_2$
 - 95% Confidence Interval half-widths ($\pm \text{CI}$)
 - Total battle count ($N$)
2. Compute Elo delta:
 $$\Delta \text{Elo} = \text{Elo}_{T_2} - \text{Elo}_{T_1}$$
3. Evaluate statistical significance:
 - Check if $\Delta \text{Elo}$ exceeds the combined 95% CI error margin:
 $$\text{CI}_{\text{combined}} = \sqrt{\text{CI}_1^2 + \text{CI}_2^2}$$
 - Flag drift claims that are merely sampling noise ($|\Delta \text{Elo}| \le \text{CI}_{\text{combined}}$).
4. Detect deflation/inflation: Calculate the average rating movement across all benchmark anchor models.

## Synthetic Input Format

JSON object with:
- `schemaVersion`: 1
- `benchmark`: string
- `anchorModels`: list of objects, each containing:
 - `model`: string
 - `eloT1`: decimal or integer
 - `eloT2`: decimal or integer
 - `ciHalfWidthT1`: decimal
 - `ciHalfWidthT2`: decimal
 - `battles`: positive integer
- `assertions`:
 - `maxAverageAnchorDrift`: optional decimal string (e.g. "15.0")

## Exit Codes

- `0`: Valid observations and drift analysis completed within bounds.
- `1`: Assertion failed (e.g. anchor drift exceeded threshold).
- `2`: Malformed input or safety violation.

## Verification

```sh
python3 scripts/check.py examples/synthetic.json
```

The download contains SKILL.md . Extract the whole bundle; the supporting files are required. Inspect the included MANIFEST.json for file hashes.

By **Undominated.ai**. [Exact upstream source ↗](https://undominated.ai/resources/skills/undominated-elo-drift-audit/definition.md) · [Licence](/resources/skills/undominated-elo-drift-audit/LICENSE.txt) · [Attribution](/resources/skills/undominated-elo-drift-audit/ATTRIBUTION.txt)

SHA-256 4900cbb36b8163e1dc4afb28de44cd970920af4ef537dbe396a12245da1f6861

 Read the applicable licence MIT License

Copyright (c) 2026 Undominated.ai

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.

04 / FOLLOW THE EVIDENCE

## The source trail.

Our notes are separate from the original resource. Check upstream before adopting a new version.

 - [Original Undominated definition ↗](https://undominated.ai/resources/skills/undominated-elo-drift-audit/definition.md) Checked 2026-10-07 https://undominated.ai/resources/skills/undominated-elo-drift-audit/definition.md Supports: Workflow and input requirements, Stated limitations
- [MIT licence ↗](https://undominated.ai/resources/skills/undominated-elo-drift-audit/LICENSE.txt) Checked 2026-10-07 https://undominated.ai/resources/skills/undominated-elo-drift-audit/LICENSE.txt Supports: Redistribution terms and attribution
- [Complete source bundle ↗](https://undominated.ai/resources/skills/undominated-elo-drift-audit/bundle.zip) Checked 2026-10-07 https://undominated.ai/resources/skills/undominated-elo-drift-audit/bundle.zip Supports: Full local source, supporting files and integrity manifest

KEEP COMPARING

## Other approaches to consider.

Related by category or shared topics. These are alternatives to inspect, not a measured quality order.

 [### Undominated · Benchmark cohort audit ↗ Audit a benchmark comparison for matched model identity, missing coverage, restricted populations and misleading confidence-interval interpretations.](/skills/undominated-benchmark-audit/)[### Undominated · Concurrency cliff audit ↗ Audit inference throughput degradation, queue build-up, and throughput cliff saturation as concurrency scales.](/skills/undominated-concurrency-cliff-audit/)[### Undominated · Eval contamination and judge audit ↗ Audit benchmark evaluation sets and prompts for training set contamination, n-gram overlap, verbatim memorization, and judge prompt skew.](/skills/undominated-eval-contamination-audit/)

 [AI Tools ↗](/tools/)[Skills ↗](/skills/)[Agents ↗](/agents/)[MCP Servers ↗](/mcp-servers/)[Workflows ↗](/workflows/)

## Continue your investigation

 - [Choose a toolkit](/tools/)
- [Inspect execution hosts](/agents/)
- [Check tool access](/mcp-servers/)
- [Check project dependencies](/check-repo/)
