--- title: "Undominated · Needle-in-a-haystack density audit: review, setup & limitations · Undominated.ai" canonical: https://undominated.ai/skills/undominated-needle-in-haystack-density-audit/ description: "Audit long-context retrieval recall across varying context lengths, needle depth placements, and distractor similarity levels." --- # Undominated · Needle-in-a-haystack density audit: review, setup & limitations · Undominated.ai > Audit long-context retrieval recall across varying context lengths, needle depth placements, and distractor similarity levels. [← Explore all skills](/skills/) EVIDENCE AND EVALUATION / Undominated.ai # Undominated · Needle-in-a-haystack density audit Audit long-context retrieval recall across varying context lengths, needle depth placements, and distractor similarity levels. [See setup guidance ↓](#setup)[Original source ↗](https://undominated.ai/resources/skills/undominated-needle-in-haystack-density-audit/definition.md) SOURCE REVIEW Reviewed 2026-10-14 Evidence 3 linked sources Publisher Undominated.ai Licence [MIT ↗](https://undominated.ai/resources/skills/undominated-needle-in-haystack-density-audit/LICENSE.txt) Revision fa4fe232507f [Read what was—and wasn’t—checked ↓](#review) “Audit long-context retrieval recall across varying context lengths, needle depth placements, and distractor similarity levels.” [Undominated.ai · upstream description ↗](https://undominated.ai/resources/skills/undominated-needle-in-haystack-density-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: Detect severe accuracy dips when needles are placed between depth 30% and 70%. ### A good fit for - Verify long-context recall across depth percentiles and detect lost-in-the-middle degradation anomalies. - Audit long-context retrieval recall across varying context lengths (32k. ### Weigh up before choosing - Analyzes supplied retrieval test results; does not execute long-context inference runs. - 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 matrix and all recall thresholds met. 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-needle-in-haystack-density-audit/SKILL.md - skills/undominated-needle-in-haystack-density-audit/scripts/check.py - skills/undominated-needle-in-haystack-density-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: `minMiddleDepthRecall`: optional decimal string (for depths between 30% and 70%). - One condition that blocks a pass is: tests[{idx}].recalled must be a boolean. ### 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-needle-in-haystack-density-audit/definition.md) - Download the complete bundle from https://undominated.ai/resources/skills/undominated-needle-in-haystack-density-audit/bundle.zip and extract it into a new directory. - Place the extracted directory at .agents/skills/undominated-needle-in-haystack-density-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-needle-in-haystack-density-audit/bundle.zip)[Raw Markdown ↗](/resources/skills/undominated-needle-in-haystack-density-audit/definition.md) # Needle-in-a-Haystack Density Audit Audit long-context retrieval recall across varying context lengths (32k, 128k, 256k, 1M, 2M tokens), needle depth placements (0% to 100%), and distractor similarity levels. ## Workflow 1. Ingest retrieval test matrix: - Total context length in tokens ($L$) - Needle depth placement percentile ($D \in [0.0, 1.0]$) - Retrieved fact exact match / semantic accuracy ($A \in [0.0, 1.0]$) - Distractor domain (e.g. random text vs high semantic overlap) 2. Aggregate retrieval accuracy by context tier: - Near context ($\le 32\text{k}$) - Mid context ($32\text{k} - 128\text{k}$) - Deep context ($> 128\text{k}$) 3. Identify "Lost in the Middle" anomalies: - Detect severe accuracy dips when needles are placed between depth 30% and 70%. 4. Verify long-context recall meets application threshold (e.g. $\ge 95\%$). ## Synthetic Input Format JSON object with: - `schemaVersion`: 1 - `model`: string - `provider`: string - `tests`: list of objects, each containing: - `contextLength`: positive integer - `depthPercent`: non-negative integer or decimal (0 to 100) - `recalled`: boolean - `assertions`: - `minOverallRecall`: optional decimal string (e.g. "0.9000") - `minMiddleDepthRecall`: optional decimal string (for depths between 30% and 70%) ## Exit Codes - `0`: Valid matrix and all recall thresholds met. - `1`: Retrieval recall failure. - `2`: Invalid input format 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-needle-in-haystack-density-audit/definition.md) · [Licence](/resources/skills/undominated-needle-in-haystack-density-audit/LICENSE.txt) · [Attribution](/resources/skills/undominated-needle-in-haystack-density-audit/ATTRIBUTION.txt) SHA-256 fa4fe232507ff3cecbcf1be924c2da4725dbc3ec03f4b8eacea1d6f58aa18a4a 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-needle-in-haystack-density-audit/definition.md) Checked 2026-10-07 https://undominated.ai/resources/skills/undominated-needle-in-haystack-density-audit/definition.md Supports: Workflow and input requirements, Stated limitations - [MIT licence ↗](https://undominated.ai/resources/skills/undominated-needle-in-haystack-density-audit/LICENSE.txt) Checked 2026-10-07 https://undominated.ai/resources/skills/undominated-needle-in-haystack-density-audit/LICENSE.txt Supports: Redistribution terms and attribution - [Complete source bundle ↗](https://undominated.ai/resources/skills/undominated-needle-in-haystack-density-audit/bundle.zip) Checked 2026-10-07 https://undominated.ai/resources/skills/undominated-needle-in-haystack-density-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 · Long-Context Distraction Audit ↗ Audits retrieval accuracy degradation as irrelevant distractor documents are added to prompts (evaluating signal-to-noise ratio in long contexts).](/skills/undominated-context-distraction-audit/)[### Undominated · A prompt is not the catalogue ↗ Refuse a prompt whose scope is the whole catalogue. Retrieval stays top-k. No catalogue is copied.](/skills/undominated-topk-not-dump/)[### Undominated · KV Cache INT4 Outlier Drift Audit ↗ Audits key/value cache INT4 quantization channel scaling to prevent catastrophic attention drift on long sequences.](/skills/undominated-kv-cache-int4-outlier-drift-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/)