# Reproducible corpus analysis: v2026.10

**Release date:** 2026-10-02  
**Records:** 190

This report is generated by `scripts/analyze-corpus.mjs` from the frozen public snapshot. It describes the Atlas corpus and must not be interpreted as prevalence in the global cyber-deception field.

## Resource types

| Category | Count | Share of corpus |
| --- | ---: | ---: |
| software | 99 | 52.1% |
| paper | 46 | 24.2% |
| product | 13 | 6.8% |
| service | 11 | 5.8% |
| case-study | 9 | 4.7% |
| dataset | 8 | 4.2% |
| community | 2 | 1.1% |
| framework | 2 | 1.1% |

## Review basis

| Category | Count | Share of corpus |
| --- | ---: | ---: |
| repository-metadata | 99 | 52.1% |
| source-page | 45 | 23.7% |
| full-text | 24 | 12.6% |
| publisher-metadata | 21 | 11.1% |
| abstract | 1 | 0.5% |

Review basis identifies the material examined by the editor. It is not a quality or effectiveness score.

## Sensitivity views

| View | Records |
| --- | ---: |
| all published | 190 |
| with critical reading | 50 |
| full text only | 24 |
| source or full text | 69 |
| excluding publisher metadata | 169 |
| excluding repository and publisher metadata | 70 |
| excluding repository metadata | 91 |
| papers with full text | 24 |
| non github sources | 89 |

Removing repository-metadata records reduces the visible corpus from 190 to 91. Restricting it to full-text reviews reduces it to 24. This demonstrates that corpus volume and evidentiary depth answer different questions.

## Missingness

| Field | Missing | Applicable | Missing share |
| --- | ---: | ---: | ---: |
| year | 43 | 190 | 22.6% |
| doi papers | 10 | 46 | 21.7% |
| source language | 124 | 190 | 65.3% |
| organization country products services | 22 | 24 | 91.7% |

## Source concentration

The snapshot contains 101 GitHub source URLs and 89 sources hosted elsewhere. Host concentration is a property of the curation process and source availability, not evidence that software dominates real-world use.

## Interpretation boundary

- Technique and environment tags are multi-valued.
- The baseline corpus lacks a complete historical candidate/exclusion ledger.
- Public sources systematically omit procurement-confidential and negative operational evidence.
- Vendor-authored sources remain attributed and are not treated as independent evaluations.
- Sensitivity subsets are transparency views, not formal evidence-quality grades.