How often is the Zush methodology updated?

The protocol is reviewed monthly and the page shows its latest review date. Product releases are logged in the public changelog. Benchmark results are published only when a reproducible run and underlying evidence are ready for review.

What weights does Zush use to score AI file renaming?

Semantic filename accuracy (35%), batch consistency (20%), automation depth (15%), safety controls (15%), and operational fit on Mac and Windows (15%). Weights are reviewed monthly and changes are published in the changelog.

Which file types are covered in the benchmark set?

The protocol covers screenshots, design files (Sketch, Figma, Adobe Illustrator, Photoshop), RAW and HEIC photos, PDFs, iWork, Office documents, audio, and video. A published result must use the same mixed-file corpus rather than a synthetic single-format sample.

How does Zush verify that a rename is safe and recoverable?

The protocol requires rollback, preview, and revert checks across a full batch run on both Mac and Windows builds. A workflow that lacks recovery behavior is downgraded regardless of naming quality.

Do you test cloud AI and Offline AI under the same protocol?

The protocol requires the same mixed-file corpus, prompts, templates, and scoring rules for cloud providers via BYOK and local models via Ollama. Results are not claimed until those controlled runs and their evidence are published.

AI File Renaming Methodology & Benchmarks

This page explains how Zush evaluates AI file renaming on Mac and Windows. We publish the scoring rubric, benchmark protocol, and review cadence so recommendations stay reproducible, auditable, and useful for real batch operations across mixed folders.

Reviewed monthly5 weighted checksUpdated 2026-07-10

Governance

Editorial Accountability

Primary author

Kirill Isachenko (lirik)

Technical reviewer

Kirill Isachenko, product owner

Update cadence

Criteria are reviewed monthly, with material changes logged in the changelog.

Rubric

Scoring Rubric

Each workflow run is graded against a weighted model that balances naming accuracy, operational safety, and day-to-day usability inside real Mac and Windows workflows.

Dimension	Weight	What we measure
Semantic filename accuracy	35%	Does the new name describe file meaning, not only generic patterns?
Consistency in batch runs	20%	Stable naming style across mixed files in one task.
Automation depth	15%	Batch workflow, folder monitoring, and repeatability.
Safety controls	15%	History, rollback, preview-before-apply behavior.
Operational fit	15%	Platform workflow quality on Mac and Windows, speed, and setup friction.

Validation

Benchmark Protocol

01
Use a mixed-file benchmark set covering screenshots, design files, RAW and HEIC photos, PDFs, iWork, Office documents, audio, and video — no synthetic single-format tests.
02
Run controlled rename jobs with identical prompts, templates, and Naming Blocks across cloud AI (Gemini, Groq, OpenAI, Claude) and Offline AI via Ollama.
03
Score outputs against the rubric and record false positives, ambiguous labels, and any collisions inside batch runs.
04
Validate rollback, preview, and one-click revert behavior on both Mac and Windows builds before marking a setup production-safe.
05
Publish results only after the run inputs, review date, tested versions, and supporting evidence can be reproduced.

Evidence status

No unpublished scores presented as facts

This page publishes the evaluation protocol, not a completed benchmark leaderboard. Zush does not claim benchmark scores until the underlying controlled run, tested versions, review date, and evidence can be published together.