meow LLM Detector 4.5.4

128 requests recommended for the new GPT baseline. The 32- and 64-request tiers distinguish 6.1 Sol from Astra less reliably and may return insufficient evidence more often. Tiers and validation

Compare model behavior using distributions of answers to short probes. Fingerprint evidence is not identity authentication and does not establish why a provider's behavior differs.

Download and launch

• Windows 10/11 Intel/AMD x64: use the windows-x64-portable-en.zip, extract into a new folder and run start.bat. Python and dependencies are isolated and bundled.
• Source packages require Python 3.11+ on Windows, macOS or Linux. Run start.bat on Windows or sh start.sh on macOS/Linux. Dependencies are installed in a private environment, not globally.
• The default address is http://127.0.0.1:8765/ . Pass --port 8790 for a different port.
• The website requires no installation; reports are public and upstream APIs must use public HTTPS.

Use GitHub Releases for official downloads; temporary acceptance-branch builds are not official release assets.

First detection

Choose a protocol, enter URL/Key or select a saved connection, choose a baseline and claimed model, then review the request budget and start. Fetch the site's model list to fill the request alias, or enter it manually. API charges belong to your account. Search history by URL, claimed model or request alias.

Claude aliases default to relay-style hyphens such as claude-fable-5-1; OpenRouter retains its precise alias. Model-list entries are not identity evidence.

Comparing old and new scores

Updated scoring accumulates valid-answer evidence and displays its average advantage, reducing sensitivity to small samples. Thresholds are recalibrated and are generally lower. A lower displayed score does not mean weaker identification. Use the current threshold; do not directly compare percentages across versions.

For a simplified scale example, ignore differences in fitting and assume two candidates and six independent, equally weighted evidence terms, each with average advantage δ. An old-style summed scale would be approximately sigmoid(6δ); the new average scale approximately sigmoid(δ).

Average advantage δ	Illustrative old scale	Illustrative new scale
0.4	91.68%	59.87%
0.8	99.18%	69.00%
1.2	99.93%	76.85%
2.3	about 99.9999%	90.89%

This illustrates display scale only: it is not a conversion formula for real reports, an accuracy rate or an identity probability. Actual fitting, request counts, candidate/family structure and thresholds also differ. See the technical report.

Updates and migration

Updates are checked at startup and every 24 hours while running. One click prepares and verifies a new version, waits for current tasks and restarts. Failed startup rolls back. Existing reports and frozen tasks retain their original baseline.

Version 4.5.2 has no automatic installer, so entering 4.5.3 requires launching the new package once. Stop the old backend and use start.bat --migrate-from "old installation" or sh start.sh --migrate-from "/old installation". The new data directory must be empty. Original data is kept; OS-vault credentials do not migrate between computers. Modified developer checkouts are not overwritten: use a separate official installation.

Privacy and interpretation

Temporary keys stay in page/task memory, not ordinary files. Refreshing the page or restarting the backend loses them. Explicitly saved credentials use the OS vault. Reports and non-secret settings live in meow_runs.

Local HTTP is for trusted local networks only. On the website, requests including keys pass through Cloudflare and the server; URLs, groups and sanitized errors are public.

A mismatch does not establish deliberate model substitution. Routing, service changes or baseline limitations may contribute; this tool alone cannot determine the cause. Community discussion is context, not verified attribution.

Usage · Technical scope · Baseline workflow · Changes