v0.1.0

First tagged release of stealthprint — a model-agnostic toolkit for fingerprinting anonymous / stealth LLMs behind any OpenAI-compatible (or Anthropic /messages) endpoint.

What's in 0.1.0

Fingerprint layers (CLI + library)

• tokenizer (L1) — token-count differential against local candidate tokenizers (HF tokenizer.json, o200k_base, cl100k_base), with source/sha256 provenance
• wrapper (L2) — chat-template overhead constant
• context (L3) — verified context bound + needle retrieval
• errors (L4) — error-envelope family of the serving stack
• vision / video (L6) — vision ground truth, repeated-probe backend-mix check, video content-block shapes
• catalog (L7) — same-gateway A/B against named catalog siblings
• survey — resume-safe catalog census with reachability classification and /responses fallback
• --record for timestamped raw request/response logs (auth headers stripped), --json output, en / zh UI
• --transport anthropic for /messages endpoints; opencode-free shell for Zen free-tier checks

Evidence archive — case studies (omen-alpha, pixel-canary, space-bunny, union-alpha, cat-jam, deepseek-v4.1-flash) and a static investigation site (scripts/build_site.py).

Fixes since the last untagged main

• Wheel now ships data/opencode_free_system.txt, so installed wheels import and the CLI starts (#24)
• Probe files reject empty / non-string probe names (#22)
• A malformed local tokenizer candidate is skipped with a warning instead of aborting the whole comparison (#23)
• Earlier: explicit endpoint key required (#14), vision control-failure stop (#15), verified context bound (#16), wrapper constant-payload deltas (#17), survey Responses retry (#18), measured tools baseline (#19), measurement integrity + traceable evidence (#21)

Install

pip install "stealthprint[all] @ git+https://github.com/majiayu000/stealthprint.git@v0.1.0"

or grab the wheel attached below.

Results are measurement evidence, not proof of weights or vendor — see the README's evidence limits.