Diagnostic instrument
Death Certificate
Statistical signature analysis for a single document. Calibrated. Versioned. Cryptographically signed.
What this tool does
Submit a single English-language document. Death Certificate returns:
- IMS_text score from 0 to 100
- 95% confidence interval (bootstrap-derived)
- Verbal label: Live, Hybrid, Synthetic, or Dead
- Methodology version with hash for reproducibility
- Cryptographic signature (Ed25519) for verification
- Permanent URL for sharing
The tool reflects statistical signature, not authorship proof. It is a diagnostic instrument, not a verdict.
Run Death Certificate
Paste your document below.
Minimum 200 words. Maximum 5,000 words. English only.
30–60 seconds · Permanent URL · Cryptographically signed
What this tool cannot do
It cannot prove authorship. It can suggest a statistical pattern consistent or inconsistent with AI generation, with calibrated confidence. Proof of authorship requires multiple independent forms of evidence.
It cannot detect AI text under recursive paraphrasing. This is a fundamental limit, not a flaw of any specific tool. See the methodology page §10 (Sadasivan boundary) for the formal result and its implications.
It cannot detect AI text in mixed human-AI documents reliably. Hybrid authorship produces ambiguous signatures. The verbal label Hybrid covers this case explicitly.
It cannot evaluate truth. Whether a document's claims are factually accurate is a separate question requiring fact-checking, source verification, and corroboration.
It cannot replace editorial judgment. Use Death Certificate as a diagnostic signal in your investigative workflow, not as a substitute for journalistic verification.
How this works
A four-block analysis pipeline:
- 1.Stylometric vector covering function word usage, sentence rhythm, POS distribution, punctuation patterns, lexical diversity, and character n-gram frequencies.
- 2.Compression-based novelty detection through normalized compression distance against a reference reservoir of known-human and known-AI reference texts.
- 3.Internal uncertainty estimation through perplexity distribution when processed by a small reference language model.
- 4.Discourse cohesion measurement through inter-sentence semantic relationships.
These four signals combine through a calibrated generalized linear model with isotonic regression for probability calibration. The output is mapped to the IMS_text score from 0 to 100.
Bootstrap confidence intervals are computed from text fragment resampling at 200 iterations. Wider intervals indicate shorter or more uncertain documents.
The full methodology specification is published at deadmeter.com/methodology.
Frequently asked questions
How long does analysis take?
30 to 60 seconds for a typical document. Longer for documents near the maximum length, or when the model is busy with other requests.
Why does my Cert have a wide confidence interval?
Wider intervals occur for shorter documents, documents in underrepresented domains in our calibration corpus, or documents showing patterns our model is uncertain about. The interval is honest about uncertainty.
Can I run the same document twice and get different results?
The same input plus same methodology version produces the same output. Cryptographic hash verifies this. If you submit twice within the same methodology version, you get the same Cert with the same hash. After version updates, results may differ. Historical Certs remain valid through their original version.
Is my submitted text stored?
Yes, for 30 days after submission to support permanent URL rendering. After 30 days the original text is deleted; the Cert metadata (score, verdict, hash) remains. Do not submit confidential content. The privacy policy details retention policies.
Can I delete my Cert?
Contact us with the Cert URL and reason. We honor deletion requests within 14 days for Cert authors who can verify control of the original submission.
Why are some texts rejected?
Death Certificate rejects: text under 200 words, text over 5,000 words, non-English text, code (use Code DNA instead), text with extreme repetition, encoded or obfuscated content. Each rejection explains the specific issue.