We measure your process.
We don’t judge your authorship.

Every write to the screenplay text is recorded in a dedicated ledger, with:

• Origin — who wrote it: you, the AI (with the model, when the path that wrote it knows it) or an import of your own file. What arrives through a version merge or a restore inherits the authorship of where it was born. Pasted text is identified: copied from Telos’s own AI text counts as AI; pasted from outside counts as unknown origin — never as yours
• The text — the full version at each point: every AI write and, of your edits, at most one checkpoint every 5 minutes per scene — the latest version is sealed before any AI write, on export and once a day
• Timestamp — the server time of each record
• AI decisions — AI suggestions you rejected, and passages you accepted in the editor
• Time investment — time between editing sessions, number of revisits per scene

From this data, we compute per scene: percentage reworked (lexical distance between the current text and the first AI-written version), dominant origin (original composition, refined from AI, reworked from AI, pure AI) and temporal patterns. Each record carries the hash of the previous one, in a chain per project; the system refuses to alter or delete records — the only exception is deleting the project or the account at your request — and any alteration made outside the system breaks the chain and shows up on verification. Once a day the chain is stamped at independent anchors — RFC 3161 time authorities (DigiCert; FreeTSA as a complement) and a Bitcoin block via OpenTimestamps —, which proves the record already existed on that date. We also show, line by line, which passages of today’s text came from the AI, from you or from outside (authorship map). What the ledger did not see is not asserted: text that predates it shows as “predates the record”, and writes without declared authorship show as such.

No biometric data. The lineage record does not capture:

• The individual content of keystrokes
• Typing cadence or neurological patterns
• Audio, video, webcam or microphone — the process record does not use them. Voice transcription in the Creation Room does send the audio you record; it is described in the Privacy Policy.
• Location or device fingerprinting beyond a standard IP

Everything this record knows comes from operations you already trigger naturally as you write. No invisible instrumentation.

Levenshtein distance (edit distance) between the scene’s current text and the first AI-written version, normalized by the larger of the two — approximated by lines for very long text, and the record says when it was. Open implementation in src/lib/utils/levenshtein.ts.

A heuristic threshold of 30% separates refined from AI from reworked from AI. It’s a number chosen through internal experimentation, not regulatory authority. It can be recalibrated as a public corpus of AI-assisted writing evolves.

When the system opens, you’ll be able to export a “Creative Process Record” at any time: a PDF/A plus the data to verify it without relying on Telos. It contains:

• The full record ledger chain, verified, with the daily anchors (RFC 3161 and Bitcoin), an independent verifier anyone can run with Node.js, and ready-made reports for U.S. Copyright Office registration, awards and editorial responsibility
• Per scene: records, decisions, time, % reworked, dominant origin and the current text’s composition by origin
• A conservative technical statement (not a legal one)
• Optionally, an RFC 3161 timestamp of the PDF itself, issued by DigiCert’s time authority (or FreeTSA’s, if DigiCert does not respond; the file says which) — technically attesting the PDF’s existence at that moment

The PDF does not declare authorship. It presents facts. Whoever judges still judges: lawyer, judge, copyright office, festival juror. Telos just hands you the notebook.

The Creative Process Record PDF does not constitute:

• A legal declaration of authorship under any legislation
• An electronic document equivalent to an ICP-Brasil signature (MP 2.200-2/2001) — FreeTSA is academic, not a regulatory certificate authority
• A certificate of copyright, literary material, or moral rights
• An AI detector — Telos does not try to “guess” whether others used AI; it only records what happened in here

For full ICP-Brasil legal force, take the generated PDF to a digital notary and re-notarize it with an ICP certificate. Telos serves as baseline auditable material.

Every AI-assisted writing tool will eventually face the same question: “how do you prove they’re the author?”. Those without a credible answer will be barred from festivals, refused at copyright offices, rejected by studios.

Telos doesn’t wait for the backlash. It builds the answer before the question dominates the industry. Not as a legal ceremony, but as the author’s daily tool: you see your process in real time, refine where you want, record evidence automatically.

Every scene you open carries its notebook built in. When they need to judge, you have a way to tell the story.

The full ADR is public in docs/adr/0016-provenance-processo-criativo.md in the repository. Key points:

• Granularity per scene, not per token. Token-level requires a complex ProseMirror plugin, marginal value.
• No emotional labels (L0–L4, etc.). Only a factual numeric score + a neutral descriptor. Avoids shaming AI use.
• No keystroke biometrics. LGPD Art. 11, marginal value.
• No ZK proofs in V1. Over-engineering, negligible audience.
• No ICP-Brasil in V1. Cost + complexity. FreeTSA covers baseline auditable material.