Linda Trace - model pack and signed update list

This repository is the update source of Linda Trace, a local text investigator for Windows (compare the writing style of documents, check questioned texts against samples of known authors; English, Russian, Polish). The application itself is distributed from GitHub Releases.

Contents

Path What it is
models/char_ngrams_model.json the character 3-4-gram TF-IDF resource used by the app: 2,000 features and IDF weights per language (en, pl, ru), about 152 KB
latest.json, latest.json.sig the signed update list: model pack version and file hashes, plus the installer of the newest app version. The signature is Ed25519 over the exact bytes of latest.json; the public key is built into the app (1uPfXe1IO9egH/t+j0hAqzneMdtT8tmqaLSRcFnIe0E=)
provenance/metadata.json, provenance/TRAIN_ATTRIBUTION.json how the resource was built and per-work attribution of its training data

Model

Not a neural network: a vocabulary of character 3-4-grams (analyzer=char, ngram_range=(3,4), lowercase, sublinear_tf, smooth_idf, L2 norm, 2,000 features per language) with IDF weights, fitted on 300-word windows of documented literary prose. The app uses the vectors, together with about 250 stylometric features and a directional cosine-delta score, to rank documents by relative style distance.

Training data: 407 works (147 English, 58 Polish, 202 Russian) that are public domain originals and/or CC0 contributions; per-work source, license basis and attribution are listed in provenance/TRAIN_ATTRIBUTION.json. No third-party text is stored in this repository.

Limits (read before relying on it)

  • Results are relative style distances inside the corpus a user supplies. They are not authorship probabilities and prove nothing about identity, authenticity or leaks.
  • Historical literary prose only. An independent held-out test (92 works, authors who died 1852-1925) gave top-1 9/17 (en), 9/19 (pl), 14/24 (ru) for the shipped fused ranking; modern institutional prose (US Supreme Court opinions 2024-2025) gave 4/10. Modern and corporate texts are not covered by any claim.
  • Character n-grams can reflect topic or source as well as the author. Documentary rights fields do not establish worldwide legal clearance: 147 works rely on a declared US public-domain status.

Updates

The app checks latest.json / latest.json.sig only after you press Check for updates or, if enabled, once at start. It verifies the signature before reading the list, downloads only changed files, checks SHA-256, and applies a model pack atomically (previous pack restored on failure). Details: https://github.com/Lendarixon/Linda-Trace/blob/main/UPDATES.md

License: proprietary, see the link above. Contact: GitHub Issues of the Linda Trace repository.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support