Skip to content

Forensic toolkit

Forensic authorship research needs more than attribution — it needs one-class verification, topic-invariant features, and evidential output framed as a likelihood ratio, calibrated against a reference population.

tamga.forensic ships these as a cohesive layer on top of the analysis methods. Every tool is classifier-agnostic — pair it with any tamga feature set and any Delta / Zeta / classify method.

What's in the package

from tamga.forensic import (
    # Verification
    GeneralImpostors,            # Koppel & Winter 2014
    Unmasking,                   # Koppel & Schler 2004

    # Topic-invariant features
    CategorizedCharNgramExtractor,   # Sapkota et al. 2015
    distort_corpus, distort_text,    # Stamatatos 2013

    # Calibration + LR output
    CalibratedScorer,                # Platt / isotonic
    log_lr_from_probs,
    log_lr_from_probs_with_priors,

    # Evaluation metrics (PAN menu)
    auc, c_at_1, f05u, brier, ece, cllr, tippett,
    PANReport, compute_pan_report,
)

from tamga.report import build_forensic_report   # LR-framed HTML template

The forensic workflow

flowchart LR
  Q["Questioned<br>document Q"] --> V
  K["Known docs K<br>(candidate author)"] --> V
  I["Impostor pool I"] --> V
  V["Verification<br>(GI or Unmasking)"] --> S["Raw score"]
  S --> CAL["CalibratedScorer"]
  CAL --> LR["log<sub>10</sub>-LR"]
  LR --> PAN["PANReport<br>auc, c@1, F0.5u,<br>Brier, ECE, C_llr"]
  LR --> REP["Forensic HTML report<br>(LR + custody block)"]

  style V fill:#FBF3DE,stroke:#C9A34A
  style CAL fill:#FBF3DE,stroke:#C9A34A
  style LR fill:#FBF3DE,stroke:#C9A34A
  style PAN fill:#FBF3DE,stroke:#C9A34A
  style REP fill:#FBF3DE,stroke:#C9A34A

Why LR framing matters

Forensic journals (IJSLL, Language and Law) and courtroom gatekeeping (Daubert in the US, R v T and its successors in the UK) expect evidence framed as a likelihood ratio:

$$ \text{LR} = \frac{P(E \mid H_1)}{P(E \mid H_0)} $$

— the probability of the evidence under the same-author hypothesis divided by its probability under different-author, with a calibrated underlying scorer. Raw classifier posteriors are rarely calibrated and are easily abused as "probability of guilt" in ways that misrepresent forensic semantics.

tamga's CalibratedScorer + log_lr_from_probs pipeline converts any raw score to a log-LR defensible in a forensic report, and the bundled ENFSI / Nordgaard verbal scale in the report template gives a transparent verbal reading.

Chain of custody

Every Provenance record accepts six optional forensic fields:

  • questioned_description — free-text description of Q
  • known_description — free-text description of K
  • hypothesis_pair — e.g. "H1: authored by X; H0: authored by someone other than X"
  • acquisition_notes — how source material was obtained (warrant, image, etc.)
  • custody_notes — any further chain-of-custody annotations
  • source_hashes — dict[document_id, SHA-256] of the original files

These land in the rendered forensic report under a dedicated Chain of custody block.