How it works

Methodology — measured, not predicted

Gramavec's forensic engine is a calibrated computer-vision stack. Every claim it surfaces is backed by a reproducible measurement, a benchmark number, and visual evidence a reviewer can audit.

Calibrated measurement engine

Eight families of deterministic features are extracted from each ink trace. No external model, no network call — same input, same output, every time.

Stroke width — Medial-axis estimate of pen-stroke thickness and its variation.
Slant angle — Vertical projection profile across rotation angles to estimate writing slant.
Skeleton & topology — Endpoints, branch points, and length of the inked skeleton — a stable shape signature.
Geometric descriptors — Scale-invariant aspect ratio, ink extent, and solidity of the bounding region.
Loop count — Number of closed loops in the binary mask — discriminates writing styles.
Hu moments (log) — Seven invariants that fingerprint shape independent of translation, rotation, and scale.
Ink density (proxy) — Optical-density proxy over ink pixels — clearly labelled NOT a true pen-pressure measurement.
Texture (LBP + GLCM) — Local-binary-pattern and grey-level co-occurrence properties — texture beyond shape.

Calibrated on a public forensic benchmark

Thresholds are derived from the public CEDAR signature dataset. The numbers below are what the engine reports — no marketing inflation, no hidden tuning.

0.72

Forensic AUC

0.37

Equal error rate

0.77

FRR at FAR 1%

60+40

Genuine + impostor

CEDAR is the public Center of Excellence for Document Analysis benchmark; 60 genuine + 40 impostor (n=100) is a small sample. We report all four metrics including the worst-case (FRR at FAR=1% = 0.77) so reviewers can audit the gap between best- and worst-case performance. Source: gramavec/forensics/calibration_cedar.json. Pinned by automated tests so a calibration regression cannot ship without a code change.

The VLM is a scribe, not the engine

Where we use Google's Gemini in the analysis flow, it acts strictly as a language-and-vision scribe: it transcribes context and explains a measurement, never produces the underlying numeric score. The score is computed by the deterministic CV engine above. That separation is structural — it's what lets us say 'measured, not predicted' and mean it.

Scope — EU AI Act Annex III alignment

Gramavec's lane separation is not decoration; it's a compliance posture. We treat the discovery lane and the forensic lane under different regulatory classifications, and our copy, tier gating, and disclaimers reflect that explicitly.

Out of scope (Annex III high-risk uses) — Employment, eligibility, creditworthiness, education access, and similar decision-making. The Discovery modules (HR style, learning style, wellness, partner compatibility) are presented as reflective entertainment only — never as inputs to a hiring, admissions, lending, or clinical decision. Tier-gated to Free precisely so they cannot be sold as a paid HR/medical screening tool.
In scope (forensic pre-screening with human-in-the-loop) — Signature comparison, document forgery, document semantic checks, document triage. These produce a measured, calibrated reading with a clearly-labelled confidence band. A certified Forensic Document Examiner (FDE) keeps the final say — we surface evidence, the expert renders judgment.

If you are evaluating Gramavec for a regulated procurement, ask us for the assurance pack: model card, sample-size disclosure, retention configuration per tier, and the audit log schema.

Visual evidence — reviewer-replicable

Every forensic comparison renders the same four artefacts. A reviewer (or a certified Forensic Document Examiner) can regenerate them locally from the original images and audit each claim.

Signature SSIM heatmap showing per-pixel structural similarity between a reference and questioned signature

• Side-by-side crops of the reference and questioned ink, at common height.
• Aligned absolute-difference overlay drawn in red over the reference.
• Per-pixel SSIM heatmap (bright = more different).
• Skeleton overlay with endpoints (red) and branch points (blue).

Replication & cross-checks

Every claim on this page is grounded in code or a public benchmark. The engine is in gramavec/forensics/, the CEDAR numbers are in calibration_cedar.json, and the visualisation artefacts come from visualize.py. For security, retention, and data-handling, see the surfaces below.

Security & data handling Privacy policy Try signature verification

Calibrated measurement engine

Eight families of deterministic features are extracted from each ink trace. No external model, no network call — same input, same output, every time.

Stroke width — Medial-axis estimate of pen-stroke thickness and its variation.

Slant angle — Vertical projection profile across rotation angles to estimate writing slant.

Skeleton & topology — Endpoints, branch points, and length of the inked skeleton — a stable shape signature.

Geometric descriptors — Scale-invariant aspect ratio, ink extent, and solidity of the bounding region.

Loop count — Number of closed loops in the binary mask — discriminates writing styles.

Hu moments (log) — Seven invariants that fingerprint shape independent of translation, rotation, and scale.

Ink density (proxy) — Optical-density proxy over ink pixels — clearly labelled NOT a true pen-pressure measurement.

Texture (LBP + GLCM) — Local-binary-pattern and grey-level co-occurrence properties — texture beyond shape.

Calibrated on a public forensic benchmark

Thresholds are derived from the public CEDAR signature dataset. The numbers below are what the engine reports — no marketing inflation, no hidden tuning.

0.72

Forensic AUC

0.37

Equal error rate

0.77

FRR at FAR 1%

60+40

Genuine + impostor

The VLM is a scribe, not the engine

Scope — EU AI Act Annex III alignment

Out of scope (Annex III high-risk uses) — Employment, eligibility, creditworthiness, education access, and similar decision-making. The Discovery modules (HR style, learning style, wellness, partner compatibility) are presented as reflective entertainment only — never as inputs to a hiring, admissions, lending, or clinical decision. Tier-gated to Free precisely so they cannot be sold as a paid HR/medical screening tool.

In scope (forensic pre-screening with human-in-the-loop) — Signature comparison, document forgery, document semantic checks, document triage. These produce a measured, calibrated reading with a clearly-labelled confidence band. A certified Forensic Document Examiner (FDE) keeps the final say — we surface evidence, the expert renders judgment.

If you are evaluating Gramavec for a regulated procurement, ask us for the assurance pack: model card, sample-size disclosure, retention configuration per tier, and the audit log schema.

Visual evidence — reviewer-replicable

Every forensic comparison renders the same four artefacts. A reviewer (or a certified Forensic Document Examiner) can regenerate them locally from the original images and audit each claim.

• Side-by-side crops of the reference and questioned ink, at common height.

• Aligned absolute-difference overlay drawn in red over the reference.

• Per-pixel SSIM heatmap (bright = more different).

• Skeleton overlay with endpoints (red) and branch points (blue).

Replication & cross-checks