AI Act transparency

Redline is a generative-AI tool that writes first-pass contract analyses. The EU AI Act (Regulation 2024/1689) asks providers and deployers of such systems to tell users what is happening under the hood. This page is that disclosure.

How Redline maps to the AI Act

Two articles apply squarely. The rest of the regulation is orthogonal to the portfolio-scale deployment Redline runs at, but the record below is kept deliberately short so it stays honest.

Article 13 — Transparency to users of high-risk AI systems

Art. 13

Every analysis carries a provenance block (model, snapshot, region, prompt-template version, reasoning-effort policy per pass, timestamp) rendered in the editorial footer below every report. The machine-readable receipt on this page packages the same information as downloadable JSON.

/transparency, AnalysisFooter

Article 50 — Disclosure of AI-generated output

Art. 50

Every report opens with a disclaimer banner that names the analysis as AI-generated and closes with a footer naming the model that produced it. The report export (Markdown and PDF) carries the same disclosure so the AI-generated nature of the output survives the handoff.

Disclaimer banner, AnalysisFooter

The pipeline an analysis traverses

Every contract analysis runs through the same five-stage pipeline. LLM calls go to Mistral in Paris; redaction happens in the browser before any text leaves the device.

Pass 0 — overview: Mistral extracts contract type, parties, jurisdiction, and a clause inventory. Pattern-based PII redaction runs first so names and IDs never travel.
Redaction (client-side): The user reviews detected entities in the browser and confirms which ones to redact before any full-text call. No raw PII leaves the device.
Pass 1 — extraction: Mistral receives the redacted full text and returns structured clauses anchored to the Pass 0 inventory.
Pass 2 — risk: Each clause is classified, scored, and grounded against a jurisdiction-specific statute allowlist. No free-text citations — only entries the catalog whitelists survive.
Chat (optional): The user may ask follow-up questions. Context is selected by keyword overlap today; top-k clauses are sent with the question to Mistral.

Operator rollback levers

These environment variables let the deployer change observable behaviour at runtime without a redeploy. They exist because transparency needs to survive incident response.

Force analysis locale

ANALYSIS_LOCALE_OVERRIDE

When set to a routing locale (e.g. `en`), every pipeline request ignores the user's UI locale and uses this one instead. Unset in production.

Pass 2 short-array retry

PASS2_RETRY_ENABLED= true

When `true`, Pass 2 reissues a batch once if the model returns fewer than half the expected clauses. Set `false` to disable during a provider incident.

Retention window (days)

RETENTION_DAYS= 30

Saved analyses auto-delete after this many days unless the user pins them. Default 30. Aligns with GDPR Art. 5(1)(e).

Known limitations

Redline is a first-pass assistant, not a substitute for a lawyer. The list below is the honest floor of what the model can and cannot do.

Mistral may invent facts that look plausible. The analysis should always be cross-checked against the clause text before relying on it. The product grounds output against a statute allowlist to limit one common failure mode (invented case law), not to eliminate hallucination in general.
The applicable-law catalog covers six EU member states (DE, NL, FR, ES, IT, PL) with dedicated statutes, plus EU-wide directives. Other EU jurisdictions receive EU-wide law only; non-EU jurisdictions fall back to general principles with no statute citations.
Redline is informational only and does not constitute legal advice. A human lawyer licensed in the relevant jurisdiction should review anything that materially affects your rights or obligations.
The six supported locales (EN, FR, DE, NL, ES, IT) are emitted natively by Mistral — no post-hoc machine translation. Other EU languages fall back to English analysis until they are reviewed.

Machine-readable transparency receipt

Every analysis can be exported as a small JSON document that records the provenance block, the pipeline, the operator levers, and the AI Act article references. The receipt is stable (`schema_version: "1"`) so auditors can diff receipts across releases. Download it from the expanded footer on any report.