A review method built to show its work—and its limits.

Structured review

You do not have to know every question before uploading. Aletura combines universal checks across WHAT, MONEY & COMMITMENT, RISK and EXIT with contract-specific checks, kill questions and applicable scenarios.

• Every defined question receives an explicit state.
• Objective facts are extracted once and reused across findings, risks and reports.
• The framework is application configuration, not an improvised hidden prompt.

Full accepted-document analysis

Accepted readable content is extracted locally in full. Aletura rejects a document before analysis when a current acceptance limit is exceeded; it does not intentionally truncate an accepted document to fit a model context.

• 20 MB and 300 pages per document.
• 1.5 million extracted characters per document.
• Comparisons accept 2–3 files, up to 40 MB and 600 pages in total.
• Additional segment, table, cell and runtime safety limits also apply.

Evidence

Important factual findings retain source references where possible. Evidence identifiers are checked against known document segments, and quotations remain in the source language. Source matching and support checks reduce unsupported findings, but do not independently establish interpretation correctness, legal accuracy or enforceability.

• A finding can be traced to a passage and page or section.
• Referenced documents not supplied are marked as not provided.
• Evidence retrieval is benchmarked separately from model answer quality.

Privacy

Guest use requires no account; guest contracts and analysis reports are not saved. Signed-in users may explicitly save an encrypted masked structured result or, where policy permits, separately choose to retain the original contract with it. Unmasked extracted text and temporary identifier mappings are never saved. Detected personal identifiers are replaced locally before Mistral, and the complete outbound payload passes a privacy guard.

• Saving is explicit and limited by the account or workspace policy.
• Optional identifier restoration occurs locally after the response and restored names are not persisted.
• An explicit save may retain bounded masked Ask Aletura turns and the consumed follow-up count; raw questions and prompts are not saved.
• Automated masking cannot guarantee that every sensitive value will be detected.

Regional inference

The default Aletura endpoint is Mistral EU regional inference at https://api.eu.mistral.ai. The active deployment configuration must still be verified, and this statement does not establish complete end-to-end EU residency, provider Zero Data Retention or legal compliance.

Languages

EN, DE, FR, IT and RO are configured report-language and benchmark targets. Contracts are reviewed in their original language and evidence remains in the original wording.

• All five languages are currently not yet validated for full end-to-end quality.
• Deterministic multilingual benchmarks exist, but live fact extraction and follow-up quality are not yet certified.
• Other detected languages receive English reports and are likewise not yet validated.

Document text is untrusted data

Uploaded text is JSON-isolated from Aletura’s trusted instructions. Prompt-injection-like language is preserved as evidence and locally flagged; model calls expose no tools and do not follow document URLs.

• Suspicious output is rejected with one bounded retry before safe failure.
• Referenced external material is not fetched automatically.
• Aletura does not claim that prompt injection is solved.

Download rendering, not authentication

PDF downloads render report data submitted by your browser; JSON downloads contain browser-held report data. Neither format proves authenticity, completeness or that Aletura generated the submitted content. Downloads are not signed certificates.