Understand what a finding can—and cannot—prove.

Scope of the assessment

VeriTrustLab provides AI-assisted risk analysis and structured evidence review. Results are not legal conclusions, independent forensic certification, cybersecurity certification, financial advice, final proof, or a substitute for incident response.

An HTTP 200, a healthy host, a progress display, a completed model call, a saved result, and confirmed recipient delivery are different evidence levels. The health response explicitly does not replace private model-readiness or actual inference evidence.

Classifier scores and coverage limits

MailGuard and Swift are existing pretrained classifiers with pinned model contracts. The integrated public cards do not publish verified accuracy, precision, recall, or F1. False positives and false negatives are possible; confidence and risk labels are not calibrated guarantees.

MailGuard samples up to three approximately 1,200-character chunks, preserves tail context within that sampling approach, and runs a token-bounded classifier. The AWS runtime truncates tokenization to 512 tokens. Accepting a large original .eml does not mean its entire content is classified exhaustively.

Correlation uses usable scored evidence, deterministic risk floors, and policy rules rather than averaging every input into a calibrated probability. Model failures remain failed/degraded evidence and must not become a fabricated benign default.

Acquisition and authentication boundaries

Pasted text lacks original headers, MIME, attachment, and transport facts. Raw .eml supports message-carried evidence and available DKIM/ARC/DMARC evaluation, but ordinary uploads cannot reconstruct trusted SPF. Uploaded Authentication-Results headers are not a substitute for independent receiver evaluation.

Trusted SMTP observations require an authorized configured boundary. Authentication may establish a protocol result or domain alignment without establishing safe content or a trustworthy person. DNS records can change after analysis.

A stronger acquisition stage increases available evidence classes, not necessarily detection accuracy. Progressive linkage records a user-linked upgrade with same_message_verified=false; it does not independently prove both submissions are the same email.

Infrastructure context and attribution

Received headers can carry attacker-controlled relay claims. GeoTrace separates those claims from directly observed boundary facts. Approximate IP/provider location and illustrative map lines are not a measured physical route or the location/identity of a person.

Registration, geolocation, and reputation providers can return partial, stale, unavailable, or missing data. Missing coordinates stay missing. A successful model or valid Passport does not make an unsuccessful threat-intelligence lookup complete.

Campaign correlation and identity

MailGraph Campaign Memory is a tenant-scoped weighted overlap heuristic. Exact attachment hashes can qualify strongly; other matches require sufficient weight across entity families. It is not a learned adversary graph or proof of a shared malicious actor.

Common provider domains and ASNs can correlate benign messages. Explainable overlap still requires analyst context and false-correlation evaluation. Identity/BEC signals do not prove mailbox compromise, a sender’s intent, or a person’s identity.

URLs, attachments, and unsupported media

Swift analyzes URL strings and supporting patterns; the workflow does not render arbitrary destinations, follow redirects as a browser, or establish what a destination will serve later.

Attachment checks cover filenames, MIME/signature patterns, sizes, metadata, and hashes. They do not execute attachments or provide a complete malware verdict. Image, audio, video, deepfake, learning, and WebUSB capabilities mentioned in historical catalog/archive records are outside the active user workflow.

What an Evidence Passport verifies

Full-package verification checks the signature, both signed digests, key fingerprint, and configured issuer trust. It detects modification under that policy; it does not validate every original claim, establish a safe message, attribute an actor, or guarantee legal admissibility.

The original-byte hash is bound into the manifest. JSON verification does not separately retrieve the original .eml and recalculate its digest. Independent byte preservation requires the original and a hash comparison.

Redaction changes signed content. Historic hash mismatches or missing provenance must remain disclosed rather than being repaired by attaching an old signature to reconstructed current report data.

Availability and pilot scope

The documented model schedule is 07:00–19:00 IST, with startup readiness checked separately. SMTP itself is a separate runtime. Quotas use UTC-day/month boundaries; capacity is finite, and outages, limits, or unavailable required evidence can defer processing.

The reference describes one explicitly assigned pilot organization/mailbox/reporting route. Self-service multi-organization SMTP provisioning, broad provider validation, integrated detection benchmarks, sustained traffic guarantees, and a fresh hosted disaster-recovery certification remain outside the verified scope.

Human review before action

Use independent trusted channels before high-impact financial, legal, employment, moderation, or security decisions. Preserve source evidence, inspect gaps and provenance, and follow your organization’s response process.

For the workflow and scoring context, read Documentation and model evaluation disclosures. For a complete exported package, use the Evidence Passport verifier.