Frequently Asked Questions

Aletheia is a hierarchical transformer classifier (ModernBERT backbone) trained on our own infrastructure on a fixed corpus.

The training corpus is about 2.5 million conversations from three sources:

1. More than 20 public and private research datasets (the bulk).
2. Real public U.S. federal court records (grooming and solicitation excerpts from CourtListener/RECAP).
3. Synthetic grooming scenarios generated in-house.

We do not store raw conversation content by default. A request is scored and the input is discarded. Where logging is enabled, we store a SHA-256 hash of the conversation and metadata, never the text itself.

We treat every input as potentially containing a minor's data and design against COPPA, GDPR, KOSA, and the DSA. We do not build profiles of children and do not link scores to identities. The platform sending us content remains the data controller for that content; Aletheia acts as a processor that returns a score.

There is no standing ability for anyone to read customer conversations, because conversation content is not retained. Serving logs contain hashes and scores, not text. In self-hosted deployments the content never leaves the customer's environment at all.

No conversation content is retained by default. The message is scored and thrown away. Optional audit logging is available, and that logging records a hash, score, threshold, and model version, not raw text, unless the customer explicitly chooses otherwise inside their own environment. Operational logs (auth and access) are kept on a short rolling window for security monitoring.

No. Models are trained and tuned on our own corpus. Customer traffic is not used for training or fine-tuning. If a customer opts into detection-improvement telemetry, it is limited to hashed, non-reversible signals, it is off by default, and it is opt-in per contract.

Aletheia produces a score and a predator identification. The customer platform decides what action to take. The product is built to support fast escalation of high-confidence threats, including a triage path aimed at reporting to NCMEC within 24 hours where the customer carries that obligation. Aletheia does not file reports on a customer's behalf by default; the reporting entity is the platform. Contracted reporting support can be arranged separately.

Because conversation content is not retained, there is no store of conversations for anyone to access.

Yes. This is a deployment tier and the intended path for data-residency and COPPA-sensitive platforms. It is a single-tenant container that runs inside the customer's own VPC, so conversations never egress to us. This tier is in active development and is offered to pilot partners on a case-by-case basis. The default today is the hosted API, which applies the same no-retention and hash-only logging behavior.

Every prediction is inspectable at the request level: it returns a calibrated score and the predicted predator participant. Optional audit logs record the conversation hash, score, threshold, and model version, so a decision can be reproduced without keeping the text.