Technical evidence and limits

Model and execution

The Python prototype uses scikit-learn with word and character TF-IDF features and logistic regression, trained on 120 synthetic messages for eight intent labels. The full Python workflow adds explicit rules. The browser lab executes the exported weights in JavaScript and exposes the raw model only. It does not send typed text to an inference server. Scores are not calibrated probabilities of correctness.

Fresh evaluation

Twenty AI-authored English messages were fixed after training, with no model or rule tuning to these results. Raw classifier: 13/20 correct against synthetic expected labels. Existing classifier plus rules: 14/20. JavaScript/Python classifier parity: 20/20 at reported precision. These results are not independent human or real-world validation. Earlier tuned development sets are development evidence, not an independent estimate of accuracy.

Event-chain boundary

The local prototype includes a per-thread SHA-256 event chain and a tampering check. Comparison with a retained chain or trusted reference can expose changes. An unanchored chain can be rewritten; it does not authenticate senders or prove the truth of a message.

Showcase and verification boundary

Case messages and cross-channel mappings are synthetic. Live account connectors, automatic identity matching, bank verification and production access controls are not demonstrated. Programmatic DOM checks cover selected interactions; they are not browser visual QA. Owner feedback is genuine usability evidence, while automated and simulated scenarios are not independent human tests. The videos illustrate the product and workflow; they are not evidence of live account connections.

AI use

AI assisted implementation, writing and synthetic data creation. The present classifier's generalization weakness remains visible. New consented, human-labelled cases and a production integration pilot are needed before claiming real-world reliability.

Context-aware evidence review

Explicit events are separated from quotes, negation, conditions and reported speech. Multiple events remain review-only; text never authenticates a receipt.

Context review: the original first-pass check scored 35/36. After fixing the known miss, all 36 development regression cases pass. These are not new unseen results. Amount-format and clause-safety checks are recorded separately. This is a deterministic context layer alongside the unchanged model; it is not independent validation or a new raw-model accuracy score.

Complete result with fingerprints · Try the analysis