About the implementation and examples
Engineering notes draw on author-supplied MailTrace development reports through 13 October 2026. They describe the documented revision, including local work, rather than independently verified live deployment. Experiment designs and architecture proposals describe work to evaluate. Examples are educational, use synthetic data and are not copied production code.
Different questions require different mechanisms
“What does this message appear to mean?” is an interpretation task. “Does this source belong to this case?” is a database and authorisation task. Asking a model to answer both mixes a probabilistic suggestion with an enforceable rule. In MailTrace’s documented organisation workflow, AI proposes meaning while backend validation controls what can be applied.
The same distinction applies to dates. A model may explain an apparent delay, but the saved source timestamp remains authoritative. If a message mentions a planned appointment, that extracted date needs a link to the exact passage and must remain distinguishable from the email’s sent time.
Treat the model response as a proposal
A well-formed JSON response can still contain a foreign source ID, an unsupported family or a stale interpretation. Schema validation is only the first check. Scope, membership, source hashes and the proposal’s base revision must be checked again at application time.
def validate_proposal(proposal, permitted_ids, current_revision):
if proposal["base_revision"] != current_revision:
raise ValueError("Evidence changed; review again")
proposed = set(proposal["source_ids"])
if not proposed or not proposed <= permitted_ids:
raise ValueError("Unknown or out-of-scope source")
return proposed
assert validate_proposal(
{"base_revision": 3, "source_ids": ["email-example-1"]},
{"email-example-1"}, 3
) == {"email-example-1"}Keep the useful uncertainty
A proposal can say “insufficient evidence” without failing the product. Forcing a confident answer makes uncertainty disappear from the interface rather than from the underlying case. The reviewed consolidation design keeps unassigned evidence visible for further review instead of treating it as resolved.
Acceptance criteria should cover adversarial and ordinary mistakes: a source removed after generation, duplicate IDs, an unknown family, and a reply that merely acknowledges receipt. None of these should be rescued by a high confidence score. Confidence can help prioritise review; it cannot grant permission or establish a fact.
References & further reading
Official documentation for the methods discussed. Application-specific details come from the development records described in the series note.