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.
A candidate group is a starting point for review
The engineering record names the algorithm evidence-cosine-complete-link-v1. It derives source representations from unique substantive chunks and excludes subject-context chunks from voting. A duplicated quote or an attachment-derived alias should not multiply a source’s influence merely because it appears several times.
The reported pair rules combine cosine similarity with either a shared thread within 30 days or sufficient participant overlap within seven days. The recorded thresholds are 0.82 for the thread route and 0.90 for the participant route, with participant Jaccard overlap at least 0.5. These are versioned implementation choices, not universal values or calibrated probabilities.
Prevent a chain from becoming an accidental merge
Complete-link compatibility requires every cross-pair to pass before two candidate groups join. A message that resembles both sides cannot bridge otherwise incompatible requests. Conflicting explicit request references or due dates can reject a pair before semantic similarity is considered.
compatible_pairs = {frozenset(("a", "b")), frozenset(("b", "c"))}
def can_join(left, right):
return all(frozenset((a, b)) in compatible_pairs
for a in left for b in right)
assert can_join({"a"}, {"b"})
assert not can_join({"a", "b"}, {"c"})Evaluate both incorrect merges and missed connections
A conservative algorithm can leave too many singletons. That trade-off must be visible in evaluation: count separate requests incorrectly merged, related evidence left apart and the reviewer work needed to correct each result. Synthetic vectors test the merge mechanics, but real labelled examples are necessary to judge semantic quality.
The reported implementation bounds candidate scope and saves group identities, scores and provenance. This supports repeatable review and stable navigation. It does not make a candidate group an approved issue. The distinction between candidates, AI suggestions and applied issue membership is part of the architecture, not just a label in the interface.
References & further reading
Official documentation for the methods discussed. Application-specific details come from the development records described in the series note.