Detecting near-duplicate functions in large repos

The literature on detecting near-duplicate functions in large repos is about accuracy. The people doing it spend their time on filing, sequencing, and what to say to someone who disagrees.

Deciding what the output is for

Staff turnover is the real adversary. The person who chose the threshold, knew which languages were parsed properly and remembered why one course was excluded will leave, and what remains is a number nobody can defend. Written-down reasoning is not bureaucracy here; it is the only mechanism by which a practice outlives the individual who set it up.

What the reviewer actually needs

Measure the reviewer's time, because it is the resource that runs out. Everything else — licence cost, compute, storage — is small and predictable. Reviewer minutes per submission is the number that decides whether the process survives contact with a busy term, and it is almost never instrumented, which is why so many pilots are judged a success and quietly abandoned within a year.

Making it survive the year

The expensive failure is not a missed match, it is an unexplainable one. A missed match costs you a case you never knew about; an unexplainable one costs an afternoon, a complaint, and a permanent reduction in how much anyone trusts the next result. Optimising recall while leaving the explanation thin trades a cheap failure for an expensive one, which is exactly backwards.

Putting it into practice

The right time to set up a Codequiry is before the term that needs it, not during.