An outsourced analyst's triage clicks become your model's labels — do you accept that write path?

## What the decision actually is The question sounds like a vendor question and it is not. Whether the outsourcing partner is reputable is a procurement matter with procurement answers. The engineering decision is that a click in somebody else's queue is now a write into the weights of a production model, and it is going to be read by the next retrain whether or not anybody wrote that down. So the call to own is: this is a production authorization surface. What are you willing to fund on it, what does that cost the people who depend on the model, and what can you honestly claim afterwards? ## Why the vantage is unusual Three properties, stacked, make this path worth a principal-level decision rather than a line in a runbook. **The writer set is not yours.** You do not hire, offboard or individually authenticate the people producing the writes. Outsourced tiers frequently authenticate as one shared identity, so the enumerable writer set is "whoever the vendor has staffed this week", which changes without notifying you. **The write is authorized and looks exactly like the work.** There is no anomaly to detect. A disposition that is wrong on purpose is byte-identical to a disposition that is wrong through fatigue, and both are indistinguishable from a correct one that you happen to disagree with. **Consumption is automatic.** The value of an outsourced triage tier is throughput, and throughput usually means the dispositions flow straight into the next training table. Reach is therefore immediate and durable, with no step in between where anybody looks. ## The levers, and who pays for each **Attribution.** Per-analyst identity on each write, so that a change has a principal against it. This is mostly a contract and identity problem rather than a technical one, and the vendor absorbs the process cost. Without it every later question about that corpus has the same answer: "the vendor", as one undifferentiated principal. **A gate before consumption.** A retained snapshot of each training table with a diff against the previous one, owned by somebody on your side who is accountable for looking. This is the single highest-value item, because it is also what makes any future investigation answerable. It costs a named person's attention on a recurring basis, which is a real budget line and the one most likely to be quietly dropped. **Dwell.** A window in which writes are not yet eligible for a retrain. This one is genuinely expensive in a way the others are not: it is paid entirely by the detection team in staleness. A pattern dispositioned today does not shape the model this week. In a fast-moving detection context that trade may be unacceptable, and if so the organisation has chosen the exposure — which is fine, provided it chose it on paper. **Scope narrowing.** Limit what dispositions from this tier are allowed to influence. Perhaps they inform triage ordering but a separate, smaller, attributable population produces the labels a retrain consumes. This preserves the throughput you outsourced for while shrinking the writer set that reaches the weights, and it is often the best answer nobody proposes. ## The claim you can stand behind This is the part a lead has to get right, because somebody will eventually ask for a statement in writing. With attribution, a gate and a retained diff, the honest claim is a scope: these principals held write over this window, every batch that reached a training table was diffed and the diff was reviewed by this owner, and no write sat unexamined for longer than this. That is a strong, defensible statement. What it is not, and never becomes, is "the training data was not tampered with". No amount of process supports that sentence, because it asserts something about intent behind writes that were all individually legitimate. Offering it is worse than offering the scope, because it invites reliance on an assurance nobody can back. ## When to refuse There is a version of this where the answer is no. If the contract cannot deliver per-analyst attribution, nobody will own a recurring diff, and the retrain cadence leaves no room for dwell, then you have a production surface with an unenumerable writer set, no gate and no record. That combination should go up as a finding with an explicit decision attached, not be absorbed quietly by the team that inherited the pipeline. Escalating it is the job; the organisation may still accept it, and that acceptance is exactly what you wanted on record. ## What an interviewer is listening for That you convert an uncomfortable arrangement into a stated authorization surface with named costs and a claim you can defend, rather than either refusing on principle or waving it through on the strength of a vendor's reputation.