A moderation score is evidence; shipping it as a verdict is how a classifier rejects an applicant

The failure has a shape, and it is a data model problem before it is a policy problem. A three-route contract emits one of allow, review or block ^[2]. A downstream field typed `approved: true | false` carries two values ^[9]. One of the three states cannot survive the trip, and the one that dies is review, because a Boolean has no way to say not yet ^[1]. That is the mechanism by which an uncertain classifier output reaches the rubric scorer looking settled: the note describes a continuous score being rounded into a Boolean because the schema downstream has nowhere else to put it, after truncation and normalization have already thinned the context ^[9].

The tenant argument is the one that kills the global cutoff outright. In the note's framing, one tenant may permit redacted case studies while another forbids pasted customer messages for confidentiality reasons; the model score is identical and the correct route is not ^[8]. A single threshold cannot encode that, which is why the author puts the route in a tenant-aware policy layer rather than a score comparison ^[8].

What the policy function has to keep is unglamorous: raw category scores, the policy version, the model identifier, the hash of the evaluated text, and the route taken, stored only as far as retention rules allow ^[10]. The stated invariant is that an operator can reconstruct why an item entered the queue without inferring it from one blocked flag ^[11]. That is a debuggability requirement dressed as logging, and it is also the only artefact that exists if a rejected candidate ever asks how the decision was made.

Measurement is where two honest teams can both be wrong. The note points out that "false-positive rate" can mean flagged items reviewers clear, allowed items that should never have been flagged, or every submission unnecessarily interrupted, and that without one definition in the evaluation spec plus the confusion matrix, both teams can report improvement over different populations ^[12]. The test set is supposed to be candidate-generated content segmented by tenant, locale, role family and content length, with the hard cases preserved: quoted abuse from a safety role, security terminology from a penetration tester, medical language from a benefits specialist ^[13].

The author is careful about what this does not settle. It is a decision note, not a claim that one set of numbers travels, and the author says outright they are unsure any static country switch stays correct without policy and legal owners reviewing the workflow ^[16]. Region can select a reviewed policy bundle, but geography alone does not decide which employment, privacy, platform or retention obligations apply; those depend on the marketplace, tenant, candidate location, role, and how the output is actually used ^[14]. Promotion of a threshold change is gated on replaying labeled tenant data and on both error rates and review workload staying acceptable ^[15], which means a tuning win that halves false positives while flooding the queue does not pass ^[2].