A new policy line does not entitle its subject matter to a category in the triage model. Three measurements decide that: how much of this content the queue actually sees, how accurately the model separates it on labelled examples, and how much extra human review the category adds at current staffing. Run those three numbers first, then choose an entry mode. Policy intent alone is not an admission criterion.
The pressure runs the other way, of course. A policy owner writes a new line, and the reasonable-sounding next step is to give it a category, wire it to the same auto-route the mature categories use, and find out about the reviewer load afterwards. We see that sequence often enough that it deserves a name: admission on intent. Its failure signature is consistent — a low-confidence signal enters a working triage layer, false positives climb, and the human team the workflow was built to relieve absorbs the difference.
Everything below is scoped to workflow decisions: routing, thresholds, entry mode, reviewer load. What the policy should say is not an engineering question, and we do not treat it as one. That line is drawn deliberately in operational moderation versus political-speech adjudication.
What makes a new category worth automating?
Volume first, because volume bounds every other benefit. A category that appears in a handful of items per week cannot meaningfully change queue depth even with a perfect classifier, and the labelled sample needed to prove the classifier works will take months to accumulate. Sample the existing queue for the new category’s content before building anything — general review already surfaces it, and that history is your volume estimate.
Then accuracy, measured against human adjudication on a labelled sample of that category specifically — not the model’s aggregate figure. Aggregate precision hides per-category variance, and a new category is by construction the weakest one in the set: the fewest labels, the least mature guidance, the most reviewer disagreement about edge cases.
A category should only auto-route when its own measured precision against human adjudication clears the floor already met by the categories it joins; otherwise it enters as a review-only flag. That floor is workflow-specific rather than universal, because it depends on the cost of a false positive in that severity class and on what the platform’s existing categories achieve. What is universal is that the new category is held to the standard of the layer it enters, not to a lower one because it is new.
Third, marginal reviewer load. Estimate it before go-live rather than measuring it after: take the observed weekly volume, apply the measured false-positive rate at the proposed threshold, multiply by the average adjudication time for items in that severity class, and express the result as reviewer-hours per week. The estimate will be rough. It will still be the difference between a staffing conversation and a surprise, and mis-sized reviewer capacity is the most common way moderation pilots come apart — a pattern developed further in why pilots fail when reviewer load is mis-sized.
The three-input admission rubric
| Input | How it is measured | Admission signal |
|---|---|---|
| Signal volume | Sample of the existing general-review queue over a representative window; items per week matching the new category | Enough weekly items that queue depth or severity mix changes measurably, and enough to build a labelled set |
| Triage accuracy | Per-category precision and recall against human adjudication on a labelled sample of that category | Clears the accuracy floor set by the categories already auto-routing in the same severity class |
| Marginal review load | Weekly volume × measured false-positive rate at the proposed threshold × mean adjudication time for that severity class | Fits inside current reviewer capacity without displacing high-severity time-to-first-review |
Read the rubric as a conjunction. Volume without accuracy admits noise. Accuracy without volume spends engineering effort for no queue effect. Both without headroom in reviewer capacity means the category is affordable in principle and unaffordable this quarter — a scheduling answer, not a rejection.
Choosing the entry mode
Admission is not binary. There are three modes, and picking the right one is usually more consequential than the yes/no.
- Shadow mode. The category scores every item and writes its verdict to telemetry, but influences no routing. Use it when volume looks sufficient but the labelled sample is too thin to trust an accuracy figure. Shadow mode is how you build that sample cheaply, against live traffic distribution rather than a curated set.
- Review-only flag. The category annotates items in the queue and may raise them within their existing severity band, but never triggers automated action. This is the default for a category that clears volume but not the accuracy floor. False-positive cost stays bounded because a human still adjudicates every case.
- Ranked triage signal. The category participates in scoring and routing like the mature ones, including auto-action inside its high-confidence band. Reserved for categories that clear all three inputs, with the load estimate validated against the first weeks of production telemetry.
Default to the most conservative mode the evidence supports, and promote on measurement. Promotion is cheap; a rollback after a false-positive spike has already cost reviewer hours and, on sensitive content, credibility with the policy owners who asked for the category.
Where the new category lands also has to agree with the tier structure it enters — a review-only flag that quietly changes which reviewer specialisation sees an item is a tier change wearing a category’s clothes. Escalation tier design treats that layer directly.
Admission is a production change
A category addition modifies a live routing path, which makes it a model release, not a configuration tweak. It gets the same release-readiness gate as any other production model change: a version stamp, a rollback path, and per-category telemetry live before traffic shifts. The instrumentation this depends on — queue depth by severity band, time-to-first-review, model–reviewer agreement per category — is the same telemetry described in instrumenting a moderation workflow for review latency and accuracy. Without it, a category cannot be admitted on evidence and cannot be retired on evidence either.
Retirement matters more than teams expect. Categories decay: the content shifts, adversaries adapt, reviewer guidance is revised, and agreement between the model and the human team drifts. Set the downgrade condition at admission time — if per-category precision falls below the floor for a defined window, the category drops from ranked signal to review-only flag automatically, and the drop is logged rather than debated. A workflow that can only add categories accumulates weak signal indefinitely.
The audit record is what makes any of this reviewable later. A platform-trust reviewer asking why a category exists should find the volume sample, the accuracy measurement, the load estimate, the chosen entry mode, and the named decision owner — dated, and tied to the model and policy revisions in force. In our engagements the per-category precision, recall, and reviewer-load figures come out of the monitoring harness we ship with the workflow, which is also where the decision itself is recorded. The broader shape of that record is set out in what an operational moderation evidence pack contains.
For the wider workflow context this decision sits inside — triage design, human-in-the-loop routing, and the broadcast and platform work it supports — see our media and telecom practice, and the engineering scope we take on is described under services.
The honest open question is the volume floor. We have no defensible universal threshold for “enough items per week”, because it depends on adjudication time, severity, and how much of the queue the category can plausibly displace. What we do know is which direction the error runs: teams admit categories at volumes far below the point where automation helps, and pay for it in reviewer hours. If you cannot state your own floor in items per week, shadow mode is the answer until you can.
Frequently Asked Questions
What signal volume threshold makes a new category worth automating rather than leaving to general queue review?
Universal numbers do not exist; — it depends on adjudication time per item, the severity class, and how much of the queue the category could displace. The practical test is whether the observed weekly volume in your existing general-review queue is large enough to change queue depth or severity mix measurably, and large enough to build a labelled sample within a reasonable window. Below that, general review is the cheaper answer.
What triage accuracy floor should a new category clear before it is allowed to auto-route?
The floor is the per-category precision already achieved by the categories auto-routing in the same severity class, measured against human adjudication on a labelled sample of the new category specifically. Aggregate model accuracy does not qualify, because it hides per-category variance. A category that misses the floor enters as a review-only flag instead.
How do we estimate the marginal human-review load a new category will add before we turn it on?
Multiply the observed weekly volume by the measured false-positive rate at the proposed threshold, then by the mean adjudication time for items in that severity class. The result is reviewer-hours per week, which you compare against current headroom and against the effect on time-to-first-review for high-severity items. The estimate is approximate; its value is that it exists before go-live rather than after.
How do we retire or downgrade a category that stops performing after admission?
Define the downgrade condition at admission: if per-category precision falls below the accuracy floor across a defined measurement window, the category drops from ranked triage signal to review-only flag, and the change is logged with its telemetry. Automating the downgrade keeps decay from turning into a negotiation, and the audit trail shows a reviewer both the admission and the retirement on the same evidence basis.
Four signals that justify category expansion
Add a new content category only when manual review volume exceeds 200 items per week for three consecutive weeks and existing categories cannot absorb the edge cases. Everything else is detail.
