How a review is decided
A review can be closed by a policy or by a person, and which one happened changes everything about what the player can appeal. In the sampled desk 3,200 of 4,000 reviews were resolved against a written policy and 800 were referred to a human. The automatic path is faster and more consistent; the human path is where a rule that read the wrong thing can be caught.
- reviews
- 4,000
- resolved by policy
- 3,200
- share automatic
- 80.0%
- referred to a person
- 800
- share referred
- 20.0%
- reviews with a record
- 4,000
Read the signals
Five comparisons the account already supports, run continuously rather than after a complaint. 6,000 flags in the sample, 96.0% of them relative to the account own history.
Choose a response
Nothing, a message, a limit or a break, or a restriction - one of four, graded by how much it takes from the player. 960 of 4,000 reviews in the sample changed access.
Write it down
Signals, crossing dates, rule version, path, response and outcome - one record per review, including the 1,600 that ended with nothing done.
Take the appeal
A route outside the desk that decided. 480 appeals in the sample, 176 of them upheld, which is 36.7% - high enough that the route does real work.
A protection review is closed either by an automatic decision against a written policy or by a person. In the sampled desk 3,200 of 4,000 reviews (80.0%) were resolved automatically and 800 (20.0%) were referred to a human. The record names which path was taken, because an automatic decision is challenged against the rule and a human one against the judgement.
The two paths, and what each is for
An automatic decision does one thing very well: it applies the same rule to the same facts at the same speed, thousands of times, with no drift. That consistency is exactly what a protection duty needs, because a rule that produced a review in one week and not the next for the same pattern would be indefensible. The cost is that a written policy cannot recognise an explanation it was not written to hold.
A referral to a person exists to catch the cases the rule read wrongly. In the sampled desk 800 reviews were referred, 20.0% of the total, and the referral is triggered by the shape of the case rather than by the size of the pattern: a signal that conflicts with another, a spend ratio that fails only slightly, a self-reported flag with no behavioural flags behind it. The two paths are complementary and the record keeps them apart.
What the deciding rule actually says
| The case | Path | Why |
|---|---|---|
| Two behavioural flags, no explanation on file | policy | the rule was written for exactly this shape |
| A self-reported problem flag, alone | person | a statement by the player is not a behavioural inference |
| One flag from each of two families | person | conflicting evidence is the case a policy is worst at |
| A spend ratio inside 10.0% of the trigger | person | a near miss is where a threshold is doing the deciding |
| Three or more flags from one family | person | repetition of the same signal may be one cause, not three |
| Two flags, one already explained on file | policy | the explanation is on the record and the rule reads the record |
That table is the whole design in miniature. The policy is trusted with the routine and the person is reserved for the shapes where reading a rule is not the same as reading a case - which is also why the referral rate is a number worth publishing: a desk that refers nothing is applying its rules blind, and one that refers everything has no rules at all.
The decision and the record are separate acts
Even an automatic decision has to be written down in a way a person can later test, because the appeal reads the record rather than the rule. In the sampled desk all 4,000 reviews carry a record, and it names the signals, the crossing dates, the rule version that was applied, the path, the response and the outcome. That is why the 1,600 no-action reviews are the most useful rows in the whole log: they are the cases where the rule fired and a human or a policy then said no, and a log that only contains actions taken cannot show that the second obligation was performed at all.
Worked example / sample D and the referral rate
- reviews: 4,000
- resolved against the policy: 3,200
- referred to a person: 800
- 3,200 / 4,000 = 80.0% automatic
- 800 / 4,000 = 20.0% referred
- the 320 restrictions and closures, being the heaviest response, were all on the referred path
- 320 / 800 = 40.0% of referrals ended in the heaviest response