SYSTEM ADMIN

System admins see a read-only operational mirror, not one line more than the impact analyst. Both roles can read these aggregate, de-identified checks; only the Impact Analyst role can close a reporting period or export the aggregate pilot report. Holding the system admin role does not expose an individual profile.

Manage · Measurement & impact reporting

Pilot figures & impact reporting

Three-month pilot · 1 April 2026 – 30 June 2026 · single arm, no control group. Everything below is aggregate and de-identified.

Three conditions to read before using any number on this page

  • De-identified. Behaviour logs hold no identity. This page has no search by person, no profile button, and no path back to a particular account.
  • Minimum group size of 10. Any group smaller than 10 people is hidden, even when the total is large enough, because a small group can be re-identified.
  • Self-reported figures are an indicative proxy, not an objective measure. They must not be presented as evidence of effectiveness.
Period: Q2 2026 64 enrolled Updated 27/07/2026 06:00 🗂️ Compare with the previous period
Pilot sample
64
enrolled · target 50–100
Baseline–endline pairs
41
64% of those enrolled
Still using at day 30
31%
20 of 64 · NOT a success measure
Continued after a risk warning
68%
average across the period

Quiz score, before and after

n = 41 matched pairs

A separate knowledge assessment runs at the end of onboarding and again two weeks after each participant's baseline. Marked out of 20.

Baseline (before)11.4 / 20
Endline (after two weeks)15.2 / 20

A difference of +3.8 points (+19 percentage points). That is a change within the same group. With no control group it cannot be stated as "the app increases knowledge" — practice effects from sitting the test twice and self-selection into the pilot have not been ruled out.

Broken down by age group

Age groupnBaseline EndlineChangeStatus
18–241810.815.0+4.2shown
25–341412.115.6+3.5shown
35–446hidden · n < 10
45 and over3hidden · n < 10

The 9 people in the two hidden groups are still counted in the total of 41 above; only the breakdown is withheld.

Retention at day 1, 7 and 30

base sample 64
Day 172% · 46 people
Day 745% · 29 people
Day 3031% · 20 people

Retention is not a success measure for this product and must not be treated as one. For a tool about gambling harm, high retention can be a bad sign. Read it only alongside the self-reported survey and the risk-warning measure below. A report that leads with retention is reporting against the wrong goal.

Self-reported survey (opt-in)

INDICATIVE PROXY

Periodic questions about real-world gambling. Entirely voluntary; declining costs nothing and changes nothing about the experience. Stored de-identified.

Agreed to take part: 38 of 64 (59%) Declined: 26
Report betting less than before15 · 39%
Report no change18 · 47%
Report betting more than before5 · 13%

The only permitted way to present this figure: "38 people reported the following", never "the app reduced real gambling for 39% of users". These are self-reported answers, subject to social desirability bias, collected from a group that chose to take part.

The "more than before" group is only n = 5, so it must not be broken down any further. As it stands this survey is not sufficient to conclude anything about a gateway effect; answering that question would take an independent study design.

Continued after a risk warning

main behavioural measure

Counts how often the risk warning appeared and how often the player continued anyway after reading it. Measured by week of use, not by calendar week. 1,842 warnings shown in total.

78%
74%
70%
66%
62%
59%
Week 1Week 2Week 3 Week 4Week 5Week 6

A fall of 19 percentage points over six weeks is the most interesting signal in this period. But only about a third of the original sample is still there in week 6, so the fall may simply reflect that the people most easily drawn in left earlier. Do not read this chart as cause and effect.

Weeks 7 to 12 are not plotted because each week falls below the minimum group size of 10.

Choosing to learn, or choosing to reset

244 decision points

Recorded at the turning points: losing an asset, reaching a critical state, and the end of a run.

Back to learning · 132 · 54% Voluntary reset (avoidance) · 71 · 29% Left the app without choosing · 41 · 17%

The recovery draw is not live, so this period has no three-way comparison. The group that left without choosing is the one we most need to watch and the one we know least about — they are no longer in the sample to ask.

Support link taps

214 taps · 37 people
Where the link sitsTapsShare
Critical state screen6832%
End of run screen5224%
Seven-day cooldown4421%
Footer & support centre3114%
AI companion199%

This is a count and nothing more. We do not store who tapped, we do not store AI conversations, and nothing is passed to the helpline. A tap does not mean the person called or got help — it measures reach, not outcome.

Previous closed period · Q1 2026

Read-only aggregate

This system view exposes only the closed aggregate needed to check platform operations. Cohort interpretation, reporting and instrument review remain with the impact analyst.

MeasureQ1 2026Q2 2026Comparison note
Enrolled1264Q1 was an internal trial; Q2 is the pilot cohort
Matched assessment pairs041Q1 had no endline; no knowledge-gain claim
Day-30 retentionnot measured31%Operational signal, not a success measure
Continued after a risk warning71%68%Descriptive only; no causal claim

Small groups remain hidden at n < 10. No row can be opened into an individual account.

Reporting hand-off

This role can verify platform aggregates but cannot close the period or export the report. Those actions remain with the Impact Analyst and its limitations block.

View closed periods

The first period is only meant to establish a baseline, not to hit a threshold.

Limitations of this pilot — these travel with every external presentation

  • Single arm, no control group. There is no comparison group that did not use the app, so no before-and-after change can be attributed to the product.
  • Self-selected participants. People who enrolled may already care about gambling harm; the results do not generalise to the wider population.
  • A sample of 64 over three months is too small to detect a real effect, especially within subgroups.
  • Attrition. Only 41 of 64 have a complete baseline–endline pair; people who left early are barely measured at all. Later figures always lean toward those who stayed.
  • The survey is self-reported and cannot be checked against real-world betting.
  • Commitment to publish. Negative or neutral results are published in full. If we see a signal that people bet more after using the app, the agreed response is to pause or change the product and publish the finding.

What must not be done with this data: reading it back into one person's gameplay; using it to personalise odds, rewards or notifications; or implying that the product helps anyone win real bets.