MARPLA growth Signal: how to use a 0–100 triage score

Inputs and safeguards

The current Model 3 uses the observed 24-hour CCU peak relative to cumulative Visits, CCU relative to Universe age, positive one-hour, 24-hour and seven-day growth, a logarithmic peak contribution, a conservative 95% lower rating bound, and small supporting contributions from positive favorite and vote changes. Low-Visits games can receive a bounded early-traction contribution. The formula caps every component so one spectacular percentage cannot consume the entire score.

A rapid rise followed by a collapse is labeled as a spike, removing the first ratio contribution and applying a penalty. Last-hour coverage below 65% also reduces the score. The candidate card exposes reasons, risks, coverage, observed peak, and stage. Stages require repeated evidence before transition, and gaps can freeze the prior stage instead of manufacturing a collapse.

Worked example: equal score, different risk

Game A rises from average 20 to 40 CCU in the supported hourly windows, has 80% coverage, positive 24-hour growth, 5,000 Visits, and 200 votes. Game B jumps from 200 to 500 at one point but falls to 180, has 40% coverage, and no reliable vote response. Both may briefly show eye-catching percentages. Model safeguards can reward A’s supported growth while penalizing B for sparse evidence and a spike.

Suppose A receives Signal 58 and B receives 42. This does not mean A has a 58% chance of success or is worth 38% more. It means A currently ranks higher under the documented public-evidence model. A direct play-test or owner diligence can still reverse the business decision.

Use the score to allocate attention

A useful workflow ends in a decision card, not a screenshot of the leaderboard.

1. Filter candidates by the thesis, such as age, genre, and scale.
2. Open the score reasons, coverage, and risk labels.
3. Check the underlying timestamps and absolute bases.
4. Play the game and collect a second independent signal.
5. Save a next action and review date or pass with a reason.

Misuses that the score does not support

Do not compare scores from undocumented model versions as if the formula never changed. Do not hide a sparse-data penalty, copy the score into a valuation model, or call it predicted growth. Cumulative Visits, age, and public voting behavior vary by genre and lifecycle; peer context still matters.

The score sees only available public and derived fields. It does not know acquisition spend, retention cohorts, revenue, liabilities, source code, rights, team quality, or moderation risk beyond observed public signals. Use owner-data diligence before a consequential decision.

Decision record

Capture the model version, score, reasons, risks, coverage, and underlying timestamps. On the next review, explain whether the candidate moved because observations changed or because a newer formula changed the weights. The distinction protects the scouting history.

Primary sources

• Roblox Discovery
• Roblox Analytics