How we order the queue

1. Scope

This page documents the production request queue and its shadow evidence models. The deterministic PPS v1 score continues to order eligible levels. Bayesian models observe outcomes; they do not change request decisions or queue order.

2. Six separate concepts

Review decides whether staff recommends a level. Priority orders accepted levels. Outreach records contact work. Confirmed submission means a moderator was actually reached. Outcome records later official state. Probability summarizes eligible evidence only when safeguards pass.

3. Human review

Reviewers choose Rate, Feature, Epic, Legendary, or Mythic. This is an Avenue recommendation, not an official Geometry Dash rating and not a prediction generated by the model.

4. Priority Point Score

The queue score is additive:

Higher complete scores are considered first. Unknown Creator Points stay unknown and never receive the CP zero benefit.

5. Prestige component

The tier values are Rate 0, Feature 1.25, Epic 2.5, Legendary 3.75, and Mythic 5.

6. Creator opportunity component

C is the current Creator Points total of the Geometry Dash account that uploaded the online level. It is never inferred from the requester, verifier, host, or credited collaborators.

Unknown CP is not 0 CP. A Rate level from a verified 0 CP uploader receives approximately 3.22 points from G, while an uploader with 4 or more CP receives 0. If Avenue cannot verify current CP, the level remains temporarily outside the ranked queue with priority pending instead of receiving a guessed score or rank.

7. Waiting component

W increases only after a successful outreach cycle in which the eligible level was not submitted. Recommending a different tier does not reset waiting.

8. Ranking and public bands

Only complete scores enter the ranked queue. They are ordered by total score, raw waiting cycles, queue age, and stable queue ID. Levels still waiting for verified creator data appear separately as Priority pending and receive no rank.

Public pages translate positions within the complete population into Top priority (top 10%), High priority (next 20%), Standard priority (next 40%), or Lower priority (remaining 30%). Pending entries are excluded from these percentiles. A public band is not an exact internal rank; exact rank and raw score remain private for stability and operational privacy.

9. Outreach episodes

An episode is one bounded attempt to move a level from recommendation toward an official outcome. Requeueing after an unsuccessful episode starts a new episode, records a reason, resets W to zero, and refreshes CP.

10. Normalized opportunities

Evidence is counted by episode + level + target. This prevents one conversation from being counted repeatedly while allowing genuinely different targets to remain separate Bernoulli opportunities.

11. Follow-ups

Follow-ups to the same private target update the same opportunity. They do not create extra trials. A different target creates a separate opportunity.

12. Target privacy

Target labels and normalized comparison keys stay private. Public responses never expose moderator identities, routes, target counts, or staff notes.

13. Confirmed submission

“Reached a moderator” is recorded only after a staff member explicitly confirms that fact. Planned outreach, attempts, and messages without confirmation are not promoted into submission evidence.

14. Fixed outcome window

The main rating outcome is observed exactly 30 days after the episode’s first confirmed submission. Provider downtime defers the check rather than silently recording failure.

15. Eventual outcome

The system also preserves whether a rating was observed later and the official observed tier. This descriptive field is separate from the fixed 30-day model outcome.

16. Exclusions

Levels already rated before outreach, missing target identity, provider conflicts, and authorized manual exclusions do not train the affected model. Every exclusion is durable and auditable.

17. Network eras

Exactly one network era is active. A new era marks a meaningful change in outreach conditions. Owner or Dev starts it; old observations remain attached to their original era.

18. Priors and carryover

New eras use a weak discounted prior based on the previous era, with an effective sample size of 4 by default. This provides continuity without treating an old network as current fact.

19. Access model

access_model_v1 estimates the chance that one real, eligible outreach opportunity reaches a moderator in the current era. It begins with a deliberately broad Beta prior and updates with resolved successes and failures:

20. Rating model

rating_model_v1 uses one completed, eligible episode as its observation unit. Success means the level was observed rated within the fixed 30-day window. Multiple moderator submissions in one episode still produce one rating outcome.

21. Subgroup shrinkage

Route and recommendation-tier subgroups borrow a weak prior from their global model. Small groups therefore remain conservative instead of producing extreme percentages from a few cases.

22. Credible intervals

Staff can inspect 50%, 80%, 90%, and 95% intervals. Public pages use the 90% interval and rounded whole percentages.

23. Evidence strength

Evidence strength combines eligible observation count, uncertainty width, and freshness. The public labels are Limited, Moderate, and Strong. They describe the model evidence, not the quality of a level.

24. Model statuses

Collecting lacks evidence. Provisional has useful but insufficient evidence. Active passes all safeguards. Degraded is stale or poorly calibrated. Paused was stopped by Owner/Dev.

25. Public activation

An access or rating estimate appears only while its relevant model is active. The combined overall path appears only while both component models are active. Activation and removal are automatic, audited, and reported privately to Owner/Dev. A missing estimate means its safeguards are not currently met.

26. Calibration

Predictions are saved before their outcomes resolve. The system tracks Brier score, reliability bins, and expected calibration error. A lower Brier score means the saved probabilities were closer to what happened, but small samples are explicitly marked insufficient.

27. Capacity forecast

capacity_model_v1 repeatedly samples from the access posterior and simulates the chosen primary opportunities. Follow-ups are never extra trials. It reports expected submissions and conservative commitments:

K80, K90, and K95 are lower commitment numbers at those posterior predictive probabilities. They are internal planning aids, not public promises while evidence is immature.

28. Overall path probability

The overall path combines sampled access and rating probabilities through Monte Carlo draws. Interval endpoints are never multiplied together because that does not produce a valid interval.

29. PPS calculator

Explore the deterministic score. This calculator does not submit or reprioritize a level.

30. Governance, glossary, FAQ, and changelog

Model versions are immutable. Changes to priors, likelihoods, features, readiness rules, or prediction meaning require a new version and a human-readable changelog entry.