OPEN ALGORITHM · VERSION 2.0.0 · EFFECTIVE 2026-10-17
Everyone deserves a chance to be discovered.
This is exactly how Hoppps decides what to recommend and how it protects creators and audiences — generated from the code that runs it. Subscriber counts never decide who gets seen; criticism is never punished; restrictions need evidence, are temporary, and can be appealed.
1. The recommendation score
H = 0.30R + 0.25S + 0.15E + 0.15Q + 0.15D
FactorWeightWhat it measures
R Viewer relevance30%Recommend videos matching viewer interests
S Viewer satisfaction25%Reward content viewers genuinely appreciate
E Meaningful engagement15%Measure valuable audience interaction
Q Content quality15%Evaluate originality and viewing experience
D Discovery opportunity15%Give underexposed creators a chance
Every factor is between 0 and 1. Small samples are pulled toward typical values, so three early views can’t make or break a video. Subscriber and follower counts are not inputs. Discovery opportunity (D) is 1.0 for a creator nobody has been shown yet and falls with how often their work was shown in the last 30 days, to a floor of 0.1.
Established creator · 500,000 subscribers0.733R .90 · S .80 · E .80 · Q .85 · D .10
New creator · 50 subscribers0.900R .90 · S .90 · E .85 · Q .85 · D 1.00
That doesn’t guarantee the new creator more views — it gives their video priority when the two compete for the same viewer.
Every page of recommendations
At most 2 videos from one creator, and no more than 3 in a row from one topic.
Every 4th place is saved for an emerging creator (discovery opportunity ≥ 0.7).
Places 5n−2 (3rd, 8th, 13th…) go to a video in its discovery test, earliest tests first — only when it fits the viewer’s interests.
Anything you mark “Not interested” is hidden for you. Videos you’ve finished are shown less, not hidden. Every recommendation can tell you why it’s there.
2. Fair exposure for new creators
Stage 1Initial exposure100 impressions
Stage 2Expansion 1500 impressions
Stage 3Expansion 22,500 impressions
Stage 4Expansion 310,000 impressions
ThenContinued distributionBased on audience response
An impression counts once a recommendation is actually on a signed-in viewer’s screen. To move on from a stage, a video’s “satisfied” viewers (watched at least half, or liked it) must make it 75% likely it satisfies at least 90% as well as typical, with at least 10 views. Targets are tests, not promises: they only fill when there are interested viewers. Videos that stop advancing stay recommendable on their score — older videos don’t disappear just for being older.
How many people are watching never decides the order. A stream with 10 viewers can lead one with 10,000 when its viewers stay (retained = 3+ minutes), enjoy it and join the conversation (counted as people, not messages). Stream quality is an eligibility check, so expensive equipment can’t buy a better spot.
Recommended Live Personalized livestream recommendationsEmerging Live Discover smaller streamersGaming Live Gaming broadcastsEntertainment Live Entertainment and variety streamsEducational Live Learning and informational streamsCreative Live Music, art, and creative broadcastsFollowing Live Streams from followed creators
4. Criticism is allowed. Harassment loses amplification.
ContentClassificationConsequence
"Why I disagree with another creator"Legitimate criticismNone
"Their content is overrated"Negative opinionNone
"Five problems with their streaming style"CommentaryNone
Encouraging viewers to threaten themHarassment incitementRestrictions if proven
Organizing an audience harassment campaignCoordinated harassmentRestrictions if proven
Fabricating serious allegations about themDeceptive targetingRestrictions if established
No proof, no penalty
Evidence that counts (once verified by a second reviewer)
Authenticated livestream recording
Verified video statement
Corroborated messages
Evidence of coordinated harassment
Reliable moderation record
Other independently verifiable documentation
Screenshot (counts only once corroborated)
Never grounds for a penalty
Mass reporting alone
Dislikes
Negative comments
Unverified screenshots
Rumors
Personal disagreements
A creator's popularity or lack of popularity
Every decision is made by a person with a moderation role, written down, and explained to the person affected. AI may point a reviewer at something; it never decides a contested case.
Temporary discoverability suspensions (creators)
Violation 1Warning
Violation 224 hours
Violation 33 days
Violation 47 days
Violation 514 days
6+30 days
Counted over a rolling 90 days; overturned decisions never count. A serious case can start further up the ladder. Channels are never banned for ordinary harassment violations, and when a suspension ends eligibility returns in full — the ranking never reads expired suspensions, so no hidden penalty remains.
Coordinated harassment7–30-day interaction restriction7–30 days
Credible threatImmediate protective restriction while it is investigatedUntil reviewed
Encouraging swattingSevere temporary account restriction30–90 days
SwattingLong-term restriction and potential referral to authorities180–365 days
Deliberately false emergency reportSevere restriction30–90 days
Abuse of reportingLoss of standard reporting features30–90 days
Restrictions limit interaction, never watching — and never the ability to report an emergency. Nobody is penalized for a legitimate police report, and an emergency report that turns out to be wrong isn’t treated as malicious. Appeals are open for 30 days and reviewed by someone who didn’t make the decision.
5. Emergency Chat Lockdown
Streamers (and moderators they choose) can lock chat instantly. A viewer’s report alone can never lock someone’s chat — it’s reviewed urgently by HOPPPS safety staff. 15 minutes is a review point, not a reopening time: chat stays closed until the streamer reopens it (+15 min, +30 min, until restored).
HOPPPS — Safety LockdownChat has been temporarily disabled for safety reasons.Please respect the streamer's privacy.Chat will be restored when the safety restriction is lifted.
Panic Mode does it all in one press: disable public chat, pause paid messages, disable new clips, stop automated highlights, preserve evidence, open the safety panel. Footage from an incident stays out of recommendations until it’s been reviewed, so a dangerous moment can’t become trending content.
The algorithm is pure, auditable code: src/lib/hfds (recommendation, moderation, safety, transparency), with tests for every formula and rule in tests/hfds-*.test.mjs and the written policies in docs/hfds. Moderation never changes a score — it only decides where something may be recommended (homepage recommendations, suggested videos, trending, discovery feed, shorts feed, livestream recommendations, emerging live, search results), so restrictions can be audited on their own.
How to audit the numbers
Creator opportunity — impressions per creator over 30 days; Gini coefficient of that distribution (0 = equal).
Exposure concentration — share held by the top 10% of creators, and the Herfindahl–Hirschman index.
New creator discovery — share of impressions to creators with under 1,000 impressions before the period.
Viewer satisfaction — average share watched for views that started from a recommendation.
Enforcement accuracy — share of decided appeals that overturned the decision.
Safety — median time from an urgent viewer report to a chat lockdown.
Version history
2.0.0 · effective 2026-10-17 — First open release of the Fair Discovery System and the Creator & Audience Protection System.
Videos ranked by H = 0.30R + 0.25S + 0.15E + 0.15Q + 0.15D; subscriber and follower counts are not inputs.
Staged discovery for new uploads (100 → 500 → 2,500 → 10,000 impressions) with a Bayesian performance test.
Creator-diversity limits and emerging-creator slots on every recommendation page.
Live streams ranked without concurrent viewer counts; Emerging Live for smaller streamers.
Temporary discoverability suspensions with a 90-day ladder, automatic restoration and appeals.
Audience restrictions, swatting incident management and Emergency Chat Lockdown with Panic Mode.
Testing: Unit tests in tests/hfds-*.test.mjs cover every formula, threshold and permission rule in this release. Rollback: Ranking falls back to the newest-first order of Hoppps 1.x if this version is withdrawn.
