Problem Statement
The under-the-hood export provides visibility into explicit policy penalties (accountLabels, postLabels), but silent reach degradation frequently stems from semantic misclassification rather than punitive enforcement.
If an account’s graph placement drifts into an irrelevant community cluster (e.g., brigading, off-topic viral interactions, or noisy graph neighbors), outbound posts enter the wrong candidate generation pipeline (SimClusters, phoenix). Creators and developers currently have zero diagnostic tools to verify if their account is topicalized correctly.
Proposed Solution
Expose the precomputed macro-level cluster affinities already mapped to the author's user ID within the under-the-hood report.
Because User-to-SimCluster embeddings are precomputed offline in periodic batch jobs, surfacing the top $K$ cluster affinities requires an $O(1)$ read from the feature store rather than real-time impression aggregation.
Proposed Schema Addition
{
"period": {
"startDate": "2026-08-01",
"endDate": "2026-08-31",
"timezone": "UTC"
},
"accountAffinityProfile": {
"modelVersion": "simclusters-v2",
"primaryClusters": [
{ "clusterId": 1420, "label": "Software & Open Source", "weight": 0.82 },
{ "clusterId": 804, "label": "AI / Machine Learning", "weight": 0.54 },
{ "clusterId": 31, "label": "General Tech News", "weight": 0.28 }
]
}
}Why Engineering & Leadership Win Here
-
Kills the "Shadowban" Paranoia with Math: The primary mandate of the
under-the-hood export was to eliminate false shadowban accusations. When reach drops, creators inevitably blame silent censorship. Showing their cluster affinities objectively proves that an account isn't suppressed—it is simply rooted in a different or colder graph cluster than the author assumed.
-
First-Principles Compute Efficiency (O(1) Lookup): Unlike real-time impression telemetry, this requires zero streaming pipeline overhead or runtime event logging across billions of feed views. Because user-to-SimCluster embeddings are already precomputed offline in batch feature stores, exporting the top $K$ keys is a trivial key-value fetch during export generation.
-
Bot-Proof & Anti-Gaming: Macro cluster embeddings do not expose per-post ranking coefficients, decay curves, or Heavy Ranker features. It gives creators high-level graph directionality without handing engagement farmers a micro-optimization blueprint.
-
Completes the Open-Source Mandate: Open-sourcing the candidate generation logic in this repository is half the equation; giving developers visibility into their own graph inputs completes the feedback loop.
Problem Statement
The
under-the-hoodexport provides visibility into explicit policy penalties (accountLabels,postLabels), but silent reach degradation frequently stems from semantic misclassification rather than punitive enforcement.If an account’s graph placement drifts into an irrelevant community cluster (e.g., brigading, off-topic viral interactions, or noisy graph neighbors), outbound posts enter the wrong candidate generation pipeline (
SimClusters,phoenix). Creators and developers currently have zero diagnostic tools to verify if their account is topicalized correctly.Proposed Solution
Expose the precomputed macro-level cluster affinities already mapped to the author's user ID within the
under-the-hoodreport.Because User-to-SimCluster embeddings are precomputed offline in periodic batch jobs, surfacing the top$K$ cluster affinities requires an $O(1)$ read from the feature store rather than real-time impression aggregation.
Proposed Schema Addition
{ "period": { "startDate": "2026-08-01", "endDate": "2026-08-31", "timezone": "UTC" }, "accountAffinityProfile": { "modelVersion": "simclusters-v2", "primaryClusters": [ { "clusterId": 1420, "label": "Software & Open Source", "weight": 0.82 }, { "clusterId": 804, "label": "AI / Machine Learning", "weight": 0.54 }, { "clusterId": 31, "label": "General Tech News", "weight": 0.28 } ] } }Why Engineering & Leadership Win Here
under-the-hoodexport was to eliminate false shadowban accusations. When reach drops, creators inevitably blame silent censorship. Showing their cluster affinities objectively proves that an account isn't suppressed—it is simply rooted in a different or colder graph cluster than the author assumed.