Fair ranking philosophy¶
iFixedX is an algorithm R&D vehicle: learn how to improve discovery at X-scale without shipping a separate “boost slot” product.
Principles (product north star)¶
No artificial boost lanes — do not reserve feed slots that only small accounts can win; that feels rigged.
Keep stochastic surfacing — the feed stays partly “lotto-like”; better candidate generation and scoring improve odds for merit, not guaranteed impressions.
Merit over pity — smaller accounts should win when posts are genuinely strong, not because of diversity injection.
Richer signals (directional):
Semantic quality / substance (Grok seasoning, bounded cost)
Originality vs noise
Conversation quality
Author consistency over time
Who engaged (network quality), not raw follower count alone
“Exp sharing” (design space)¶
Ideas under discussion (not all implemented):
Reputation-weighted engagement
Discovery credit when quality accounts amplify smaller ones
Author-level features accumulated in Pipetrix
mixerMetaJsonand optionally Account rollup (seenPostCount,lastSeenAt)
Implementations should extend existing types (Tweet, canonical ingest, search.ts) — not a parallel social graph DB unless necessary. Replay raw Grok batches via Grok ingests when re-scoring without new xAI calls.
Two rankers today¶
Mode |
Location |
Character |
|---|---|---|
Default (engagement-style) |
|
Closer to velocity / engagement heuristics |
iFixedX rank |
|
Text match + recency + anti-spam sliders |
Grok supplies candidates and context (mixerNote, topic batches); local code owns ordering.
Evaluation¶
Use Ranking lab sliders; ingest snapshots rankingLab into canonical mixerMetaJson for offline comparison.
Grok collaboration¶
When Grok on the box proposes scoring changes, start from:
src/mixer/merge.tssrc/mixer/buildFeedPosts.tssrc/feedtrix/canonicalIngest.tsserver/forYouCorpus.tsserver/search.ts