Ranking system¶

iFixedX ranks the same corpus two ways so you can compare X-style engagement ordering with a text + recency + anti-spam ranker. Scoring runs on the API server (server/search.ts); the client displays score and optional breakdown per tweet.

Ranker vs mixer (boundary)¶

Stage	Where	What it does
Ranker	`server/search.ts` → `scoreTweet` / `rankTweets`	Assigns `score` + `breakdown` per tweet for a given query and sliders
Corpus cap	`server/index.ts`	Slices tweets before `rankTweets` (`maxResults` query param)
Mixer	`src/mixer/merge.ts`, `buildFeedPosts.ts`	Merges live batch + Pipetrix canonical, unseen-first, tab filters — does not re-score

Home Top shows one ranked list (platform or refined) after mixer merge and time-window filter. Changing sliders requires a new API fetch (or re-rank from stored corpus on Explore).

Modes¶

UI label	API `mode`	Character
Default (engagement-style)	`platform_style`	Stronger weight on likes/retweets (log-scaled)
iFixedX rank	`refined`	Query overlap + phrase match + recency − spam; weaker engagement

Home Top tab uses one of these lists (after mixer merge and time filter). Explore always computes both for comparison.

Scoring components¶

Each tweet gets a numeric score and breakdown object:

Key	Meaning
`overlap`	Token overlap between query and post text
`phrase`	Bonus when full query appears as substring
`recency`	Exponential decay, 14-day half-life, × `3.2 × recencyWeight`
`engagement`	`log10(10+likes)×0.35 + log10(10+retweets)×0.25`, then × 1.65 (platform) or × 0.45 (refined)
`spam`	Negative — hashtags, promo words, tag-based bait (refined only)

Final score: overlap + phrase + recency + engagement - spam (spam stored in breakdown as negative spam).

Recency formula (code)¶

ageHours = max(1, hours since createdAt)
decay = exp(-ln(2) * ageHours / (24 * 14))
recency = decay * 3.2 * recencyWeight

recencyWeight is clamped to [0, 1] on the API.

Refined-only rules¶

With a non-empty query, if overlap + phrase < 0.8, score is multiplied by 0.35 (weak textual match sinks).
Strict match enables hashtag/promo penalties and extra penalty when tags look engagement-bait shaped.

Home / empty query¶

On For You, the query string is often empty. Overlap/phrase are ~0; recency and engagement (platform mode) still sort the feed. This is intentional — search ranking rules do not demote the whole timeline.

Ranking lab sliders¶

Controlled in UI, sent to /api/compare and /api/for-you:

Parameter	Default (server)	Effect
`recencyWeight`	`0.78`	0 = ignore age; higher = prefer newer posts
`strictMatch`	`1` → true	Anti-spam / hashtag penalties
`maxResults`	Compare: 30; For You: 35; xAI home: 12 (max 25)	Corpus size before `rankTweets`

Explore also supports omitRetweets (Recent Search query -is:retweet).

SearchParams in search.ts only includes q, mode, recencyWeight, strictMatch — maxResults is enforced in route handlers, not inside scoreTweet.

Slider values are snapshotted into Pipetrix mixerMetaJson.rankingLab on ingest — see Ranking lab snapshot.

API flow¶

corpus[]  →  rankTweets()  →  platform_style[] + refined[]

Explore: GET /api/compare?q=…&source=mock|live
Home: GET /api/for-you (live / mock / xai paths)

Fairness direction¶

Product goal: better candidate quality and seasoning, not reserved “boost slots.” See Fair ranking philosophy.

Future work (exp-sharing, Grok semantic scores) should extend scoreTweet / mixerMetaJson, not a parallel ranker DB. Raw Grok batches for replay live in Grok ingests.