What raises and lowers your reach on X's For You feed: the tables

The reference tables from the report, on one page. Every value traces to a file in the repository. The calculator and simulation are on the main page.

Version 2026.10.03 · commit b412112 (2026-10-03) · Created by @SidneyCur

Action weights

The weight the ranker multiplies by a predicted probability.

#	Action	Weight	Parameter	Note
01	Reply	5.0	ReplyWeight	Every reply action
02	Reply on an original post from a mutual follow	20.0	5.0 + 15.0 boost	Original posts only, not replies or reposts
03	Quote	5.0	QuoteWeight	—
04	Share via DM	5.0	ShareViaDmWeight	—
05	Share via copy link	20.0	ShareViaCopyLinkWeight	Largest non-reply gain
06	Share	2.0	ShareWeight	—
07	Follow author	4.0	FollowAuthorWeight	Predicted follow after seeing the post
08	Repost	1.0	RetweetWeight	In-network only
09	Favorite	0.5	FavoriteWeight	—
10	Click	0.3	ClickWeight	—
11	Original post (no action)	none	no weight	Posting itself is not scored

Penalties (malus)

Negative weights. Each one lowers a post's score in proportion to the predicted chance of that action.

#	Action	Weight	Parameter	Example
01	Not interested	−47.52	NotInterestedWeight	A viewer taps not interested. P = 0.03 → −47.52 × 0.03 = −1.43
02	Block	−31.2	BlockAuthorWeight	A viewer blocks you. P = 0.005 → −31.2 × 0.005 = −0.16
03	Mute	−58.8	MuteAuthorWeight	A viewer mutes you. P = 0.02 → −58.8 × 0.02 = −1.18
04	Report	−234	ReportWeight	A viewer reports you. P = 0.01 → −234 × 0.01 = −2.34
05	Not dwelled (scrolled past)	−0.02	NotDwelledWeight	A viewer scrolls past. P = 0.50 → −0.02 × 0.50 = −0.01

Enforcement outcomes

What happens to an account or post after a flag, with an example of what the author sees.

#	Outcome	Effect	Duration	Example
01	SpamHighRecall (account)	Drops your posts for people who don't follow you	30 days	Your account is labelled. Strangers' For You feeds skip your new posts for 30 days. Your followers still see them.Source: enforcement_user.yaml, author_rules.rs
02	SpamHighRecall (post)	Drops that post as spam	30 days	One reply is labelled as spam and disappears from feeds for 30 days. Your other posts are not affected.Source: enforcement_post.yaml, registry.rs
03	RiskyHighVizReply (reply)	Applied to a flagged reply	30 days	A reply under a popular post is labelled risky. The repo does not trace its visibility effect, so treat it as a warning flag.Source: enforcement_post.yaml
04	Liveness check / CAPTCHA / Arkose	Challenge before continuing	Until passed	The account must complete a check before it can post again.Source: enforcement_user.yaml
05	Temporary suspension	perm: false, policy PlatformManipulation	Temporary	The account is locked under the platform-manipulation policy. The repo does not set the length.Source: enforcement_user.yaml
06	Permanent suspension	CSE and campaign cases	Permanent	Reserved for the most serious cases, such as child-safety reports and campaign abuse. The account does not return.Source: enforcement_user.yaml

Reply-related triggers

Reply behaviours the classifiers and rules flag, with an example of each.

#	Trigger	What is flagged	Label	Example
01	Fast reply spam	Replies in quick bursts	SpamHighRecall	A burst of replies in a few minutes, under many different posts. The repo does not publish the speed threshold.Source: enforcement_post.yaml
02	LLM-written text (llm_slop_post)	Machine-generated look	RiskyHighVizReply + account label	The same polished, generic line, "Great point, totally agree!", posted under many posts.Source: enforcement_post.yaml, enforcement_user.yaml
03	Gibberish	Nonsense text	SpamHighRecall	A reply like "asdf asdf 🔥🔥" that carries no meaning.Source: enforcement_post.yaml
04	Multi-step reply spam	Patterns across a thread	Classifier flag	A chain of replies under one post, each adding a little more promotion. The newest reply in the chain must be flagged for the label to stick.Source: grox/flows/reply_spam/
05	Coordinated spam	Patterns across accounts	Classifier flag	A group of accounts that reply to each other's posts within minutes.Source: grox/flows/reply_spam/

Exemptions

Accounts that skip some reply-spam labels or enforcement.

#	Exemption	Effect	Type	Source
01	High PageRank (score ≥ 50)	Skips RiskyHighVizReply	PageRank	task_write.py, constants.py
02	Grey badge	Skips RiskyHighVizReply	Badge	task_write.py
03	Very high follower count	Skips some enforcement (threshold in YAML is a mock value)	Follower count	enforcement_user.yaml

Author diversity multiplier

Applied to your 2nd, 3rd, and later posts in one feed. Decay 0.5, floor 0.25.

#	Post position	Multiplier	Calculation	Note
01	1st	1.000	`k = 0: 0.75 × 1 + 0.25`	No penalty
02	2nd	0.625	`k = 1: 0.75 × 0.5 + 0.25`	Score is 62.5% of base
03	3rd	0.438	`k = 2: 0.75 × 0.25 + 0.25`	—
04	4th	0.344	`k = 3: 0.75 × 0.125 + 0.25`	Same as the worked example
05	5th	0.297	`k = 4: 0.75 × 0.0625 + 0.25`	—
06	6th	0.273	`k = 5: 0.75 × 0.03125 + 0.25`	—
07	7th and after	→ 0.250	`Floor: decay 0.5, floor 0.25`	Never reaches zero

Final report: DO

Actions that raise the score. Numbers pair with the DON'T table.

#	Action	Score	How the code uses it	Example
01	Post original posts to mutual follows	Gain5 → 20 (boost +15)	`bidirectional_boost_eligible()` needs `!is_reply && !is_retweet && is_mutual_follow_author`. `reply_weight_for` adds 15 (xai-value-model/weights.rs).	Your original post "Shipping v2 today" reaches a mutual follow with predicted reply P = 0.10. Contribution: 0.10 × 20 = 2.00 (instead of 0.10 × 5 = 0.50).
02	Earn copy-link shares	Gain20	`ShareViaCopyLinkWeight` = 20.0 (home-mixer/params/param.rs).	A viewer copies the link to your post. P = 0.05 → 0.05 × 20 = +1.00.
03	Earn DM shares	Gain5	`ShareViaDmWeight` = 5.0.	A viewer sends your thread to a friend by DM. P = 0.04 → +0.20.
04	Earn quotes	Gain5	`QuoteWeight` = 5.0.	A viewer quote-posts your take with their own comment. P = 0.02 → +0.10.
05	Earn replies on your posts	Gain5	`ReplyWeight` = 5.0.	"Which tool did you use?" invites replies. P = 0.03 → +0.15.
06	Earn follows	Gain4.0	`FollowAuthorWeight` = 4.0.	After a post that shows expertise and a clear bio, a viewer follows. P = 0.01 → +0.04.
07	Earn shares (share sheet)	Gain2.0	`ShareWeight` = 2.0.	A viewer uses the share sheet. P = 0.02 → +0.04.
08	Earn reposts (in-network)	Gain1.0	`RetweetWeight` = 1.0. Only in-network, because out-of-network reposts are filtered.	A follower reposts you. P = 0.05 → +0.05.
09	Earn favorites	Gain0.5	`FavoriteWeight` = 0.5.	A viewer likes the post. P = 0.10 → +0.05.
10	Earn clicks	Gain0.3	`ClickWeight` = 0.3.	A viewer taps your link. P = 0.08 → +0.024.
11	Earn link opens	Gain0.2	`OpenLinkWeight` = 0.2.	A viewer opens the link card. P = 0.05 → +0.01.
12	Earn video opens	Gain0.07	`VideoOpenWeight` = 0.07.	A viewer taps to open your video. P = 0.10 → +0.007.
13	Earn photo expands	Gain0.05	`PhotoExpandWeight` = 0.05.	A viewer expands your image. P = 0.20 → +0.01.
14	Earn dwell and click-dwell time	Gain0.05 dwell · 0.4 click-dwell	`DwellWeight` = 0.05, `ContClickDwellTimeWeight` = 0.4, `ContDwellTimeWeight` = 0.004 (multiplied by predicted time).	A viewer clicks in and reads. Predicted click-dwell time 2 (model units) × 0.4 = +0.80.
15	Earn quoted-post clicks	Gain0.05	`QuotedClickWeight` = 0.05.	A viewer clicks the post you quoted. P = 0.10 → +0.005.
16	Earn in-network posts the viewer hasn't explored	Gain0.02	`PostUnexploredWeight` = 0.02, applied only when `in_network = true`.	A follower who has never engaged with your topic sees your post. P = 0.30 → +0.006.
17	Post original posts for new audiences	No weightNo fixed weight	Phoenix and SimClusters retrieve out-of-network original posts. `OONRetweetReplyFilter` removes out-of-network replies and reposts.	A stranger in France sees your original thread through retrieval. Your reply to someone else's post would be filtered for that stranger.
18	Reply to followers and mutuals	No weightNo fixed weight	In-network replies pass `OONRetweetReplyFilter`. Out-of-network replies don't, unless `EnablePhoenixOonReplies` is on (default false).	Your reply to a follower's post is visible to your followers. Your reply to a stranger's post is dropped for strangers.
19	Reply within 48 hours of the post	Filter	`MAX_POST_AGE` = 48 × 60 × 60 seconds, enforced by `AgeFilter`.	Reply at 3 hours: eligible. Reply to a post from 3 days ago: removed before scoring.
20	Post regularly	No weightNo fixed weight	Thunder keeps recent posts from followed accounts in memory (thunder/, README).	Four posts a week keeps your last 48 hours of posts ready for your followers' in-network feed.

Final report: DON'T

Actions that lower the score or trigger enforcement. Paired by number with the DO table.

#	Action	Outcome	How the code uses it	Example
01	Automate or schedule replies	LabelPost label `SpamHighRecall`, 30 days	enforcement_post.yaml: `fast_reply_spam_post` → `act_add_fast_reply_spam_post_label`, TTL 2,592,000,000 ms.	A script posts 200 replies in an hour. Flagged replies get the post label, which drops them as spam for 30 days.
02	Post LLM-written or templated replies	LabelPost label `RiskyHighVizReply`, 30 days. Account label `SpamHighRecall` if flagged `llm_slop_user`	enforcement_post.yaml: `llm_slop_post` → `RiskyHighVizReply`. enforcement_user.yaml: `llm_slop_user` → `SpamHighRecall`.	"Great point, totally agree!" under many posts is flagged `llm_slop_post`, so the reply is labeled `RiskyHighVizReply`.
03	Post gibberish or filler	LabelPost label `SpamHighRecall`, 30 days	enforcement_post.yaml: `gibberish_post` → `SpamHighRecall`.	A reply of "asdf asdf 🔥🔥" is flagged `gibberish_post` and gets the post label.
04	Repeat near-identical replies across threads	ClassifierClassifier flag (threshold not published)	grox/flows/reply_spam/classifier_multi_step_reply_spam.py. It drops flags unless the newest reply is flagged.	The same "Totally agree!" posted under 30 threads is flagged by the multi-step classifier.
05	Coordinate engagement with groups or swaps	ClassifierClassifier flag (threshold not published)	grox/flows/reply_spam/classifier_coordinated_spam.py and plan_coordinated_spam.py.	Ten accounts that reply to each other within minutes match the coordinated-spam pattern.
06	Burst many replies in minutes	LabelPost label `SpamHighRecall`, 30 days	enforcement_post.yaml: `fast_reply_spam_post` (same rule as the automation row).	Fifteen replies in two minutes, by hand, can still trigger the fast-reply rule.
07	Trigger a block	Loss−31.2	`BlockAuthorWeight` = −31.2 (home-mixer/params/param.rs).	A viewer blocks you. P = 0.005 → 0.005 × −31.2 = −0.16.
08	Trigger a mute	Loss−58.8	`MuteAuthorWeight` = −58.8.	A viewer mutes you. P = 0.02 → −1.18.
09	Trigger a report	Loss−234	`ReportWeight` = −234.0.	A viewer reports you. P = 0.01 → −2.34.
10	Trigger "not interested"	Loss−47.52	`NotInterestedWeight` = −47.52.	A viewer taps not interested. P = 0.03 → −1.43.
11	Lose dwell time (not dwelled)	Loss−0.02	`NotDwelledWeight` = −0.02.	A viewer scrolls past in under a second. P = 0.50 → −0.01.
12	Post spam or malicious links	DropDropped	visibility-filtering/rules/registry.rs: `malicious_url/drop/undesirable` and `spam_high_recall/drop/undesirable`.	A link to a flagged domain removes the post from feeds.
13	Run multiple or inauthentic accounts	DropTemporary suspension (`PlatformManipulation`)	enforcement_user.yaml: `act_inauthentic_detection_v45_suspend` → `act_suspend_user`, `perm: false`.	Two accounts that only like and reply to each other trip inauthentic detection and get suspended.
14	Keep the same pattern after a reach drop	LabelLabels last 30 days regardless	All label rules use `ttl_msec: 2592000000` (30 days).	Changing your pattern now doesn't shorten a label that's already set. Wait out the 30 days.
15	Expect replies to reach strangers	FilterFiltered for out-of-network viewers	`OONRetweetReplyFilter` with `EnablePhoenixOonReplies` = false.	Your reply to a follower's post is removed from a stranger's For You feed.
16	Expect reposts to reach strangers	FilterFiltered for out-of-network viewers	`OONRetweetReplyFilter`: `is_retweet && in_network == false` is removed.	Your repost of someone you follow is removed from a stranger's feed.
17	Count on video views	No weightAdds 0 through VQV	`VqvWeight` = 0.0 and `QuotedVqvWeight` = 0.0.	A 40-second video with 10,000 plays adds nothing through VQV.
18	Count on profile clicks	No weightAdds 0 directly	`ProfileClickWeight` = 0.0. Only the follow action (4.0) counts.	500 profile visits add 0. One follow adds 4.0 × P(follow).
19	Post into a viewer's muted keywords	FilterRemoved for that viewer	`ViewerMutedKeywordFilter`.	A viewer muted "crypto". Your crypto post is removed for that viewer.
20	Post a 4th or later item from you in one feed	LossScore × 0.344 at 4th, floor × 0.25	vm-ranker/params.rs: `EnableAuthorDiversity` = true, decay 0.5, floor 0.25.	Four of your posts rank for one viewer. The 4th is multiplied by 0.344 before the final sort.

Final report: worked examples

Illustrative probabilities, not measured values.

#	Scenario	Calculation	Result	What it means
01	Largest single gain: boosted reply on your original post	`20 × P(reply) = 20 × 0.10`	+2.00	The biggest positive single term.
02	Large gain: copy-link share	`20 × P(share) = 20 × 0.05`	+1.00	Copy-link shares are the largest non-reply gain.
03	Largest single loss: report	`−234 × P(report) = −234 × 0.01`	−2.34	The biggest negative single term.
04	Break-even report probability	`2.00 ÷ 234`	P ≈ 0.85%	About one report-likely viewer in 117 cancels one boosted reply.
05	Mute vs. favorite	`−58.8 × 0.02 vs +0.5 × 0.30`	−1.18 vs +0.15	One mute costs about 8 favorites.
06	Combined post	`2.00 (boosted reply) + 0.15 (favorite) + 0.04 (follow) − 1.18 (mute)`	+1.01	Positive and negative terms add up before the final sort.
07	Out-of-network penalty	`2.00 × 0.75 (OonWeightFactor)`	1.50	Out-of-network posts keep 75% of their score.
08	4th post from you in one feed	`2.00 × 0.344`	0.69	Multiplier from `decay^k`, with k = 3.
09	10th post from you in one feed	`0.25 + 0.75 × 0.5^9`	0.2515	Approaches the 0.25 floor and never reaches zero.
10	New-account out-of-network penalty	`0.00001 (NewUserOonWeightFactor)`	Off in repo	Only applies when NewUserAgeThresholdSecs > 0. The repo default is 0.