Fork status · 2026-09-26

What this tree actually is.

The live lab is the contract and the simulation. The Rust prior lives on a public fork of xai-org/x-algorithm. Neither one changes For You on X. Both flags default off, and the event source in the fork is empty.

Where	Pointer
Upstream	xai-org/x-algorithm @4c5cfe8
Fork	TheRetardedElon/x-algorithm---TheRetardedElon
Current branch	pulse/causal-q-4c5cfe8
Pull request	PR #3 (draft)
Writeup on the branch	README

The contract

score' = score · q_t

Phoenix writes the weighted sum. PULSE may only multiply the final score.

q_t = 1 if the rank flag is off q_t = 1 if candidate.pulse is missing q_t = 1 if σ is missing, NaN, or > 0.18 q_t = q(·) only if the posterior is narrow

weighted_score is never rewritten.
Favorite, reply, and view counts are not inputs.
Reactions at or after scoring time are dropped.
The viewer’s own click stream is the wrong shape for the event source. The source is reactions on the post and past decisions by the author being ranked.

What may narrow σ

Class	Meaning	Narrows σ in this slice?
human	Independently acting person	No
mixed	Human plus tools	No
agent	Declared model or automation	Yes, σ = 0.08
service	Declared service account	Yes, σ = 0.08
coordinated	Coupled actors, not a verdict	No

Unlabeled posteriors are floored at σ = 0.22, above the action gate. A high coordination residual is a measurement. It is not a rank change. Newsrooms, sports, and fan accounts look synchronized. Treating that as guilt would punish real humans. The lab uses the same gate.

Honest status

Item	State
main vs upstream	Synced. PR #2 merged 4c5cfe8 into this fork’s main.
PULSE on by default	No. Both flags default false.
Event source	EmptyPulseEventSource returns empty windows.
Changes live For You	No.
Can q_t move a score in the Rust tree?	Only if flags are on and narrow evidence is attached. Declared agent or service is the only class that does that today.
Old RankingScorer	Deleted upstream. The hook moved.
This lab	Documentation and a causal simulation. Not the Rust tree.

Branches

Name	What it is
xai-org/x-algorithm main	Public snapshot at 4c5cfe8, 2026-09-26.
fork main	That snapshot, plus a short landing README. No PULSE mixer files.
pulse/causal-q	Original five commits on the Sept 1 tree. Hooks the deleted RankingScorer. Superseded. PR #1 is the record. Do not patch it.
pulse/causal-q-4c5cfe8	Current work. PR #3. Hooks PhoenixScoresRankingScorer and VMRanker.

Between Sept 1 and Sept 26 upstream deleted home-mixer/scorers/ranking_scorer.rs and moved the published sum into the value model. Replaying the old commits would conflict on a file that no longer exists. The estimator moved forward. The hook moved with the score.

Where it sits

reactions already arrived + author decisions already observed │ ▼ PulseEventSource (empty) │ ▼ viewer_influence_mask drop actors in Bᵤ ∪ Mᵤ │ ▼ evidence → PulseEvidence │ ▼ candidate.pulse ┌────┴────┐ ▼ ▼ value-model scorer VMRanker score = weighted·qₜ score = score·qₜ

weighted_score stays untouched on both paths. A source error fails open: pulse stays unset, so q_t is 1.

Two systems, not one

PULSE asks whether the lagged field looks like independent humans or a coupled controller. That becomes evidence, a width, and maybe q_t.

Reject Closure is preference isolation. If viewer u blocks actor a at time t, then later:

Candidate(a, u) = 0 already true for For You authors Suggestion(a, u) = 0 product contract, other surfaces Influence(a → u) = 0 viewer mask on the lagged field Pᵤ(E, t) = P(E, t) \ Bᵤ

A blocked account is not labeled a bot. They are removed from this viewer’s field. The global source still returns every arrived reaction. The hydrator subtracts.

Surface	In the public repo	After a block
For You author	Yes	Hard drop (AuthorSocialgraphFilter)
For You lagged field	Yes, on the PULSE branch	viewer_influence_mask
Posts from people you simply don’t follow	Yes	Kept. Out-of-network is a discount, not a ban.
@ typeahead	No	Same exclusion set should apply. Compose is another service.
Who to Follow	No	Same
Search	No	Same

Flags

Flag	Default	Effect
rust_home_mixer_enable_pulse_prior	false	Allow q_t ≠ 1 when evidence is narrow.
rust_home_mixer_enable_pulse_hydrator	false	Run PulseHydrator.

File map

Path	Role
home-mixer/models/pulse.rs	Evidence, causal_q, sigma policy, lagged field, viewer mask
home-mixer/candidate_hydrators/pulse_hydrator.rs	Event-source trait, empty source, mask
home-mixer/scorers/phoenix_scores_ranking_scorer.rs	score = weighted · q_t
home-mixer/scorers/vm_ranker.rs	apply_pulse after remote or local score
docs/pulse.md	Mixer contract
docs/reject-closure.md	Block, mute, suggestion contract

The name counterfactual_independence overclaims. On this branch it is residual predictive lift of a lagged cluster term, not a randomized experiment. The fuller time-split estimator still lives on the superseded Sept 1 branch.

Still missing

Wire the candidate field, the two flags, and the empty hydrator.
A real event source: per-post reactions with timestamps before t, plus optional author decision rows. Not the viewer’s click stream. Not lifetime counts.
An offline falsification harness before unlabeled evidence is allowed to narrow σ. Measure the false-positive rate on newsrooms, sports, and fan communities. The falsify page is that contract.
Reject Closure on @ typeahead, Who to Follow, and search. Those services are not in the repository.
The click-then-block sequence, labeled as a hypothesis.

What this lab is for

The miniature feed uses the same gate as the Rust prior: wide σ stays at q_t = 1, a viral swarm updates the EKG and does not move the slate, a declared agent can. It does not compile the mixer, and it does not model @ suggestions. Read the algorithm page for the published weights, then open the lab.