[NielsRogge]

Follow-up to #867. The trending score now treats X like velocity as a lift, not a linear blend:

score = 100 × max(0.5·G + 0.5·H, w_x · X)

Why

A read-only preview on 2026-11-03 used the 12:50 production components at the agreed X weight of ⅓. Only 29 of about 2,200 scored papers carry X data. The linear blend capped every other paper at 66.7, so X-tracked papers took ranks 1–5, CaPtain at 3 likes/h among them (it jumped from rank 119 to rank 4). Under lift-only, papers without X keep their score, and only truly viral posts break in. Sparks of Weight-Space Programming at 3,000 likes/h goes from rank 265 to first.

Changes

• composite_score computes max(GitHub/HF blend, x_weight × X). GitHub and HF keep 0.5 each.
• PWC_TRENDING_X_WEIGHT now scales the lift and accepts [0, 1], where 1 is the full lift. It stays 0 (off) in production until the Nov 5 comparison. The env key and its plumbing are unchanged.
• Updated the docstrings, the CRON README and the pwc-cron-health probe.

Validation

• pytest backend/tests/test_trending_formula.py backend/tests/paper/test_trending_scores.py backend/tests/deployment/test_cron_environment.py backend/tests/deployment/test_host_env_credentials.py: 84 passed, 1 failed. The failure is the link_google_scholar_profiles 3h timeout contract, which already fails on main.
• ruff check / ruff format --check are clean.

With the weight at 0, this changes no live score.

🤖 Generated with Claude Code

---

Note

Medium Risk
Changes how paper trending ranks behave once X weight is above zero, affecting the public trending feed; at default weight 0, scores are unchanged.

Overview
Paper trending no longer linearly blends X like velocity with GitHub and Hugging Face. The composite score is now 100 × max(0.5·G + 0.5·H, w_x · X), so papers without linked X posts keep their GitHub/HF ranking, and only a strong X rank can raise the score (including to the top when w_x = 1).

PWC_TRENDING_X_WEIGHT scales that lift and is validated on [0, 1] ( 1 is allowed; previously rejected). Default 0 means no live behavior change until production turns X on.

Docs for the cron README and refresh_x_post_likes probe describe lift-only semantics; unit and cron tests assert that X never lowers a score and that weight 1 is accepted.

^{Reviewed by Cursor Bugbot for commit c534897. Bugbot is set up for automated code reviews on this repo. Configure here.}

[NielsRogge]

Status 2026-11-03: live in production (64e78dd). PWC_TRENDING_X_WEIGHT is still unset (0), so no ranking has changed.

Still open:

1. About Nov 5: send the user a read-only top 30 comparison, the current ranking against weight 1 (score = 100·max(0.5G+0.5H, X)).
2. On the user's approval only: set PWC_TRENDING_X_WEIGHT=1 in /etc/paperswithcode/trending-cron.env, then check the next calculate_paper_trending_scores run.
3. Then a frontend PR: remove the "trending (X)" tab, redirect /x-trending to /, and remove GET /papers/x-trending. Keep the admin PUT and the paper_x_trending table.

[NielsRogge]

Done 2026-11-05: the user approved weight 1, and PWC_TRENDING_X_WEIGHT=1 is set in production. The 08:45 UTC score run put paper 131739 first on the homepage. The trending (X) tab was removed in #885 (live in production as d7e7c1c). Nothing remains open.