Post Virality

@0xMovez's post (source post, 2026-09-24) described a "viral post prediction analyser":

• Jev parses 800 viral posts;
• it groups them by hook type (build demo, receipt, launch, contrarian…);
• it runs 12 typed checks on every post (hook, numbers, media, CTA);
• a new post is scored against its hook type and its 5 most similar viral posts.

This is a small, honest version with a held-out test of whether the score ranks higher-engagement posts higher.

• Data: 84 public posts. 78 are the posts behind X links the human has sent (saved link-index, the Sep 24 batch, and links in our research notes) and 6 are posts quoted inside them.
  • Fetched read-only through `, one request about every second, with no login (fetch.py`).
  • Median 199 likes, max 94,637.
  • This is not a random sample of X. It's what one person found worth sharing, mostly AI and dev posts.
• 12 checks:
  • 5 in code: has a number, has media, has a link, list format, length bucket. Code can see these for sure.
  • 7 model Nouls, one request per post: strong hook, concrete result, shows how, novelty, call to action, names a known product, hype language.
  • 1 model Choice: hook type (build demo, receipt, launch, contrarian, tutorial or list, news or commentary, other).
• The model sees only the post text: no counts, no author, no follower numbers.
• Scorer (code): the mean of three predictions in log-likes space, learned from the training folds only: (i) the hook group's mean, (ii) the mean of the 5 nearest posts in check space, and (iii) a per-check lift sum. It is reported as a 0–100 percentile against the training posts.
• Explanation: a short template (hook type, checks passed, 3 nearest posts, score), not a second model.
• Evaluation: 5-fold cross-validation, so every post is scored by a scorer that never saw it. We report the Spearman rank correlation with log(likes) and a bootstrap 95% CI, against three baselines: followers only, code checks only, and a shuffled control.

Why decomposing helps. Each check is a separate yes/no you can inspect, and the scorer is plain arithmetic. When the result is weak, you can see which part is weak. Nothing is a black-box "virality score".

Spearman rank correlation with log(likes) on held-out folds, 84 posts, with 95% bootstrap CI:

What it shows:

• The result is weak. The model-check scorer has a positive rank correlation of about 0.26, and its CI only just clears zero.
• The shuffled control reached +0.19 by chance, which shows how noisy 84 posts are.
• The code-only checks and follower count carried no signal on this set.
• So the model's judgments (hook, concrete result, novelty…) are the only part with any signal. But this is not a virality predictor anyone should trust yet.
• What would make it meaningful: a baseline of several hundred posts from one niche, drawn systematically (not hand-shared), with engagement normalised by audience and age.
• openjev hook types: {'launch': 17, 'build_demo': 19, 'news_or_commentary': 21, 'other': 5, 'tutorial_or_list': 13, 'receipt': 9}.

REAL JEV (OpenCode Zen free, jev-1.13-free): not run on real Jev: exceeds free-tier budget. No retry is scheduled. The saved results contain the openjev pass; its measured sample and interval are generated above. Free-route evaluation is reserved for the frozen held-out projects.

• openjev: 84 requests (one per post), 64,475 input tokens (about 767 per post), median 362 ms, $0.
• At list price a 10,000-post baseline is about 7.7M tokens, roughly $0.32.

• Calibration. Better-calibrated checks give cleaner check vectors and so better neighbours. Data size and selection limit this project more than the model does.
• The scorer is deliberately dumb. With a real baseline, it should be refit per niche.

Inspired by a published viral-post predictor post (2026-09-24): claims unverified. The numbers on this page are ours, from our own held-out test.

• @0xMovez (the post above): the 12-typed-checks and hook-group design.
• Latent Space: "dark data" is the biggest money maker he names, and "evaluate on your own data, not public benchmarks" is why this project has a held-out test.
• building-with-jev (dbreunig): one property per question, and counts in code (numbers, media, links, length are code checks).

cd samples/post-virality && ./run.sh      # = virality.py fixtures/posts.jsonl   (fetch.py refreshes the posts)