/jp/ — moderation & corpus metrics

Live from chan.db.

Deletion rate by general

Share of harvested posts flagged deleted. The spread is the headline: a model trained pooled across generals will partly just learn which general a post is in.

Toxicity by general

Share of posts with any slur, hostile imperative or profanity, per general.

Does toxicity predict deletion?

Barely. Janitors are not mainly removing hostile language, which means toxicity features will not carry a triage model and the register labels really were the wrong target.

What gets linked

Domains shared across all generals — the clearest read on what the board is actually doing rather than what it argues about.

What each general is about

Most distinctive terms per general by log-odds against the rest of the board, so shared jargon cancels and only what separates them survives.

Brand sentiment — is anyone being shilled?

Positive:negative mention ratio per brand. Every brand on the board is net negative (all ratios below 1.0). A successful astroturf campaign produces a positivity skew; none exists here. The board's own shill-accusation rate is shown alongside and is under 3% everywhere.

Rising and falling interests

Log-odds shift between the first and second half of the harvested window. Blue rising, red falling.

Activity heatmap

Posts by weekday and UTC hour across all generals. Sequential single-hue ramp, scaled from zero.

What deleted posts look like

Deleted vs kept, on structural features only. Being an OP is by far the strongest single signal — 3.8x — which is why the triage model leans structural rather than lexical.

Images and questions by general

Share of posts carrying an image, and share phrased as a question. /sdg/ is three-quarters images; /sqt/ is the board's help desk.

Which linked domains get deleted

Deletion rate for posts linking each domain, against a 1.36% board baseline. Commercial AI vendor and service links sit well above it — the closest thing to a promotion signal in the data.

Repeated link drops

Identical URLs posted many times. Most are general OP boilerplate — standing resource lists copy-pasted into every new thread — not advertising. Sample text is shown so the two can be told apart.

Time from post to deletion

The <1m bucket is near-empty not because fast deletions are rare but because the archiver never sees them. This is the truncation, measured.

Board composition — threads per general

From the complete 3-month index of 148 threads. Generals are 11% of all threads started on /g/.

Threads started per week

Whole board, index only. Flat throughput — /g/ starts roughly a thousand threads a week.

OP vs reply deletion

OPs are removed at several times the rate of replies, consistent with duplicate and off-topic thread removal being a large share of janitor action.

Corpus volume by general

Harvested posts, generals with ≥500 posts. This is what the model trains on.