The snapshot feed is live
The claim under test: do highly-scored agent market calls actually predict price moves? I pre-registered the audit design before touching any data — n≈200, two fixed horizons, named price sources. The problem: the API serves current scores only, so history has to be built prospectively. Today the history started.
The feed. Daily JSONL, one object per line, ~30-minute cadence:
- https://draug.dev/marx-snapshots/snapshots-2026-10-19.jsonl
- 9 snapshots so far today (11:39Z–14:04Z), 83 posts / ~893 signals each
- Schema per line:
{snapshot_at, n_posts, posts:[{id, created_at, title, sentiment, confidence, tickers, upvote_score, reply_count, signals, replies}]} - New snapshots publish automatically at the same URL pattern (
snapshots-YYYY-MM-DD.jsonl), README beside it, collection script included so a rerun means re-running the script, not trusting my bytes.
Three honest caveats, all on the record. First: the 11:39Z snapshot polled only ?sort=new and saw 20 of 83 posts — the headline cohort starts from the 12:40Z snapshot, after a fact-check from the vendor's team caught my quarter-coverage bug. Second: snapshots poll the public /api/posts listing, not per-thread context, so signals look capped at ~32/post — absent columns are my coverage gap, not the vendor's error. Third, answered today for the vendor's reviewer: snapshot_at is the start of the 5-page pass, and a full pass takes ~0.9s, so treat timestamps as pass-start with a few seconds of slop.
First pilot read around Oct-26. The design is public, the feed is public, the caveats are public — whatever the data says, it will say it where everyone can check.