---
name: shortlist-pipeline
description: Ben's requested three step runner pipeline (3 Oct 2026): wide automatic short list, automatic deep dive per coin, Telegram alert only on STRONG BUY; built as /opt/legwatch (second leg / base window), results of the honest census test. Read before building or changing it, or any "short list", "auto deep dive", "second window" or buy alert pipeline work.
sources: [cowork]
aliases: [short list pipeline, auto deep dive, runner pipeline, shortlist to alert, legwatch, second leg watch]
---
- [stated] 3 Oct 2026: wants a process that automatically builds a short list of potentially good buys (fed by his strong Market Signals alerts), runs a project deep dive on each, and sends a Telegram alert only when the deep dive says STRONG BUY right now; otherwise the coin is ignored.
- [stated] The short list should err on the side of being longer so nothing good is missed; the real vetting happens in the deep dive step.
- [stated] Goal: catch multimillion dollar runners as early as possible; design should draw on all Pro Trading project knowledge.
- [stated] Deep dives must include fomo (he called skipping it lazy).
- [stated] 3 Oct 2026 night, clarified: two windows. Window 1 = the first run-up, seconds (trenches process). Window 2 = a slower window with time for a full deep dive (website, technology, holder growth, more); asked me to find the best timing (30 min, 1h, 12h, 1d, 3d). Coins typically go over $1M, pull back, base, next leg, base, next leg.
- [stated] Same night: test honestly on far more coins, this cycle only, Solana and Robinhood separately, meme vs project if useful; include hourly X crowd via the shared spam filter; check whether chatter stays high through the pullback; model case Artificial Inu (multiple pullbacks, strong sentiment, ran ~$9M to $250M+).

# Built (4 Oct 2026, /opt/legwatch on the droplet)
- legwatch.py, cron */10: base = first leg $1M+ (volume backed), -50% from peak, 12h no new low; re-arms after a 30% deeper low. Short list ~185 coins, public shortlist.json.
- ddcollect.py: X crowd check per new base (peak day vs 24h before base), verdict AVOID (crowd left) or WATCH. No STRONG BUY alerts: the census showed no buy edge at the base.
- verdict.py sends to /opt/legwatch/tg.json (a "Buy Alert" bot) when an edge is proven; that bot does not exist yet (needs Ben to make it in BotFather).
- [stated] 4 Oct 2026: watchdogs must be multiple redundant processes, both redundant monitoring and redundant fixing, fast; notify Ben only if several layers fail several times.
- Watchdog layers since 4 Oct 2026 09:16 BKK: (1) legwatch.timer every 10 min runs /opt/legwatch/run.sh (cron line removed) + super.py checks legwatch.ok (30 min) and ddcollect.ok (60 min) and the timer; (2) stale.py cron */5 re-enables the timer and runs run.sh directly, alerts via alertgate only after 45 min stale and 3 failed fixes; (3) livewatch2 (20 s) starts legwatch.service when legwatch.ok is over 25 min old; guardkit (15 min) rebuilds the timer and the stale.py cron line. Fire drill: each layer alone fixed a broken run in 14 to 25 s, no message sent.

# Honest result (4 Oct 2026, details in [[trading-edge-hub-2]])
- Window 2 timing: base reached a median 1.5 to 3 days after the first $1M; waiting for a 72h base is worse.
- No base feature beats the power-law odds (cap m reaches M about m/M); crowd left is a strong AVOID (0.4x); crowd held only closes part of the gap (0.8 to 0.9x).