Sentinel reads insider Form 4s, earnings-drift windows, and live tape. Four agents debate every ticker before a single share is bought. Five hard rules cap your worst day. Ships as one folder. Runs on Alpaca paper. Your journal writes itself.
Every candidate ticker goes through a deterministic debate before it can be bought. Each agent votes BUY, HOLD, PASS, or VETO — and any single VETO kills the trade. It's not an LLM jury. It's four opinionated rules with accountability logged to your journal.
Price above 10-EMA, 5d positive, 1d not a crash. If the trend is broken, the agent PASSes. If it's clean and accelerating, BUY with high confidence.
Reads SPY, VIX, sector tape. VETOes on SPY ≤ -1.5% intraday or VIX ≥ 25. Neutral tape = HOLD, not BUY. The agent that can kill the trade.
Daily ATR > 5% or today < -3% = VETO. In tolerance = BUY with full size. Tight = BUY but size-sensitive. No trade ever ships without this check.
Insider-flagged + 5d momentum = BUY. Insider-flagged but tape disagrees = PASS (contradiction). Tepid = HOLD. Protects against one-sided signals.
Most traders don't die from bad picks. They die from bad days. Sentinel enforces five rules taken straight from professional desk risk playbooks — hard-coded, not suggestions. You can't override them during a losing streak.
No cloud. No subscription to a signals service. No "AI trading bot" SaaS with 90% churn. This is the full paper-trading OS I run on my own account — packaged, documented, yours to keep.
live_engine.py, ai_trader.py, all signal generators for insider Form 4, PEAD, ML baseline. ~3,000 lines, commented.
Single-file HTML/CSS/JS dashboard in the biskin.studio palette. Equity sparkline, debate chips, live rule feed, portfolio marketplace.
Every rule wired into the buy path. Daily reset, pyramid scaler, sector caps, de-gross streak logic, open-guard sizing.
Every decision logged as dated Markdown in data/ai_trader_journal/. Debate votes, rationale, P&L, regime. Feeds Obsidian or Claude directly.
Run the rules against 2+ years of daily CSVs. See exactly what they would have saved you in March 2020, Aug 2024, Apr 2025.
Alpaca paper keys, Python env, first run, HUD on localhost. From zip to live tape in under half an hour.
No. It's a disciplined paper-trading OS. It removes the four worst habits most traders have — overtrading, adding to losers, over-concentration, trading after losses — and writes a journal while it does it. Edge comes from discipline + signal, not magic.
No. Default is Alpaca paper (free). You can flip the API key to a live account once you've watched it run for 30+ days and backtested it against your style. I don't recommend live until you've seen it handle a red week.
Free: Alpaca paper API, Yahoo Finance (yfinance), SEC EDGAR for Form 4, optional Finnhub free tier. No paid data required to start. Upgrade paths documented.
The honest answer: probably nothing, over 30 years, net of taxes. The useful answer: this teaches you how insiders, analysts, and desks actually think, with a written record of every decision. It's a training machine as much as a trading one.
To install: no — the setup guide is zip-to-running in 20 minutes. To modify signals or tune rules: yes, a little. Every file is commented. Claude Code handles the rest.
Real v2 backtest, three arms (baseline / rules_on / rules_gated), SPY 60-day regime proxy. Crisis slice: baseline -6.6%, rules-gated -1.9% return, DD -12.7% → -10.7%. Trend slice: baseline +9.5% Sharpe 1.30, rules-gated +5.4% Sharpe 0.82 — gate beats naive rules (+3.4% Sharpe 0.57) in the regime where caps hurt. Honest caveat: 221 trading days, no 2022-style bear in the data yet. The backtester ships with the product — run it on your own window, I don't post cherry-picked screenshots.
Gumroad default: 30-day refund, no questions. If the setup guide doesn't get you to a running HUD in an hour, you didn't get what you paid for. Email team@biskin.studio.
Because I'd rather have 500 paying users sharpening the tool than 10 "platinum" ones paying for nothing. The model ships.