Roadmap and console: four phases, micro-capital, ship it

[Part 2](/blog/quant-zero-cost-stack) covered running both sides on zero budget. This one closes the series: how it lands, in what order, and why the interface matters. One rule up front: no fully-automated black-box trading on day one. Four phases, each with a hard acceptance bar; miss the bar, no upgrade.

Phase 1: basic quant lab and a real dashboard (the current focus)

The quant module goes standalone. Wire yfinance and CCXT, build the dynamic universe — top 20 US stocks by volume or top 10 crypto by market cap, refreshed daily. Build the backtest and stats kit: win rate, expectancy, profit factor, max drawdown — all of them, no exceptions. The dashboard runs on TradingView Lightweight Charts — open source, millisecond candle rendering, indicator overlays, entry/exit markers. The acceptance bar is plain: any ticker, a backtest with marked trades on screen within a minute.

Phase 2: decision memory and the AI layer

Mount the unified model router. The cloud model digests US-equity news and macro text into Market Context; the local decision model turns technicals plus context into signal suggestions. And stand up the decision-memory table — every day, traded or not, one row: the AI's call and the actual move, labeled with PnL and correct/incorrect. From this day the system accumulates experience.

Phase 3: unattended runs and strategy health

Scheduled evaluations (open, midday, close) produce a daily Strategy Health score — below threshold, the strategy gets a warning and is frozen from trading. Paper trading first: prove the system can run continuously with no real capital at risk. The goal is not 24/7 uptime as a badge — it is thirty incident-free days.

Phase 4: progressive authorization, L0 to L5

Strategy versioning is iron law: a live strategy is never edited in place. The AI may propose upgrades (say, V1.1) based on the memory base, but a new version must first survive out-of-sample testing and paper trading and hit the Sharpe and win-rate thresholds before an admin can promote it into the Active Pool with one click. Authorization is granted level by level: log-only, suggest-only, paper, micro-capital live. Every level is earned with data from the previous one.

Why the console matters

Those geek-chic quant terminal screenshots making the rounds — monospace type, paper-toned palette, oversized PnL and Sharpe figures, entry/exit markers on the candles — they do look good. But their real job is not looking good: a carefully designed console is the core tool for strategy explainability and health monitoring. You see exactly when a signal fired, why the AI sized up at that node, where the drawdown lived. It happens to share DNA with my design system: monospace fonts, paper palette, restrained decoration — the root-system illustration on this site's homepage and that terminal aesthetic are the same taste. Later, D3 or G6 will render "news / factors → model reasoning → signal" as a live network graph — the invisible AI thought process, made visible.

Micro-capital, run the chain

After Phase 1, bind an account with a few hundred dollars and place real orders — the point is validating slippage, API latency, and stop-loss behavior, not making money. Every live PnL row and decision log is track record: the hardest backing this system will have, wherever it goes.

That closes the series: positioning, zero-cost start, the landing path. "An AI that funds its own compute" is the far vision. Step one is still the Phase 1 dashboard. End of series.