Trading Bot.

I studied financial markets and I understood them. Trading them profitably was another thing. I'd watch a stock I'd researched all week drop sharply on a single announcement and sell within five minutes, then watch it recover by close. After a red day I'd jump into something I hadn't properly analysed, just to make the loss back.

The analysis was usually fine. It was the real-time decisions that killed it. I'd override my own rules the moment I felt any pressure. So I set out to keep the analysis and take myself out of the execution.

Every weekday at 10:35 GMT, thirty minutes after London open once the opening noise has settled, the system runs. APScheduler fires five jobs through the day. Nothing is held overnight.

1. 10:30Start 5-minute candle collection via Lightstreamer
2. 10:35Full four-agent evaluation across its FTSE watchlist. The only entry window of the day
3. 12:00Regime recheck: if the market fails the regime gate, close everything
4. 14:00Second regime recheck
5. 16:00Liquidate everything. No overnight gap risk, ever

First thing each session: kill-switch and market regime check. If either fails, nothing else runs. News gets scanned and classified by tier before the technical check: EMA, VWAP, RSI. Only setups that pass everything reach the quality gate, where Claude Sonnet rates them against a strict JSON schema. The last word belongs to Claude Opus as risk auditor. If it vetoes, the trade dies. The rails live in code, not prompts: the agents can assess and recommend, they cannot override the config.

• Structured outputs, no fallback holesThe quality gate used to run on a model whose JSON sometimes failed to parse, and the fallback defaulted the score to 3: a safety hole that could let marginal trades through. Switched to Claude Sonnet with a guaranteed schema. No parse failures, no fallback needed. Prompt caching across parallel scans cut token spend by about 90%.
• News before numbersEach ticker's RNS gets checked before the technical scan. One API call finds same-day announcements and classifies them by tier. This replaced a screen scraper that kept breaking; the scraper survives only as the fallback.
• Walk-forward backtesting60-day default, up to two years. Evaluates at 10:35 on each historical day with no lookahead bias and a 0.2% round-trip spread cost deducted.
• Atomic state, full audit trailA file-locked state manager prevents race conditions between the order manager and the kill switch. Every decision writes to a JSONL audit trail: timestamp, session, stage, ticker, outcome. The whole system is one long paper trail, which is how I like my systems.