shadow-ledger

Context

No real order, no broker, no key: just public prices and fictional money. Every decision (buy, sell, do nothing) is written to a journal you can only add to. The engine can then answer “what if…?”: how does this decision compare to its alternative a few minutes, a day, a week later, and now?

The project is inspired by a rules engine with an audit journal and a survival budget, without reusing its code.

What you can do with it

• Replay the past with strategies that never see the future, pay their fees, and are judged on a held-out period.
• Follow a real market live with a fleet of about thirty portfolios in a single process, which can be resumed after a stop.
• Let an agent play, including a Claude Code session on my usual plan, with no billed API.
• Sweep a rule across 37 archived trading days, discarding unusable market days instead of counting them as zero.
• Follow it all in a Vue dashboard that updates live.

The guarantees, checked in the code

• A journal that can’t be rewritten: SQL triggers refuse any update or deletion, and the clock never goes backwards. You can’t backdate a decision.
• No future: a strategy only sees two objects with private fields, and execution happens at the first strictly later price, with a dedicated test. It’s an interface guarantee, not a sandbox, and the README says so.
• Bounded agents: two tools, caps enforced by the code and a survival budget.
• Safe concurrent writes: immediate transactions, tested with two processes.

An honest result

I found no trading edge. This project is for measuring an idea honestly, not for making money.

The tests

264 tests across 20 files.