Current implementation and explicit deviations are in the ROADMAP and ADRs 0011–0013. Local technical acceptance is distinct from real OAuth/broker, Railway publication and prolonged financial observation.
Status: proposed. Dataset, frequency, benchmark and costs pending.
Contract of one run#
Input: strategy version, instruments, interval and calendar, immutable dataset/hash, seed when applicable, capital and currency, fill model, fees, spread, slippage, latency and risk policy. Store the code commit/hash and runtime version. Output: signals, risk decisions, simulated orders/fills, equity curve, metrics and a limitations report.
Use the same strategy/risk contract as the worker, with simulated clock and broker. Replay runs isolated from execution and has no live credentials. Durable asynchronous job, states queued/running/completed/failed/canceled, CPU/memory budget and the ability to cancel. Do not block the HTTP request while testing.
Simulation#
In the proposed default, a closed candle generates an executable signal only at the next eligible event. Fills respect order type, tick, lot, costs and modeled liquidity. OHLC does not reveal the intrabar path: if stop and target fit in the same bar, use an explicit conservative rule or finer data; never silently choose the best result. Gaps may execute beyond the stop, with the assumption recorded.
Costs and taxes are not a universal constant. Parameterize operational costs; the tax estimate is separate, if implemented, with reviewed rules. Compare slippage/liquidity scenarios and report sensitivity. Failure or absence of history makes the run fail or explicitly marks the interval as incomplete.
Report and acceptance#
Reproduce the same curve/fills within a documented tolerance with the same inputs. Identify before/after costs, benchmark, drawdown, sample size and temporal partitions. No tuning on the final test. Compare with simple manual fixtures and test zero trades, consecutive losses and insufficient equity. Internal paper and broker paper are separate results, with their own assumptions. Go is the initially recommended runtime; optional Python exchanges versioned artifacts, without changing the production strategy without a new equivalence test.