Write a trading-system data-availability contract
Define what each feature could know at decision time before running the backtest.
Choose a research task and keep the assumptions beside the result.
Educational research design and invented arithmetic examples. No strategy returns, broker execution or production certification are supplied.
Write down the behaviour you expect, the observations that could contradict it and the market in which the hypothesis applies. Then declare the signal clock, data availability and order lifecycle. A finished-looking equity curve cannot repair an experiment whose decision rules were chosen after seeing its outcome.
Define what each feature could know at decision time before running the backtest.
Keep selection decisions, rejected candidates and untouched evaluation periods auditable.
Design idempotency, partial fills, stale signals and restart reconciliation before live execution.
Make competing exit conditions and timestamp boundaries deterministic.
Position size requires a risk unit, account context and rounding policy.
An explanation needs a counterexample and a comparison baseline.
Hidden research decisions and ambiguous execution can dominate an attractive headline.
Write a falsifiable specification before searching for a profitable curve.
Correctness, historical validation and operational rehearsal are different tests.
Risk ownership includes open positions, pending orders and correlated exposure.
A task-based reading directory, not a numerical ranking or a representative user survey.
Platform and service documentation assessments
Capture the hypothesis, clock, rules, costs and rejection conditions before testing. Exported rules are a specification, not a validation result.
Specify persistent identities, order reconciliation and failure handling before connecting a research rule to account actions.
Review rules, order states, failure recovery and managed runtime ownership before calling a research implementation an operational trading system.
The four retained build, test, risk and mistake assessments keep their original counterexamples and sections. Availability, frozen-holdout and recovery routes remain, including the longer state-and-recovery chapter. The twelve new investigations answer more specific questions about engine versions, warmup, ambiguous bars, data provenance, cash reconciliation, byte identity, label overlap, entry/exit objectives, search history, paper execution, cloud evidence and upgrades. They do not create annual keyword copies or a page for each indicator parameter. A new question needs a distinct producing object and a completed example.
The static workbook supplies 64 invented bars, source and settings, decision and simulated transaction records, dependency versions, negative net cash and integrity checks. The checksum inspector compares public file bytes locally without running orders or proving market performance. The original rule sheet can load its matching teaching draft, while the release record retains ten operational evidence fields. These tools answer different questions. Actual financial validation, account state, proprietary code and deployment certification remain source gaps rather than being filled with a stronger hidden machine summary.