Practice & capstones

Every level ends with a capstone. Do it for real — produce the artifact, not a mental sketch. Grade yourself against the rubric, or hand the artifact and rubric to someone skeptical.

Level 1 capstone — The measurement report

Gate: measure a trade log honestly.

Brief. Take two broker statements — your own demo/live history, or any two real trade logs you can export (a demo account you run for two weeks generates one). Produce a measurement report: R-multiple distribution, expectancy, profit factor, max drawdown (depth and duration), win rate vs payoff ratio, and the gross-vs-net gap using your Lesson 1.5 cost model.

Deliverable. One page of numbers plus ~200 words of interpretation, and the pandas script that computed them.

Not yet	Metrics computed on gross numbers, or win rate treated as the headline; no cost itemization.
Passing	All metrics correct and net of itemized costs; drawdown reported with duration; the interpretation says what the sample size does and doesn’t allow you to conclude.
Strong	Passing, plus: per-session cost breakdown, a variance statement (what range of outcomes this expectancy and trade count is consistent with), and one identified measurement trap you avoided (e.g. floating losers, cropped history).

Level 2 capstone — The pre-registered verdict

Gate: take one strategy to a pre-registered pass/fail verdict on net costs.

Brief. Formalize the London Breakout from a plain-language description into a deterministic rule spec (Lesson 2.2). Write the pre-registration before testing (Lesson 2.8): hypothesis, deciding metric, required sample size, pass threshold. Backtest event-driven with your full cost model (Lesson 2.4). Walk it forward (Lesson 2.7). Compare against a dumb baseline (random entries, same sizing and costs). Publish the verdict you pre-registered — whichever way it comes out.

Deliverable. The rule spec, the dated pre-registration, the backtest code, and a one-page verdict.

Not yet	Rules tuned after seeing results; costs missing or flat-rate; verdict hedged (“promising, needs more work”) instead of matching the pre-registered criteria.
Passing	Spec is deterministic (two people would get identical trades); pre-registration written first and honored; net costs modeled by session; in-sample and out-of-sample cleanly separated; verdict states pass/fail against the pre-registered threshold.
Strong	Passing, plus: walk-forward efficiency reported; the dumb baseline beaten or the failure against it stated plainly; a list of every degree of freedom you spent, and what a deflated view of the result looks like.

Expected outcome, stated up front: a simple session breakout on retail CFD costs usually does not survive. If your verdict is “no edge after costs,” and your process was clean, you passed.

Level 3 capstone — The attribution report

Gate: run a system with breakers, a ledger, and pre-registered kill criteria.

Brief. Run any rule-based strategy (the Level 2 spec is fine, even if its verdict was negative — this capstone tests operations, not alpha) on a demo account for 4+ weeks through your MT5+Python stack. Circuit breakers on (daily-loss lock, anomaly halt). Broker-side stop on every position. Attribution ledger logging every candidate: state snapshot, decision, expected vs actual fill, itemized costs, R-multiple. Kill criteria written and dated before the first trade.

Deliverable. The ledger (CSV/DataFrame), the breaker code, the signed one-page operator contract, and a report a skeptic would accept: what ran, what it cost, how fills compared to expectation, whether any kill criterion tripped, and one thing the ledger caught that the equity curve alone would have hidden.

Not yet	Positions without broker-side stops; gaps in the ledger; a manual intervention that isn’t logged; kill criteria written (or edited) after drawdown started.
Passing	Complete ledger for every candidate including skipped/blocked ones; breakers demonstrably fire in a forced test; restart-reconciliation shown to work once; fill-quality series reported.
Strong	Passing, plus: slippage distribution vs the cost model’s assumption compared explicitly; a documented breaker or kill-criterion event handled per contract; a written post-mortem in the change-log format.

After the capstones

You now hold the three artifacts that make anyone’s trading claims testable — including your own: a measurement report, a pre-registered verdict, and an attribution ledger. The loop from here is the improvement engine (Lesson 3.6): one change at a time, pre-registered, out-of-sample, with symmetric demotion. Run it as long as the system earns its place.