Terms and evidence boundaries
Educational research design and invented arithmetic examples. No strategy returns, broker execution or production certification are supplied.
Educational research design and invented arithmetic examples. No strategy returns, broker execution or production certification are supplied.
Population and denominator
The population is the set of eligible observations under a declared rule. A denominator describes which of those observations enter a particular rate. Closed trades, all calls and account days are not interchangeable.
Price return and account return
A price move concerns an instrument. Account return depends on capital, exposure, cash flows and costs. Summing unrelated percentages does not produce a compounded portfolio return.
Simulation and execution
A simulation applies a model to observations. An execution record identifies actual orders and fills. A software test can validate logic without supplying an independent strategy-validation sample.
Source and verification
Describe the cause, expected effect and observable alternative explanation. Decide which experiments can reject the idea before searching for parameters. Record every material research choice, including unsuccessful runs; the best final curve is not the whole research population.
Sources and scope
Experiment identity and byte identity
Experiment identity includes data, source, parameters, timing, execution, cost policy, eligible population and selection history. Byte identity is narrower: a digest can show equality of a specified file under the chosen hash algorithm. The same nickname does not make corrected data or a new exit the same experiment. A matching manifest hash cannot prove that a person never inspected a holdout. A source-file revision is also not the original cutoff of an underlying model-performance snapshot. Those meanings should travel with any machine summary or exported research note.
Warmup, decision, intention and fill
Warmup builds the state required for a rule and is not automatically an evaluated outcome period. A decision consumes declared available information. An intention requests an order; a simulated fill follows the engine's policy, while an actual fill needs a broker transaction. The example's first decision is bar 5, not the first CSV row. Its entry intention uses a completed close and fills at the next synthetic open. Its entry-fill bar counts in the three completed holding bars before the exit request. None of those invented timestamps describes an actual market session.
Reference checks and strategy evidence
Prefix invariance means changing later input leaves earlier decisions unchanged under the tested conditions. Cash reconciliation means declared fills and costs reproduce an amount. A software reference fixture means deliberately constructed test data and expected behaviour. Economic validation needs a different producing market population, evaluation design and independent evidence. An exposed holdout remains exposed even if a new filename or hash is created. A negative software replay is not a verified no-model master, and a passing replay is not production promotion.
Operational identity and authority
Research state, source-code state, registry gate, deployed revision, subscriber route and certification are separate records. A deployment receipt should identify actual cloud execution and durable output, not only code existence. A kill control may stop new entries while leaving pending orders or held positions, depending on its stated authority. Restart reconciliation needs authoritative account state, not an empty local cache. The website's checksum comparison and release notes concern its static publication; they neither manage orders nor certify a proprietary trading runtime.