You send a return chart to a friend. They ask: “Which data version? What costs? Can I run it again?” If the answers live only in an old conversation, the handoff becomes a memory test. A screenshot may establish that someone saw a result without showing how to reconstruct it.
Start instead with a small package whose recipient can check key accounts without guessing inputs or editing code. Deliberately remove one input, observe an honest stop, then restore it. A documented failure can make a package easier to take over than a summary saying everything worked.
Define the evidence this exercise can provide
You will assemble a minimum checking package, recalculate in a separate directory and locate a failure, then restore the original and preserve a one-condition variant. Begin with a fixed mechanism before extending the process to real data and operating environments. A neat folder is not itself completed research.
The National Academies distinguishes consistent calculations using the same data, methods, and analysis conditions from studies collecting new data for the same question.[1] This chapter primarily practices the former. Reproducing a faulty implementation can reproduce its mistake; G38's independent expectations and G46's evidence limits still matter.
Use C4's five-day synthetic ledger, supplemented by the C3-50 twelve-period package. Neither is real ETF performance. C4 is not shorthand for C2; this exercise changes none of C2's month-end signal, execution, or funding rules.
Make the minimum package answer practical questions
Open the C4 materials, with Chinese supporting notes, and retain instructions, specification, price inputs, a calculable XLSX workbook, fixed expectations, and a blank comparison sheet. Keep three identities separate: original prices supply inputs; a workbook or calculator executes rules; the expected ledger is a comparison target, not newly generated output.
C4 begins with $1,000 cash. A close strictly above $11 targets 20 shares, adjusted at the next listed trading day's open. Each successful order costs $1. Insufficient cash rejects the whole order without a fee. There are no external flows, distributions, slippage, taxes, or modeled settlement delays. Equality with $11 does not target shares.
State acceptance points before running: D3 cash $739, 20 shares, closing equity $959; D4 cash $938, no shares; D5 remains $938. Separate expectations from inputs and rules. Predict first, then inspect actual outputs. Checking already-viewed answers is retrospective verification, not a blind test.
Begin with the ready-made spreadsheet
Save your own baseline copy of the C4 interactive ledger. Confirm B7 initial cash is 1,000, B8 commission is 1, and the five price rows match the original CSV. Preserve formulas. Find D3 and D4 in the calculation area and check cash, shares, fees, and equity individually.
Change only B8 to 2. Expected D3 cash is $738, equity $958, and final equity $936. Save a fee-two variant. Restore the baseline fee of 1 and confirm 938 returns. Close and reopen, repeat the check, and record software name, version, recalculation mode, and actual values seen. “The formulas are present” is a different observation.
If changed inputs produce unchanged outputs, check calculation mode and formula support. Do not type expected answers over formulas. Microsoft distinguishes automatic and manual recalculation.[2] Its documentation is a reason to inspect your software, not proof that this book tested your desktop Excel or WPS.
For comparisons across tools, retain a numerical CSV snapshot and check headers, row order, signed transactions, and decimals. CSV ordinarily preserves displayed text and values rather than complete workbook behavior; Microsoft documents these format losses.[3] Keep both XLSX and CSV. A value snapshot is not a refreshable-formula replacement.
Read the existing author run with its actual scope
The recorded October 29, 2026 author check copied materials into a new temporary directory on Windows 11, then read prices, specifications, and the existing calculator from that copy. It recalculated all five C4 rows against seven fixed expected fields per row, then compared all fields in the calculator CSV and workbook's numerical export. They agreed.
The independent hand check is short: D3 buys 20 shares at $13, spending $260 plus $1 and leaving $739. At the $11 close, shares are worth $220 and equity $959. D4 sells 20 at $10, receiving $200 less $1, leaving $938. The original prices and fees—not an AI interpretation of a chart—explain the result.
The supplemental C3-50 check copied the whole existing package. Its generation and independent verification commands both exited zero in the copy. All 24 period-ledger rows, ten trades, and both rule summaries matched. Quarterly ending equity was approximately $11,053.9253; threshold-rule equity $10,984.3270, preserving the $1,000 external contribution before M7 and every teaching fee.
Those are archived author-side runs, using decimal and rational arithmetic. Environment, dependencies, commands, and outputs remain in the original author record, currently in Chinese. They provide a second numerical route, while your spreadsheet route tests whether you can repeat the work yourself. Neither signs for the other.
A new directory is not a new operating system, fresh dependencies, or another human recipient. That check did not recalculate XLSX inside desktop Excel/WPS, test CoTrading import/run/export, or acquire real long-term C2 data. G38's earlier workbook-engine check is separate evidence; the G60 run compared its exported CSV values, rather than repeating that engine test.
Let missing input invalidate the current run
Next, the recorded C3 copy experiment temporarily moved inputs.csv, leaving other files intact. Its generation command exited 1, reporting the missing input. The previous ledger.csv still existed. That old file did not turn a failed new run into a successful one.
Record the current working directory and command, nonzero exit, missing filename, provenance of old output, and blocked status. The problem is missing material, not deteriorating strategy performance or a reason to choose another model. AI-generated “similar returns” would change the research object instead of restore it.
After the original input was restored to the copy and its identity checked, generation and verification both exited 0 again. Original package files were neither moved nor rewritten. Recovery establishes that the process can complete with restored material; it does not erase the preceding failed run.
Separately, C4's existing calculator changed only its fee parameter from 1 to 2. It reproduced D3 cash 738 and endpoint 936, then returned to 938 with fee 1. Two successful trades each cost another dollar, explaining the two-dollar endpoint difference. Prices, quantities, times, and C2 rules were unchanged. This is a controlled sensitivity, not a search for a prettier winner.
Specify hashes and tolerances before seeing differences
The recorded copy experiment checked hashes before copying and after finishing: all 19 related original files retained identical byte digests. SHA256 is one way to summarize bytes, supported by Python's official interface.[4] It can reveal a changed file, but cannot prove that data are genuine or a conclusion is correct.
Line endings are bytes too. Saving the same price CSV once with LF and once with CRLF produced different byte hashes while all five parsed data rows agreed. State whether your package compares original bytes or a specified normalized text representation such as UTF-8 with unified newlines. Those digest conventions are not interchangeable. XLSX is binary and must not undergo text newline conversion.
Define numerical tolerances separately. C4's integer accounts require exact equality. C3-50 uses 28-digit decimal arithmetic and a verifier tolerance of $1e−18; the recorded exported fields matched original values directly. Finding a one-dollar cash difference and then raising tolerance to two dollars would not establish reproduction. Keep four-decimal display values distinct from full internal values.
Matching endpoints alone is insufficient. Compare transaction time, direction, quantity, commission, cash, and holdings before performance metrics. A first disagreement may identify different models, implementations, numerical precision, or merely column meanings. Trace that first point instead of arguing over charts.
Expand the package without making the recipient repair code
A real project also needs original material or lawful acquisition instructions, availability times and corporate actions, T05 rules and boundaries, settings, dependencies, tool versions, complete output, every attempt, and restoration points. Where redistribution is not permitted, retain acquisition steps, versions, and checkable fields instead of sharing unauthorized raw files.
If the tool generates code or environment files, preserve them with operating instructions and dependencies. The learner need not edit every source line. Distinguish incompatible environments from implementation disagreements: a missing library needs a verifiable environment route; a wrong balance needs intermediate accounts. Repeated source edits until the expected endpoint appears are not a reproducible handoff.
This deterministic arithmetic used no random training, so random seed is not applicable. Later sampling or modeling needs seeds, algorithms, library versions, and data splits. For AI services, save service/model identity, call time, prompts, responses, and tool outputs. If a service does not expose an underlying revision, record “not provided,” not an invented version.
The same prompt does not guarantee the same reply. Nor can a prompt or screenshot substitute for inputs and actual ledgers. An AI claim that you changed a position from anxiety remains an inference without your record or other evidence; do not insert it into decision history. G59's information and action-authorization boundaries still apply.
Diagnose a version mismatch
Explain why reproducing 938 does not establish effectiveness in real markets. Then assess a claimed successful run with missing input but an old ledger still present. Write what evidence to preserve and how the status should change.
The first result establishes only consistency within the supplied synthetic data and implementation. Real data, economic value, and operating conditions are different requirements. The second needs command, directory, exit and error, and provenance of the old result. Mark that run failed; restore the original input and rerun. An old file cannot represent newly completed work.
Now consider a transfer package whose instructions specify $1 per order, but D3 cash is 738, equity 958, and endpoint 936. Prices and quantities match the baseline. Only one settings screenshot shows fee 2. Is reproduction established? Must the error be duplicate fees?
Those outputs fit the separate fee-two variant, not the fee-one baseline. Obtain actual run settings, order-fee records, and the output's version, checking whether the screenshot belongs to that run. Duplicate charges, mismatched settings, and stale output are different hypotheses. The endpoint alone cannot choose among them.
If fee 2 is confirmed, register a controlled variant, then restore fee 1 and test the original. If disagreement remains, stop at its first occurrence. Do not change the baseline answer to 936 to make the comparison pass.
Attach a handoff card to T05
T05 remains the strategy specification. Its attachment records question, data/rule versions, file inventory, hash convention, environment, a path requiring no source edits, key expectations and tolerances, actual runs and outputs, failure/recovery, and one controlled change. Mark read-only steps, file-generating steps, old-version locations, and sensitive material excluded from sharing.
Ask the recipient to record actual operations and blockers in their own independent directory. Reading without recalculation is “readable for inspection.” If a tool cannot export cash and orders, operational reproduction remains unaccepted. Do not sign “handoff complete” for another person or transform author-machine success into human success in a clean environment.
Freeze minimal inputs, rules, tools, and expectations; compare intermediate accounts in a copy; preserve failures, restore, and change one condition under another version. Hand over environment, randomness, integrity conventions, and limits with the outputs. Reproduction supports consistency within that stated scope, not all research value, financial capacity, or operating readiness.
References
- National Academies: Reproducibility and Replicability in Science, 2019, Summary. Computational consistency with the same inputs and conditions versus studies with new data.
- Microsoft: Change formula recalculation, iteration, or precision in Excel. Calculation modes, not evidence of this reader's desktop result.
- Microsoft: Excel formatting and features that are not transferred to other file formats. CSV values and workbook features have different preservation scopes.
- Python 3.12: hashlib. Byte inputs and SHA256 digests, not data authenticity.
Sources checked October 30, 2026. Archived author execution scope is distinguished from this chapter's teaching tasks. AI-assisted draft; independent AI full-text and bilingual review is complete. Human financial review and target-reader testing have not been conducted and are not scheduled for this production round. Native Excel/WPS, a clean environment, and human handoff remain unverified.