“I rushed last time; next time I will stay calm” sounds reflective but is difficult to inspect. Which decision, which omitted step, and which changed behavior does it describe? Another month can pass with only a memory that reflection already happened.
A decision journal retains inputs, contemporaneous statements, actual actions, and later outcomes. Follow four complete records and a stage review before choosing one observable improvement. Begin with evidence rather than an empty form or a self-description.
Preserve a decision rather than write an autobiography
You will separate advance information, personal statements, and AI inference; assess four process observations before adding outcomes; and turn a specific gap into a later check. All four observations may involve no trade. Do not create transactions to fill T10.
MoneySense investor education recommends writing plans and conditions for changed decisions in advance.[1] Here that becomes a dated record: a new judgment can change while the old version remains recoverable. More writing is not the objective. Another reader must be able to identify when each reason appeared.
C5-58's person, company, dates, sources, and behavior are synthetic and independent of earlier characters. Lan has twenty research minutes weekly for one question about fictional North Bank Software: will its specified product complete its first paid commercial sale to an unrelated external customer by the end of October 16, 2026?
Settle against the exercise's complete sales ledger and later company explanation, checking payment date, customer scope, and product. Test accounts, intentions, and internal trials do not qualify. This is neither revenue growth nor a price increase. Without a complete real ledger, a real researcher would record insufficient evidence rather than equate not found with did not occur.
Real funding remains zero. Every record retains source-card ID, publication date, actual viewing time, source location, interpretation, decision, behavior, later result, and exceptions. AI notes carry their own label rather than become words Lan supposedly said.
September 7: define the event first
F1 was published September 4 and obtained September 7 at 10:00. Its second paragraph offers free testing; paragraph four says prices and charging dates will follow. Lan's dated statement notes that a usable product does not establish a paying customer.
She assigns the original event a working 60% probability, explicitly lacking a suitable reference class and making no calibration claim. Her plan is to investigate payment terms without producing a simulated order. Actual work takes eighteen minutes: save F1 and both locations, list questions, and place no order.
Charging information remains missing and the event unsettled. There is no exceptional trade. AI's suggestion that she may favor a new-product story is a candidate explanation without confirmation or independent support. It stays out of facts. Later comparisons can use her dated statement instead of invented motives.
September 14: a known gap remains unfilled
Forwarded F2, dated September 13, reaches her at 10:00 on September 14, claiming paying customers. Its image lacks customer scope, payment date, and original source. The exercise supplies her contemporary statement that she fears missing the opportunity. That does not establish the cause of every later action.
She plans to find the original. Actual behavior is thirty-two minutes repeatedly viewing the forwarded message and making an unsubmitted simulated buy-order draft without locating the source. No actual or simulated fill occurs. This date's probability is missing; the earlier 60% cannot silently become this observation's estimate. The event remains unsettled.
She exceeds the twenty-minute budget by twelve and crosses the plan's evidence-before-drafting gate. AI writes that fear caused her to buy. Buying contradicts the supplied actions and must be removed; motive remains an unverified inference. A smooth narrative cannot turn a draft into a transaction.
September 21: add a source, not a rewritten past
F3's price page was published September 18 and saved September 21 at 10:00. Line one gives a charging plan; line three expects commercial orders to open October 5. It does not disclose an actual customer payment. Lan records that she is checking the original and that a quotation is not a completed sale.
Her new working probability is 70%, based on charging and opening plans rather than supposed payment evidence. She plans to retain F3's locations and seek ledger information. In sixteen minutes she does so without submitting a simulated draft or trading. The plan is verified; actual sales remain unknown and the event unsettled. No exception occurs; September 14 stays blank.
AI's statement that her emotions are more stable lacks measurement. The record establishes source-location saving before action, not that AI or this tutorial made her calmer or that 70% is more accurate than 60%.
September 28: revise the judgment within its original deadline
F4 was published September 25 and obtained September 28 at 10:00. Paragraph two delays expected commercial opening until October 31, still subject to change. This weakens a sale-by-October-16 case without establishing legal or technical impossibility of an earlier sale.
Lan keeps the deadline and lowers her working estimate to 20%. She plans to pause new simulated requests and check commercial-sales information October 8. Actual work takes fourteen minutes: retain paragraph two, close the unsubmitted draft, and record the pause reason, with no fill.
Closing a draft is not canceling a submitted order. The event is not due and no exception occurs. No October 8 follow-up record is supplied, so that task remains awaiting verification, not completed by assumption.
AI calls this finally learning to admit error. That is an evaluation rather than an observed event. A reviewer can say the update's direction matches new information while retaining questions about unexamined channels. One revision does not establish personality change.
Review the four processes before revealing October's outcome
The stage review is October 20. Hide settlement first. F1, F3, and F4 have original locations and support scope; F2 does not. Completion is three of four, or 75%. Times of 18, 32, 16, and 14 minutes total eighty, equal to the four-week budget, concealing the second week's twelve-minute overrun.
The identifiable defect is crossing a drafting gate without F2's original evidence, followed by AI reporting a nonexistent purchase. Emotion by itself is not the defect. Nor are the other entries automatically sound: the 60%, 70%, and 20% estimates remain weakly supported, and F4's channel coverage needs examination.
Outcome-bias research finds that results can influence judgments of earlier thinking, without establishing that outcomes can never reveal relevant new evidence.[2] Save this process assessment first, then add later material. If it exposes a previously discoverable omission, revise that specific process rather than dismiss everything as luck.
Now reveal F5, published October 19 and obtained October 20. The supplied complete ledger and subsequent explanation confirm no qualifying paid commercial sale by the end of October 16. Free testing does not count. Record nonoccurrence of the original event.
All four records concern that one final event; they do not create four independent forecast results. There were no holdings or fills, so investment profit and loss is not applicable—not 0% earned or successful loss avoidance. Delay helps interpret the result without rewriting the original 60% as 20% or filling the September 14 gap to improve calibration.
Select one change with an observable check
The sole new improvement is to complete original location plus support scope before new material changes a judgment or produces a simulated draft. If either is missing, the verification is incomplete and cannot justify higher confidence or a new draft. Questions, acknowledgment of uncertainty, and timely real-risk handling remain possible.
Predefine four comparable material-review opportunities in the next cycle, rather than require four trades. Retain verification completion time, judgment/draft action time, gate violations, and exception reasons. Only verified-before-action counts. A week with no new applicable material is not an opportunity; do not manufacture success or ask AI to prefill four completions.
A later move from 3/4 to 4/4 can describe the records across those eight observations. Without a control, random assignment, or other identification, it cannot establish that the tutorial or AI caused a 25-percentage-point improvement. Difficulty, familiarity, timing, chosen opportunities, and missing records can differ.
Inspect more than checkmarks. A filled location pointing to the wrong document, or an exaggerated support statement, still fails. Excluding the hardest case or relaxing completion's meaning to improve the score also changes the comparison. Return to G12's actual claim/source match.
Three probabilities still concern one settlement
The 60%, 70%, and 20% values update the same deadline event under changing information. They help examine revision reasons. Repeating one nonoccurrence three times cannot establish generally excessive confidence or that low-probability forecasts outperform high-probability ones.
For calibration, specify which forecast time counts for each event—for example an initial version at a common distance from settlement. Accumulate comparable events with fixed grouping, horizons, outcome definitions, and missing policy. Met Office reliability diagrams similarly compare grouped forecasts with observed frequencies.[3] Investment application still needs dependence and environment checks; four journal entries are not ability certification.
Range forecasts need another field. Revenue growth between 5% and 15% is not a 70% probability of a commercial sale. Without range coverage, fiscal period, and revenue scope it cannot share the calibration chart. Make the object precise before insisting on a number everywhere.
Leave a gap instead of inventing a history
If she later recalls being optimistic on September 14 but cannot recover a number, add an October 20 recollection with uncertainty. It is not the original day's record. AI can list questions but cannot insert 70% or 90% into an earlier date to create a complete past.
Revised documents likewise retain old visible versions and new acquisition times, with affected conclusions identified. Behavior follows evidence: draft, submission, acceptance, partial fill, and final fill differ. This case contains only a draft; do not generate simulated returns to show what could have been earned.
Use fictional exercises without names, addresses, identity documents, brokerage numbers, asset screenshots, or real trades. Keep an actual original ledger securely if needed for private review, and share only redacted fragments necessary for the particular check rather than a default full-account export.
ICO guidance on data minimization emphasizes relevant, sufficient, necessary information and clear labeling of opinion rather than fact.[4] These are borrowed record-design principles, not a determination of each reader's legal jurisdiction. Changing amounts for privacy also cannot be called verified real reconciliation; create a separately labeled synthetic demonstration.
Add new evidence without erasing the original gap
Explanation question: Why does nonoccurrence not mean four independent failed predictions?
Application question: AI reports closing the draft as an executed sale. What should be corrected and preserved? Answer before reading on.
The entries update one event and one lacks a probability. Correct the action to closing an unsubmitted draft and remove nonexistent trades and returns from facts. A marked erroneous AI version can remain for audit, without continuing as the factual account.
For a separate variation, on October 20 an original timestamped September 14 email is discovered. It explicitly assigns the same October 16 event 75%, but still supplies no verified F2 original source. May the probability be added? Does verification become complete or calibration become four independent events?
Append the original email with its September 14 timestamp and October 20 discovery, preserving the initial missing-record state. It supports a 75% contemporaneous estimate in a way AI guessing does not. It does not supply the missing source location, so verification stays 3/4. The underlying event still settles once; the email cannot automatically become another independent calibration sample.
Save four T10 entries and one stage review
Reconstruct all four using fictional inputs. Each needs ID, information cutoff and acquisition, traceable material and support, personal statement or missing status, labeled AI inference, contemporaneous probability and horizon, plan, actual behavior, then-visible outcome, and exceptions. Later evidence gets another date instead of overwriting old entries.
The stage page lists process gaps before settlement, then outcome and new evidence, one improvement, next observation scope, completion definition, missing policy, and exceptions. Check that updates were not counted as independent events, recollection as prior fact, or form-completion change as investment return or causal effect.
Know which step the next review should change
Separate then-available evidence, statements, AI inference, and behavior, retaining genuine gaps. Evaluate process under its original definition before adding outcomes. A decision involving no trade can still produce a complete record.
Improve one observable step and preserve opportunities and denominators before checking execution. Four records can locate one violation without proving forecasting skill or a therapeutic effect. Traceable revisions contribute more than an unsupported self-rating.
References
- MoneySense: Behavioural Science and Investing. Updated September 21, 2026. Written plans and changing-decision conditions only; no fixed psychological multiplier or intervention guarantee adopted.
- Baron and Hershey: Outcome Bias in Decision Evaluation. JPSP 54(4), 1988, 569–579. Experimental conditions do not represent all investors.
- Met Office: Reliability and Sharpness Diagrams. Forecast groups and observed frequency; financial record design is a teaching transfer.
- ICO: Principle (c)—Data Minimisation. Necessary information and opinion labels; design principles, not individual legal advice.
Sources checked October 30, 2026. C5-58 and the recovered-email variation are synthetic, not user history, trades, or validated teaching effects. 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.