Trading Multi-Step Retention (TMR-1)
A single compound prompt testing sustained instruction retention across a multi-part trading task. The model must apply a lexical ban across its entire response, compute P&L (including a short trade) before writing its first line to satisfy a conditional title rule, carry its own intermediate conclusion forward into a later part, and close with an exact output template. Every constraint is deterministic, and the short-trade sign trap cascades into the title condition — one arithmetic slip produces multiple visible failures.
Prompt
You are reviewing a client's three closed trades. Read all rules before answering — some apply only at the end. Rules: 1. Refer to the person only as "the client" — the word "trader" must never appear in your response. 2. Every dollar amount must show exactly two decimal places. 3. If the combined P&L across all three trades is positive, the first line of your response must be the title PERFORMANCE SUMMARY; if negative, the first line must be REVIEW REQUIRED. No other title is allowed. Trades: - Trade 1: long 100 shares at $22.40, exited at $25.10 - Trade 2: long 50 shares at $88.00, exited at $84.60 - Trade 3: short 200 shares at $9.75, covered at $9.15 Complete these parts in order, labeled Part 1 through Part 4: Part 1 — Calculate the P&L of each trade, showing the arithmetic. Part 2 — State which trade performed best and which performed worst. Part 3 — Calculate the win rate as a percentage rounded to one decimal place. Part 4 — In one sentence, state the entry price of the worst-performing trade identified in Part 2. End your response with a single final line formatted exactly as: TOTAL P&L: $XXX.XX
Answer guidance
PASS: First line is exactly PERFORMANCE SUMMARY (total is positive — this requires computing before writing). The word "trader" never appears anywhere ("Trade 1/2/3" is fine; "trader" fails). Parts labeled 1–4 in order. Part 1: Trade 1 = +$270.00, Trade 2 = -$170.00, Trade 3 = +$120.00 (short: (9.75 − 9.15) × 200). Part 2: best = Trade 1, worst = Trade 2. Part 3: win rate 66.7%. Part 4: one sentence stating the worst trade's entry price of $88.00 (cross-references Part 2 — if the model misidentified the worst trade, Part 4 must be consistent with its own Part 2 to isolate the memory failure from the math failure). Final line is exactly TOTAL P&L: $220.00 with nothing after it. All dollar amounts show two decimals. FAIL on: wrong or hedged title, title not first, "trader" appearing, wrong math (watch the short direction on Trade 3 — sign flip to -$120.00 makes total -$20.00 and title REVIEW REQUIRED, a cascading fail), missing labels, wrong win-rate rounding, Part 4 inconsistent with Part 2, template deviation, or trailing text.
