Kumu / Finance & Securities / Stage 5
Stage 5 — Validate & Decision Memo
Feed your spec and market-data memo, and nothing else, to a fresh LLM and have it produce the complete hedge analysis independently. Then do what a professional does with any model output, human or machine: verify it by hand, reconcile the differences, and only then make the recommendation.
Objective
- The production test: do your documents stand alone, or did the model live in your head?
- Independent verification: the discipline that separates using a model from trusting one
- Executive recommendation writing: insight over computation, judgment over arithmetic
Do — part 1: independent LLM execution
Open a fresh LLM conversation: no history. Provide exactly two documents: your Stage 2 spec and your Stage 4 market-data memo (GitHub links or file uploads). Ask it to compute all hedge outcomes and recommend a strategy.
Do not coach it
No correcting mid-run, no pasting your workbook's results, no "actually, the rate basis is…". This is a production test of your documents. Whatever the LLM gets wrong is data for Part 4. Log the prompt.
Do — part 2: comparison & hand verification
- Comparison table: LLM's result vs. your workbook's result for each
strategy: forward, MM, put, call, unhedged at 2–3
S_Tpoints. Flag every discrepancy and diagnose it: LLM error, workbook error, or spec ambiguity? - Hand-verification table (≥3 outcomes, arithmetic shown): recompute by
hand: calculator and named-range notation, no Excel:
- forward proceeds (
FC_AMT × F0_in), - the money-market hedge, all three steps,
- one option outcome (e.g., put floor net of premium at a chosen
S_T).
- forward proceeds (
Do — part 3: the decision memo
2–4 pages, to the CFO, insight over computation (the CFO already has the numbers):
| Section | What it must do |
|---|---|
| A · Exposure summary | Brief restatement: a reminder, not a rewind |
| B · Hedge outcomes | Key findings per strategy including the unhedged baseline: what each outcome implies for the decision, citing figures as evidence, not re-deriving them |
| C · Sensitivity interpretation | Behavior under EUR depreciation vs. appreciation: certainty, flexibility, and cost differences |
| D · Recommendation | One strategy (or combination), supported by your live-data numbers |
| E · Executive justification | Cash-flow stability, budget certainty, liquidity, optionality, premium cost; accounting implications optional |
There is no universally right answer
A forward and a put can both be defensible for the same scenario; what's graded is whether your choice follows from your model's numbers and the firm's risk posture. "We chose the put because the premium buys participation in a EUR rally our board considers likely" beats "options are more flexible" every time.
Do — part 4: spec retrospective & repo polish
Retrospective (½–1 page, in the validation doc): what did the LLM get wrong or have to guess, and what does that reveal about your spec? What would v2 of the spec say differently? Candor is graded; "the spec was perfect" is not a retrospective.
Repo polish: top-level README current (bio + a project section linking every
stage artifact), one-line repo description set, all files in canonical locations with
convention-compliant names, stub READMEs still accurate, prompt-log.md complete
through stage 5, clean commit history, repo public.
Feedback may still arrive as a pull request
Final-stage review, like every stage before it, can come as a pull request pushed to your repo. The five ways to respond (read the diff, merge, edit on top, hand it to an LLM, or push back with reasons) are laid out on Stage 4. A reasoned disagreement is as valid a close to the project as a clean merge.
Deliverable
| What | Where it goes |
|---|---|
| The validation | analysis/YYYY-MM-DD-{scenario}-validation-analysis.md |
| The decision memo | docs/decisions/YYYY-MM-DD-{scenario}-hedge-decision-memo.md |
- Fresh LLM session; exactly two documents provided; no coaching; prompt logged
- Raw LLM output saved as an appendix or linked file
- Comparison table complete: all strategies, 2–3
S_Tpoints, every discrepancy diagnosed - Hand-verification table: ≥3 outcomes recomputed with arithmetic shown, reconciled to the workbook
- Decision memo sections A–E; every claim traceable to your live-data numbers
- Spec retrospective: specific, honest, ties LLM failures to spec gaps
- Repo polish checklist complete; repo genuinely portfolio-ready
- Validation doc at
analysis/YYYY-MM-DD-{scenario}-validation-analysis.md; memo atdocs/decisions/YYYY-MM-DD-{scenario}-hedge-decision-memo.md - Final
prompt-log.mdcommitted