Runbook
The five-stage procedure, as executed. There was no orchestration code: this file was the orchestrator’s instructions.
panel/RUNBOOK.md Open the file as written
Equity Flow — Diligence Pipeline RUNBOOK
Autonomous execution plan for the full adversarial diligence run. Working
directory for every stage and every relative path in this file:
product-diligence/.
Execution rules
- Run stages strictly in order; each stage reads its inputs from disk, never from conversation memory.
- After completing each stage, append a line to
STATUS.mdwith the stage name, timestamp, output files written, and any anomalies. - If a stage fails or produces obviously degenerate output (empty report,
agent refusing, missing file), log it in
STATUS.md, attempt one retry, then CONTINUE to the next stage rather than halting — except Stage 5, which must not run unless Stages 1–3 completed. - Never ask the user for input. Where this guide says "show me," write to the file instead. All human review happens after the run.
- In Stage 4, hold the model to stated assumptions written into
models/ASSUMPTIONS.md.
Source materials
- Full product disclosure/spec:
../disclosures/equity-flow-invention-disclosure.txt - Homeowner-facing website section:
../public/equity-flow.html(the "For Homeowners" section: enrollment flow, "How Equity Flow works," cash-flow estimates, market-scenario and monthly-impact comparisons) - Investor-facing website section:
../public/equity-flow.html(the "For Investors" section: revenue streams, deal economics, ride-along/secondary- market pitch, investor FAQ, market-sizing tool, Equity Shield)
Panel agent definitions live in ../.claude/agents/. Every agent operates
under its own prompt: adversarial failure-hunting, claim tagging
(SETTLED LAW / LIKELY INTERPRETATION / STATE-DEPENDENT / AMBIGUOUS /
SPECULATIVE), and a kill / fix-before-launch / acceptable verdict.
Stage 1 — PARALLEL REVIEW
Wave 1: dispatch eight agents in parallel
Each agent writes its complete report to reviews/<agent-name>.md.
| Agent | Materials | Extra task beyond its standing prompt |
|---|---|---|
retail-customer |
Homeowner website section (primary), full spec (reference baseline) | Flag everything the page omits, buries, or contradicts relative to the spec; independently compute the homeowner's true cost / APR-equivalent across hold durations. |
investor |
Investor website section AND full spec | Flag every place the investor marketing overpromises or diverges from what the spec actually delivers, including rebuilding the ~11% IRR claim from scratch. |
tax-attorney |
Full spec | — |
securities-consumer-finance |
Full spec AND both website sections | Marketing-consistency review: any claim on either page unsupported by or contradicting the disclosure (UDAAP/deceptive-practices exposure on the homeowner side; securities-offering and misrepresentation exposure on the investor side). |
real-estate-title |
Full spec | — |
estate-bankruptcy |
Full spec | — |
product-operator |
Full spec | — |
insurance-actuarial |
Full spec | — |
Hold red-team until all eight complete.
Wave 2: red-team
Run red-team against the spec, both website sections, and all eight
reviews in reviews/. Its extra job: find the failure mode the panel
missed. Write to reviews/red-team.md.
Stage 1 outputs: reviews/retail-customer.md, reviews/investor.md,
reviews/tax-attorney.md, reviews/securities-consumer-finance.md,
reviews/real-estate-title.md, reviews/estate-bankruptcy.md,
reviews/product-operator.md, reviews/insurance-actuarial.md,
reviews/red-team.md.
Stage 2 — CROSS-EXAMINATION
Give each of the nine reviewers (the eight panel agents plus red-team) all
the OTHER agents' reports from reviews/ and ask two questions:
- Which of these findings do you dispute, and why?
- Which proposed fixes from other reviewers create NEW problems in your domain?
Each reviewer writes its response to reviews/crossexam-<agent-name>.md.
Flag every direct conflict between reviewers (finding A from one agent incompatible with finding or fix B from another). Collect the conflict list in the cross-exam files; Stage 5 must resolve or surface each one.
Stage 2 outputs: nine reviews/crossexam-*.md files.
Stage 3 — SCENARIOS
Read ../disclosures/equity-flow-invention-disclosure.txt and all reviews
(reviews/*.md, including cross-exams). Create scenarios/SCENARIOS.md
tracing the product end-to-end through each of:
- Falling home prices (−20%)
- Rapidly rising rates
- Customer death
- Divorce
- Customer bankruptcy — Chapter 7
- Customer bankruptcy — Chapter 13 (separately)
- Foreclosure by a senior lienholder
- Early sale
- Sale delayed 15+ years
- Natural disaster destroying the property
- Coordinated fraud
- Recession with 3× expected defaults
- Adverse tax-law change
For each scenario: who gets paid, who eats the loss, what the customer experiences, what the investor experiences, and whether any contract term fails or becomes unenforceable. Flag every scenario where the outcome is undefined by the spec.
Stage 3 output: scenarios/SCENARIOS.md.
Stage 4 — QUANT MODEL
Build a Python model in models/ of the product economics:
- Customer outcomes and investor returns across the scenario grid:
home-price appreciation from −3% to +8% annually, hold periods 1–30
years, the fee structure from
../disclosures/equity-flow-invention-disclosure.txt(monthly disbursement, 2–3% origination fee, 1.0x capital return, 25% appreciation share, +1%/yr ride-along ratchet capped at 50%, market-pause mechanics, settlement waterfall). - Output sensitivity tables and a summary of where the customer does worse than the obvious alternative (HELOC / cash-out refi / doing nothing) and where investor returns go negative.
- Write the key tables and findings to a results file in
models/(per the execution rules, nothing is "shown" interactively). - State every assumption the model makes in
models/ASSUMPTIONS.md, and hold the model to those stated assumptions — any assumption the code needs that is not in ASSUMPTIONS.md gets added there, not silently embedded.
Stage 4 outputs: model code in models/, models/ASSUMPTIONS.md, results/
sensitivity tables in models/.
Stage 5 — SYNTHESIS
Gate: must not run unless Stages 1–3 completed (per execution rules; a failed Stage 4 is logged but does not block, though section 8 must then say the model is missing).
Synthesize everything into ANALYSIS.md with this structure:
- Executive verdict (go / modify / no-go, one paragraph)
- Product mechanics (from
../disclosures/equity-flow-invention-disclosure.txt) - Findings ranked by severity: DEALBREAKERS, then FIX-BEFORE-LAUNCH (with cost/difficulty), then COSMETIC
- Customer value proposition assessment
- Legal/regulatory analysis with every claim tagged SETTLED / LIKELY / STATE-DEPENDENT / AMBIGUOUS / SPECULATIVE
- Tax analysis, same tagging
- Estate and bankruptcy analysis
- Unit economics summary (from the model)
- Risk register and failure scenarios
- Competitive and reputational exposure
- Unresolved questions
- Required professional opinions — the specific questions a real securities attorney, tax attorney, and consumer-finance attorney must answer before launch, in priority order
- Recommended product changes
Rules for this stage:
- Rank, don't summarize.
- Every legal/tax cite must reference the underlying reviews.
- Nothing labeled SPECULATIVE may appear in the executive verdict as if established.
Stage 5 output: ANALYSIS.md.