Data room
Nine AI reviewers were told this product had already failed. Their job was to explain why.
The idea
Millions of homeowners became equity rich almost overnight when prices surged, but they cannot reach that money without giving up a low mortgage rate or taking on a monthly payment, and many of them only need a bridge for a few years. The intent was to help those households rather than profit from their squeeze: a fair, plainly described way to turn part of that equity into regular monthly cash, settled when the home is sold.
The panel
Nine AI agents, each playing one specialist (tax, securities, title, bankruptcy, insurance, operations, an investor, a customer, and a red team), were told the product had already failed and asked to explain why. They wrote independent reports, cross-examined each other, traced thirteen failure scenarios, and built a pricing model, and the synthesis of their work was a no-go on the first design.
The outcome
The idea was shelved. A large lender, Finance of America, launched essentially the same product: a second-lien loan with no monthly payment for homeowners 55 and older, with a line-of-credit version alongside it. The one design that survived this review was the one closest to it.
This page shows how the review of Equity Flow™ worked and holds everything it produced, as written.
- Parallel review18 min
- Cross-examination1 h 12 min
- Scenario trace6 min
- Pricing model5 min
- Synthesis4 min
56 documents, 177,015 words. 94,503 of them are the reviewers' own reports and cross-examinations.
How the run worked
There was no orchestration code. One Claude Code session followed a written runbook, launched each reviewer as a subagent with its own prompt, and kept its place by writing to disk. Every stage read its inputs from files, never from the conversation, and appended a line to the run log when it closed.
- Assume it failed. Every reviewer is told the product has already failed and asked to explain why. No softened findings.
- Tag every claim. Each legal or tax claim carries one of five confidence tags, from settled law to speculative.
- End on a verdict. Kill, fix before launch, or acceptable, with reasons.
- Do not stop to ask. No questions to the human mid-run. A failed stage gets one retry, is logged, and the run continues.
-
1
Parallel review
18 min
Eight reviewers were dispatched together, each with its own prompt and the product documents, and each wrote a full report to disk. The red team was held back until all eight reports existed, then given the documents and all eight reports with one extra job: find the failure the panel missed.
Wave 1, 8 min: eight reviewers at once
All eight reports on disk
Wave 2, 10 min: the red team reads everything
voted killvoted fix before launch
What the run log recorded
Stage 1 wave 1 complete. 8/8 reports written: retail-customer (KILL), investor (KILL), tax-attorney (KILL), estate-bankruptcy (FIX w/ kill conditions), securities-consumer-finance (FIX w/ 2 embedded kills), real-estate-title (FIX, kill-equivalent in TX), product-operator (FIX), insurance-actuarial (FIX w/ 2 embedded kills) | anomalies: none — all files non-empty | dispatching red-team wave 2
Stage 1 complete. red-team.md written (671 lines, verdict KILL, incl. WHAT THE PANEL MISSED: market-pause-as-fund-escape-hatch, post-assignment covenant weaponization) | 9/9 reports on disk | anomalies: none | proceeding to Stage 2 cross-examination
Written to disk: Nine reviews
-
2
Cross-examination
1 h 12 min
Each of the nine received the other eight reports and answered two questions. Which of these findings do you dispute, and why? Which fixes proposed by other reviewers create new problems in your domain? Direct conflicts between reviewers were collected for the synthesis to resolve or surface.
RC II TA SC RE EB PO IA RT Retail customer Institutional investor Tax attorney Securities and consumer finance Real estate and title Estate and bankruptcy Product operator Insurance and actuarial Red team Each row is a reviewer. Each filled square is another reviewer's report it was handed: 72 readings in all. Interrupted, then recovered
The nine parallel agents were cut off mid-stage when the account reached its monthly usage limit. Seven of the nine cross-examinations were already on disk and were verified complete. The runbook's rule for a failed stage is one retry, then continue: the two missing ones (investor, and insurance and actuarial) were re-run, and the stage closed 18 minutes after the interruption was logged. Nothing else had to be redone, because every stage reads its inputs from disk.
What the run log recorded
Stage 2 failure. All 9 cross-exam agents terminated early: monthly spend limit on subagent usage. No crossexam-*.md files written. Per runbook failure rule: attempting one retry (single probe agent); if probe fails, Stage 2 will be executed inline by the orchestrator from the on-disk reports and the deviation logged.
Stage 2 complete. 9/9 crossexam files written. Anomalies: initial parallel run terminated by monthly spend limit AFTER 7/9 files were written (all verified complete, untruncated); investor and insurance-actuarial re-run successfully per one-retry rule. Notable: actuarial conceded both investor disputes (recovery at -45pct = 0.175x not 1.0x; Ch7 discharge channel omitted), Shield coverage revised down to 0.29-0.33x. Central conflict DC-1 (lien-vs-tax characterization fork) flagged by every reviewer. | Stage 3 (scenarios) executing inline by orchestrator
Written to disk: Nine cross-examinations
-
3
Scenario trace
6 min
The product was walked end to end through thirteen failure scenarios. For each one: who gets paid, who takes the loss, what the customer and the investor experience, and whether any contract term fails. Every outcome the specification left undefined was flagged.
- Home prices fall 20%
- Rates rise fast
- The customer dies
- Divorce
- Chapter 7 bankruptcy
- Chapter 13 bankruptcy
- A senior lender forecloses
- Early sale
- Sale delayed 15 years or more
- A natural disaster destroys the property
- Coordinated fraud
- Recession with three times the expected defaults
- Adverse change in tax law
What the run log recorded
Stage 3 complete. scenarios/SCENARIOS.md written (13 scenarios traced end-to-end, pause-paradox dual-branch applied, 27 undefined-outcome items in consolidated register) | executed inline by orchestrator | anomalies: none
Written to disk: Scenario traces
-
4
Pricing model
5 min
A Python model of the economics across twelve appreciation rates and eleven holding periods, held to a written list of assumptions. Anything the code needed that was not on the list had to be added to the list, not buried in the code.
What the run log recorded
Stage 4 complete. models/: equity_flow_model.py + ASSUMPTIONS.md (26 stated assumptions) + RESULTS.md (6 grids, 2 summaries, stress table, reconciliations; -27pct breakeven drawdown and 0.17x sand-state recovery reproduced; 43 sub-hurdle investor cells; EF worse than HELOC in 81 cells) | anomalies: 2 IRR-solver bugs fixed during run (underflow bounds, sign-change bracketing), logged here per runbook
Written to disk: Model assumptions, Model results, Model source
-
5
Synthesis
4 min
One ranked analysis from everything on disk. The rules: rank, do not summarize; cite every legal and tax claim to the review it came from; nothing tagged speculative may appear in the verdict as if established. This stage could not run unless stages 1 to 3 had finished.
What the run log recorded
Stage 5 complete. ANALYSIS.md written (13 sections; executive verdict NO-GO; findings ranked 9 dealbreakers / 13 fix-items / cosmetic list; all legal-tax claims tagged and cited to underlying reviews; SPECULATIVE items excluded from verdict) | RUN COMPLETE: Stages 1-5 all executed; Stage 2 spend-limit failure recovered per one-retry rule; Stage 4 solver bugs fixed in-run
Written to disk: Final analysis
The panel, and where each reviewer landed
Four reviewers voted to kill. Five voted to fix before launch, four of them with a kill condition attached. Each prompt, review, and cross-examination is one click away.
-
Retail customer
Reads the homeowner page the way a customer would: clarity, true cost, fairness, predatory patterns.
KillThe homeowner calculator left out the investor’s share, and the page showed none of the contract’s material terms.
-
Institutional investor
Market size, unit economics, capital needs, defensibility, structured-finance risk.
KillAt trend home-price growth the deal earns about 7.2% gross. The marketed 11% needs 4.67% a year of appreciation.
-
Tax attorney
Federal and state tax characterization, debt versus equity, reporting burden, audit risk.
KillThe tax authority the spec relied on, IRC Section 453, does not apply on its face. The likeliest reading is a shared appreciation mortgage.
-
Securities and consumer finance
Howey and Reves, the Investment Company Act, TILA, RESPA, ECOA, UDAAP, state licensing.
Fix before launchtwo embedded killsBoth marketing pages and the guarantee product were kills as written. The public investor page was itself a general solicitation.
-
Real estate and title
Recordability in 50 states, lien priority, title insurance, due-on-sale, homestead, foreclosure.
Fix before launchkill-equivalent in TexasVoid against Texas homestead as drafted, and the Florida death path fails. The claim of all 50 states was false.
-
Estate and bankruptcy
Death, probate, trusts, incapacity, divorce, and Chapters 7, 11 and 13.
Fix before launchtwo kill conditionsThe promises that the interest could not be discharged or modified in bankruptcy were false under the Bankruptcy Code’s own definitions.
-
Product operator
Origination cost, servicing, valuation integrity, decades of dormant-position administration.
Fix before launchOrigination lost about $2,000 to $4,000 a deal at the stated fee, and decades of servicing were priced at zero.
-
Insurance and actuarial
Reserve solvency, insurance classification, statutory capital treatment, the secondary-market thesis.
Fix before launchtwo embedded killsThe secondary sale the fund depended on priced at 0.69x to 1.10x of deployed capital, against the 1.41x the marketed return required.
-
Red team
Three attackers in one pass: a fraudster, a competitor, and a journalist or regulator.
KillFound what the other eight missed: the market pause works as a distressed fund’s escape hatch, and covenants can be turned against homeowners after assignment.
What cross-examination changed
The second round is where the panel stopped being nine separate opinions. Reviewers conceded points, sharpened others, and found that their repairs did not fit together.
A reviewer withdrew its strongest finding in the product's favor
The actuarial reviewer conceded both of the investor's disputes: recovery after a 45% price decline is 0.175x, not 1.0x, and it had left out the Chapter 7 discharge channel. It cut the guarantee reserve's coverage to 0.29x to 0.33x of the design-case tail. Read it
Every reviewer arrived at the same central conflict
The product needed its recorded interest to be not a debt and not a lien for tax and regulation, and to be exactly a lien for bankruptcy, probate, and homestead. The panel labeled it DC-1. The estate reviewer said its verdict becomes a kill if that fork cannot be resolved. Read it
The fixes undid each other
The five reviewers who voted to fix proposed repairs that were destructive in pairs. The lien fix breaks the tax position. Honest duration economics make the fund unraisable. A consumer-fair pause removes the fund's only limit on credit risk. Read it
The verdict: no-go
Equity Flow as specified and marketed should not be launched, and the defect is architectural rather than reparable.
The nine dealbreakers, as the synthesis ranked them
- The characterization fork is unresolvable as designed
- The tax architecture fails at its foundation
- Bankruptcy/probate representations are false and self-destructive
- Texas is closed and Florida's death path fails
- The fund economics do not clear the cost of capital at trend housing
- The liquidity thesis is circular
- The marketing is contradicted by the sponsor's own documents
- Equity Shield is unlawful and insolvent as structured
- The counterparty-credit structure fails homeowners
Behind those nine sit thirteen defects that could each be cured at a cost, twelve ranked risks, six questions the panel could not settle, and fifteen questions written for licensed securities, tax, and consumer-finance counsel.
The synthesis closes on the question the panel had been asked: whether the product could be legal, honest, and financeable at once. Its answer was that each property was achievable only by surrendering another. In its words, "That is a design verdict, not a compliance gap."
What the verdict changed
The product that was reviewed was withdrawn. The analysis ended with the only redesign the panel thought coherent, and that became the next concept the same day. Three more designs went back through the reviewers whose findings mattered most to each. One survived.
-
Version 1Killed
Equity share
Monthly payments to the homeowner in exchange for the capital back plus 25% of future appreciation.
Full panel, five stages.
No-go. Nine dealbreakers, and fixes that undo each other.
-
Version 2Survived
Deferred-interest loan
Working name HomeDraw. A second-lien loan with a monthly draw and no payments until sale, called a loan everywhere. Built from the panel’s own redesign recommendation.
Three reviewers: securities and consumer finance, investor, insurance and actuarial.
Fix before launch from all three. The securities reviewer found no embedded kills. This became the basis of the product on the investor page.
Concept: deferred-interest loanReview: securities and consumer financeReview: investorReview: insurance and actuarial
-
Version 3Killed
Symmetric index-settled share
A lump sum for a slice of metro home-price exposure, with the investor sharing losses as well as gains.
Three reviewers: investor, securities and consumer finance, tax.
Killed by the investor and securities reviewers. A sweep of 175 parameter combinations found none that paid both sides.
Concept: symmetric index-settled shareReview: investorReview: securities and consumer financeReview: tax
-
Version 2.1Killed
Open line of credit
The loan reshaped as a draw-whenever credit line with the monthly schedule as a feature.
One reviewer on one question: does the funding math hold?
Killed as specified. The corrected central-case consumer rate was 13.55% against a 10.5% ceiling.
Concept: open line of creditReview: funding and utilizationFunding model resultsFunding model source
A second full run was prepared: 55 questions drawn from an outside critique of the public site, routed to eight of the nine reviewers. It has not been dispatched. See the question list
What this is, and what it is not
- The reviewers are AI agents playing professional roles. None is licensed counsel, and no outside legal opinion has been issued.
- The statutes, cases, and rulings cited in the transcripts have not been checked by a lawyer. The confidence tags are the agents' own.
- Stages 3 to 5 were written by the orchestrating session from the reviewers' files, not by separate agents.
- The only cost control was the account's monthly spend limit, and it cut the cross-examination stage mid-run.
- The full panel reviewed a version that was withdrawn. The design that survived had a three-reviewer pass, not a full run, and it was never reviewed by licensed counsel or a rating agency.
- The documents are published as written. The only edits are absolute file paths from the machine that ran them, shortened to relative ones. The product specification the panel reviewed is not published here.
All documents
No document matches that. Try a reviewer's name or a document type.
The method
The written procedure the orchestrating session followed, and the nine prompts that define the reviewers.
- RunbookProcedure1,044 words
- Retail customerRole prompt670 words
- Institutional investorRole prompt825 words
- Tax attorneyRole prompt786 words
- Securities and consumer financeRole prompt687 words
- Real estate and titleRole prompt758 words
- Estate and bankruptcyRole prompt1,090 words
- Product operatorRole prompt934 words
- Insurance and actuarialRole prompt872 words
- Red teamRole prompt953 words
- Question list for run 1Work assignment3,092 words
Run 1, stage 1: the nine reviews
Each reviewer’s full report on version 1, ending in its verdict.
- Retail customerReview4,846 words
- Institutional investorReview4,750 words
- Tax attorneyReview5,991 words
- Securities and consumer financeReview6,707 words
- Real estate and titleReview6,586 words
- Estate and bankruptcyReview6,481 words
- Product operatorReview5,587 words
- Insurance and actuarialReview3,988 words
- Red teamReview6,040 words
Run 1, stage 2: the nine cross-examinations
Each reviewer’s disputes with the other eight reports, and the fixes it says would break something in its own domain.
- Retail customerCross-examination4,400 words
- Institutional investorCross-examination4,109 words
- Tax attorneyCross-examination4,060 words
- Securities and consumer financeCross-examination5,893 words
- Real estate and titleCross-examination5,124 words
- Estate and bankruptcyCross-examination4,800 words
- Product operatorCross-examination5,374 words
- Insurance and actuarialCross-examination4,384 words
- Red teamCross-examination5,383 words
Run 1, stages 3 to 5: scenarios, model, verdict
Written by the orchestrating session from everything the reviewers left on disk.
Successor concepts and their reviews
The three designs that followed the no-go, each reviewed the same day by the reviewers whose findings mattered most to it.
- Version 2 concept: deferred-interest loanConcept1,320 words
- Version 2 review: securities and consumer financeReview6,079 words
- Version 2 review: investorReview4,633 words
- Version 2 review: insurance and actuarialReview4,701 words
- Version 3 concept: symmetric index-settled shareConcept1,376 words
- Version 3 review: investorReview4,533 words
- Version 3 review: securities and consumer financeReview5,177 words
- Version 3 review: taxReview7,274 words
- Version 2.1 concept: open line of creditConcept763 words
- Version 2.1 review: funding and utilizationReview3,295 words
- Version 2.1 funding model resultsModel292 words
- Version 2.1 funding model sourceCode572 words
Run 2: prepared, never dispatched
A second run on the current public site was set up and has not been run.
The current model
The one model that produces every number on the investor page, read live from this site’s source.