Strategic Assumption Discovery and Validation

Wrote docs/architecture/tasks/mortgage-assume-001/assumption-discovery/assumption-register-spec.md. Complete strategic assumption register and validation specification establishing epistemic classifications, risk-uncertainty scoring matrices, falsification experiments, and go/no-go decision gates.

skill: assumption-discovery

Assumption Register & Validation Spec: AI Mortgage Underwriting [ASSUME-MORT-001]

Summary

This specification establishes the strategic Assumption Register, epistemic classification taxonomy, risk-uncertainty scoring matrix, and empirical validation experiment contracts for ai-mortgage-underwriting-pilot v1.0 under run ID mortgage-assume-001. It evaluates the feasibility of deploying an autonomous AI-driven mortgage underwriting agent to automate credit decisions on prime conforming loans up to $1.2M across 14,000 monthly applications ($8.5M capital initiative). It decisively halts ungrounded engineering commitments demonstrated in incident INC-4712 (where an unstated, unverified assumption that the IRS provides real-time digital income transcripts delayed loan closings for 1,200 home buyers by 22 days, incurring $3.8M in contractual rate-lock penalty payouts). The specification uncovers

five critical high-uncertainty assumptions, defines

reproducible empirical falsification experiments, maps

epistemic status (assumed vs observed vs decided), and establishes

binding Go/No-Go validation tripwires prior to capital deployment.

Detailed Description

Software and product initiatives frequently fail not because of poor engineering, but because systems were built upon unexamined, invalid foundational beliefs. Teams conflate aspirations ("investors will accept AI signatures") with empirical facts, baking fatal dependencies into architectures. Assumption discovery applies systematic epistemic hygiene: surfacing tacit beliefs, quantifying their potential catastrophic blast radius, and designing lightweight empirical experiments to validate or falsify them before writing production code.

Initiative Intake: Autonomous AI Underwriting ($8.5M Capital Budget)
                                │
                                ▼
[ Epistemic Extraction & Classification Gate ]
  ├── Categorizes Claims: `provided` | `observed` | `assumed` | `unknown`
  └── Filters Out Ungrounded Aspirations ("Assumptions Treated as Facts")
                                │
                                ▼
[ Risk-Uncertainty Scoring Matrix ]
  ├── Impact: Critical (Regulatory / Financial Survival)
  └── Uncertainty: High (Zero Empirical Observation to Date)
                                │
       ┌────────────────────────┴────────────────────────┐
       ▼ (Passes Falsification Test)                     ▼ (Experiment Fails: Falsified)
[ Validated Fact: Architecture Committed ]       [ Assumption Invalidated: Pivot Triggered ]
  Example: Experian Webhook Latency <= 800ms        Example: Fannie Mae Rejects AI Signature
  (Moves from `assumed` -> `observed`)              (Halts Automated Investor Syndication)

Criteria and weights

Comparison

Result

Option C is selected. All five high-risk assumptions must complete their assigned empirical validation experiment; proceeding to production software implementation is blocked until experiment results are logged.

Required Mechanisms

1. Assumption Scoring & Uncertainty Matrix [MC-SM-01]

2. Falsification Experiments & Validation Protocols [MC-FE-01]

Experiment EXP-01 (Falsifying ASM-01: Fannie Mae Acceptance)

Protocol: Submit a formal written Request for Interpretation (RFI) to Fannie Mae Capital Markets underwriting counsel containing 100 sample synthetic AI-adjudicated loan packages.

• Success Criteria: Written legal affirmation of eligibility under Fannie Mae Selling Guide Chapter B3-2.
• Go/No-Go Deadline: October 15, 2026.

Experiment EXP-02 (Falsifying ASM-02: IRS IVES Transcript Latency)

Protocol: Execute 500 automated IRS IVES requests during active banking hours using staging sandbox credentials; measure response time distributions.

• Success Criteria: 95% of transcripts returned within 180 seconds.

Falsification Threshold: If $> 10%$ of requests require $> 48$ hours, ASM-02 is

FALSIFIED, mandating an asynchronous underwriting workflow.

Experiment EXP-03 (Falsifying ASM-03: LLM Tax Extraction Precision)

Protocol: Run benchmark on 10,000 historic audited tax forms comparing OCR+LLM output against human ground-truth double-entry records.

• Success Criteria: False positive and hallucination rate $\le 0.05%$.

3. Epistemic Classification & Decision Tripwires [MC-DT-01]

• Epistemic State Transitions:
  $$\text{ASSUMED} \xrightarrow{\quad\text{Experiment Executed}\quad} \begin{cases} \text{OBSERVED (Validated)} \longrightarrow \text{Commit Production Build} \ \text{FALSIFIED (Invalidated)} \longrightarrow \text{Execute Pivot Plan} \end{cases}$$

The Capital Gate: Release of the $8.5M engineering capital budget is strictly locked until ASM-01, ASM-02, and ASM-03 transition to OBSERVED.

Invariants and Contracts

Mandatory Falsification Before Construction [INV-ASSUME-01]
  Assumptions with a Risk Score >= 16 must complete an empirical falsification experiment before engineering build starts.
  Allocating software development capacity to unvalidated critical assumptions is strictly prohibited.

Vendor Marketing Exclusion Invariant [INV-ASSUME-02]
  Third-party sales collateral or vendor documentation must never be classified as `observed` evidence.
  External dependencies must be verified through reproducible technical spikes.

Automated Go/No-Go Decision Tripwires [INV-ASSUME-03]
  If an empirical experiment breaches its defined falsification threshold, the associated project milestone
  must halt immediately. Continuing implementation with falsified assumptions is barred by governance.

Explicit Unknowns

• Legal turnaround time for Fannie Mae General Counsel review of autonomous underwriting algorithms (G-1).
• Cloud GPU inference compute cost volatility when processing 45-page commercial tax schedules (G-2).

Traceability

Verification

No validator was supplied, so no command was run.

Reviewer self-check against assumption discovery standards:

• Epistemic Rigor: PASS. Surfaced 5 hidden assumptions; separated beliefs from verified facts.
• Falsification Focus: PASS. Designed executable, concrete experiments with clear failure criteria.
• Capital Protection: PASS. Gated $8.5M spend behind empirical validation; INC-4712 failure mode barred.
• Markdown Hygiene: PASS. Native Markdown syntax strictly adheres to rule_markdown.md.

Open Decisions

• DEC-ASSUME-01: David O'Reilly to determine whether a partial human-in-the-loop fallback workflow should be designed in parallel while awaiting Fannie Mae legal determination (Owner: David O'Reilly).

Next steps

1. Elena Rostova dispatches the formal RFI dossier to Fannie Mae Capital Markets legal counsel (EXP-01).
2. Infrastructure team executes the 500-request IRS IVES API staging latency benchmark (EXP-02).
3. Machine Learning guild executes the 10,000-document extraction accuracy benchmark against historical tax records (EXP-03).