{"schema_version":1,"research_id":"eoa_inverse_innovation_exp04_external_evaluation_20260802","source_assessment_id":"invariant_mode_decomposition_design__economics_finance:DISCARDED_AUDIT:H3:v0","cell_id":"invariant_mode_decomposition_design__economics_finance","search_queries":["site:consumerfinance.gov small business lending alternative data cash flow underwriting thin file official","site:federalreserve.gov small business credit survey credit application documentation burden approval discouraged borrowers","site:occ.treas.gov alternative data underwriting small business cash flow official","site:fdic.gov small business lending underwriting alternative data cash flow","FinRegLab small business cash flow underwriting empirical research applicant data documents underwriting","small business loan application document burden research required documents application time study","PCA SVD credit scoring feature reduction rare exceptions residual fairness underwriting paper","alternative data small business credit underwriting fair lending human review residual cases official guidance","site:consumerfinance.gov adverse action complex algorithms creditors specific reasons underwriting 2023 circular","site:occ.treas.gov model risk management validation fair lending underwriting alternative data guidance","small business lender reduced application fields bank data underwriting application minutes official company","PCA credit scoring dimensionality reduction loan application prediction study DOI","principal component analysis SME credit scoring feature reduction paper","fairness dimensionality reduction rare subgroup credit scoring PCA"],"sources":[{"source_id":"S1","title":"Community Banks: Effect of Regulations on Small Business Lending and Institutions Appears Modest, but Lending Data Could Be Improved (GAO-18-312)","publisher":"U.S. Government Accountability Office","url":"https://files.gao.gov/reports/GAO-18-312/index.html","source_class":"GOVERNMENT_OR_REGULATOR","publication_date":"2018-08-06","accessed_at":"2026-08-02","claims_supported":["A generalizable survey found that 79 percent of community banks increased documentation requirements for small-business borrowers during 2010–2017.","Documentation and credit-standard burdens were reported as particularly difficult for newer small businesses.","The report also cautions that low- or no-documentation lending can increase default risk, supporting the need for a validated safety constraint."]},{"source_id":"S2","title":"Modernizing Access to Credit for Younger Entrepreneurs: From FICO to Cash Flow","publisher":"National Bureau of Economic Research","url":"https://www.nber.org/papers/w33367","source_class":"PRIMARY_RESEARCH","publication_date":"2026-06 (revised; first issued 2025-01)","accessed_at":"2026-08-02","claims_supported":["Personal credit scores can be noisy and mechanically disadvantage younger entrepreneurs with shorter histories.","Incorporating business-account cash-flow data increased approvals for younger, lower-score entrepreneurs without evidence that the result reflected increased lender risk-taking.","Cash-flow data provide an independent underwriting signal for constrained small-business applicants."]},{"source_id":"S3","title":"Transforming Small Business Credit: Technology and Data Adoption in Mission-Based Lending","publisher":"FinRegLab","url":"https://finreglab.org/research/transforming-small-business-credit-technology-and-data-adoption-in-mission-based-lending/","source_class":"PRIMARY_RESEARCH","publication_date":"2025-06","accessed_at":"2026-08-02","claims_supported":["Mission-based lenders are already adopting electronic bank-account feeds and lending platforms for underserved small businesses.","Implementations reduced paperwork and shortened processing from months to days.","Some lenders are moving toward automation and machine-learning risk assessment, while operational, cultural, and validation challenges remain."]},{"source_id":"S4","title":"Interagency Statement on the Use of Alternative Data in Credit Underwriting","publisher":"Federal Deposit Insurance Corporation","url":"https://www.fdic.gov/news/financial-institution-letters/2019/fil19082.html","source_class":"OFFICIAL_GUIDANCE","publication_date":"2019-12-13","accessed_at":"2026-08-02","claims_supported":["Federal financial regulators recognize reliable cash-flow data as an alternative underwriting input.","Regulated financial institutions may use alternative data but must manage consumer-protection, compliance, and data-quality risks.","The statement supplies a credible regulatory pathway rather than a categorical authority prohibition."]},{"source_id":"S5","title":"Fundbox + FreshBooks: Powering an Embedded Line of Credit for Service-Based Small Businesses","publisher":"Fundbox","url":"https://fundbox.com/newsroom/blog/fundbox-freshbooks-embedded-line-of-credit/","source_class":"COMMERCIAL_FIRST_PARTY","publication_date":"2026-06-03","accessed_at":"2026-08-02","claims_supported":["An operating embedded-credit product already reuses revenue, invoice, and payment-history data held in FreshBooks.","The product advertises decisions in minutes with no separate paperwork, tax-return upload, bank-statement upload, or branch visit to get started.","This directly demonstrates both implementability and substantial prior-art collision on applicant-burden reduction through existing correlated financial records."]},{"source_id":"S6","title":"Knowledge Discovery Using Neural Approach for SME's Credit Risk Analysis Problem in Turkey","publisher":"Elsevier, Expert Systems with Applications","url":"https://www.sciencedirect.com/science/article/abs/pii/S0957417411000327","source_class":"PRIMARY_RESEARCH","publication_date":"2011-08","accessed_at":"2026-08-02","claims_supported":["PCA and factor analysis have already been tested as feature-extraction methods in SME credit-risk assessment.","The study used a real SME portfolio, dimensionality reduction, classification, and cross-validation.","On its limited 512-observation sample, supervised feature selection performed better than PCA-style extraction for achieving minimal input dimension, weakening any presumption that PCA is the best field-reduction mechanism."]},{"source_id":"S7","title":"Consumer Financial Protection Circular 2022-03: Adverse Action Notification Requirements in Connection with Credit Decisions Based on Complex Algorithms","publisher":"Consumer Financial Protection Bureau","url":"https://www.consumerfinance.gov/compliance/circulars/circular-2022-03-adverse-action-notification-requirements-in-connection-with-credit-decisions-based-on-complex-algorithms/","source_class":"GOVERNMENT_OR_REGULATOR","publication_date":"2022-05-26","accessed_at":"2026-08-02","claims_supported":["Creditors using complex models must still provide applicants with accurate, specific principal reasons for adverse action.","Model opacity is not a defense for noncompliance with ECOA and Regulation B.","Latent PCA or SVD modes would therefore require a defensible mapping back to actual scored factors and decision reasons."]},{"source_id":"S8","title":"Supervisory Guidance on Model Risk Management","publisher":"Board of Governors of the Federal Reserve System","url":"https://www.federalreserve.gov/frrs/guidance/supervisory-guidance-on-model-risk-management.htm","source_class":"OFFICIAL_GUIDANCE","publication_date":"2026-04-17","accessed_at":"2026-08-02","claims_supported":["Credit-underwriting models require risk-proportionate development, validation, monitoring, governance, and controls.","Validation should assess assumptions, methods, data, outcomes, acceptable performance ranges, and deterioration over time.","Third-party and customized models remain subject to validation and ongoing outcome analysis, supporting material implementation and recurring-governance costs."]}],"problem_evidence":{"support":"MODERATE","rationale":"Documentation and processing burdens in small-business lending are externally supported, and younger or short-history entrepreneurs can be disadvantaged by personal-credit-score reliance. Cash-flow evidence can improve assessment. However, no direct source establishes the hypothesis's specific failure mechanism that coordinate-level checklists systematically duplicate correlated records and discard low-variance, decision-critical exceptions.","source_ids":["S1","S2","S3"]},"stakeholder_evidence":{"support":"MODERATE","rationale":"Mission-based lenders, fintech lenders, and embedded-finance platforms demonstrate demand for less paperwork and richer transaction data. Named operating implementations establish credible adopters for adjacent workflows, but no source shows adopter pull for PCA/SVD modes, structured reconstruction residuals, or the proposed two-percentage-point subgroup constraint specifically.","source_ids":["S2","S3","S5"]},"prior_art":{"proximity":"SUBSTANTIAL_COLLISION","closest_analogues":[{"name":"Fundbox–FreshBooks embedded underwriting","similarity":"Uses financial records already held in an accounting platform to assess small firms with no separate paperwork and decisions in minutes, directly matching the burden-reduction objective and much of the workflow boundary.","remaining_difference":"The public description does not disclose PCA/SVD compression, structured residual routing, or explicit predictive and subgroup false-negative tolerances.","source_ids":["S5"]},{"name":"Mission-based lenders adopting electronic cash-flow feeds","similarity":"Lenders already use electronic bank data and platforms to reduce paperwork and processing time for underserved small businesses, with some moving toward automated or machine-learning underwriting.","remaining_difference":"The reported implementations do not specify modal decomposition, residual reconstruction thresholds, or an applicant-field reduction endpoint.","source_ids":["S3","S4"]},{"name":"PCA-based SME credit-risk feature extraction","similarity":"PCA and related feature-reduction methods have long been evaluated on real SME credit-risk data to obtain reduced representations for classification.","remaining_difference":"The research did not test applicant burden, structured human review of residual exceptions, or subgroup false-negative preservation; it also found a supervised feature-selection method better for minimal input dimension.","source_ids":["S6"]}],"distinctive_claim_remaining":"The narrow remaining distinction is a governed intake protocol that uses modal compression while explicitly routing structured reconstruction residuals to humans and jointly constrains field reduction, out-of-sample discrimination, and subgroup false-negative changes. No direct evidence establishes that this combination is already standard, but the as-written comparison against a full checklist does not isolate its advantage over mature electronic-data, embedded-underwriting, or supervised feature-selection baselines.","confidence":"HIGH"},"implementation_evidence":{"support":"MODERATE","rationale":"Electronic financial-data ingestion, paperless embedded underwriting, PCA credit-risk analysis, human exception handling, and model monitoring are individually feasible. The central implementation gap is that PCA/SVD components ordinarily require the original input variables to compute them; decomposition alone does not identify which applicant fields can be omitted. A deployable protocol would need external data feeds or a separately validated field-selection/reconstruction rule, plus explainable adverse-action mappings and statistically adequate subgroup validation.","source_ids":["S3","S5","S6","S7","S8"]},"scores":{"meaningful_impact":{"score":4,"rationale":"Reducing documentation for creditworthy small firms could materially reduce access friction, while cash-flow research suggests benefits for younger and lower-score entrepreneurs. The prevalence of avoidable duplicated fields remains unmeasured.","source_ids":["S1","S2","S3"]},"stakeholder_pull":{"score":3,"rationale":"Lenders demonstrably want lower-friction, data-driven underwriting, but evidence of pull is for electronic cash-flow and embedded products rather than the modal-residual protocol.","source_ids":["S3","S5"]},"incremental_advantage":{"score":2,"rationale":"The hypothesis benchmarks only against a full checklist. It does not establish an advantage over existing paperless cash-flow underwriting, embedded accounting-data products, or supervised feature selection.","source_ids":["S3","S5","S6"]},"distinctiveness_plausibility":{"score":2,"rationale":"Residual routing plus subgroup tolerances are a potentially distinctive governance package, but burden reduction through correlated digital records and PCA-based SME credit reduction already have substantial prior art.","source_ids":["S3","S5","S6"]},"technical_implementability":{"score":3,"rationale":"All component technologies are mature, but converting latent components into fields an applicant need not provide is not automatic and may undermine interpretability. The proposal needs a validated observation/field-omission mechanism beyond PCA/SVD itself.","source_ids":["S5","S6","S7","S8"]},"adoption_authority_feasibility":{"score":2,"rationale":"Regulators permit responsible alternative-data use, but a production credit-model change requires lender authority, model validation, fair-lending review, adverse-action reason mapping, and operational ownership of manual exceptions.","source_ids":["S4","S7","S8"]},"evidence_readiness":{"score":2,"rationale":"Relevant loan and cash-flow datasets exist, but no located evidence contains the full combination of requested-field logs, outcomes, protected-group labels, rare residual exceptions, and an incumbent digital-underwriting comparator needed to test the claim cleanly.","source_ids":["S2","S3","S6"]},"safety_net_benefit":{"score":4,"rationale":"If effective, structured residual review could protect applicants whose decisive evidence lies outside dominant modes and subgroup tolerances could prevent average-performance results from masking exclusion. This benefit remains prospective rather than demonstrated.","source_ids":["S1","S7","S8"]},"scalability":{"score":3,"rationale":"Digital ingestion and automated scoring scale well, as existing embedded products show, but lender-specific data schemas, validation, explanation requirements, drift monitoring, and the residual-review queue limit plug-and-play scaling.","source_ids":["S3","S5","S7","S8"]}},"score_confidence":"MODERATE","costs":{"first_evidence":{"band_2026_usd":"50K_TO_250K","scope":"Retrospective three-way replay on one lender's historical applications: full checklist, current digital/supervised baseline, and the as-written modal protocol; include field counts, out-of-sample discrimination, subgroup false-negative rates, residual-review yield, and confidence intervals.","confidence":"LOW","assumptions":["Existing application-level records, requested-field logs, outcomes, and lawful subgroup labels are available under a research agreement.","Approximately 10,000 or more applications are available, with enough defaults and subgroup observations for useful estimates.","Work is limited to offline analysis and does not affect live credit decisions."],"source_ids":["S2","S3","S6","S8"]},"initial_deployment_startup":{"band_2026_usd":"250K_TO_1M","scope":"Build a shadow-mode prototype, data connectors, field-omission mapping, residual queue, reviewer interface, reason-code mapping, documentation, independent validation, and compliance testing at one lender.","confidence":"LOW","assumptions":["One lender and one loan product are in scope.","Existing intake and decision systems expose usable APIs or batch exports.","No core-banking replacement or new nationwide data-acquisition program is required."],"source_ids":["S3","S5","S7","S8"]},"operational_launch":{"band_2026_usd":"1M_TO_5M","scope":"Controlled production launch for one lender, including security and vendor review, integration, model governance, fair-lending testing, reviewer staffing and training, monitoring, rollback controls, and applicant communications.","confidence":"LOW","assumptions":["Launch covers one product and a limited geography before expansion.","The lender bears integration and governance costs but already has underwriting and compliance teams.","Manual residual cases remain a minority of applications."],"source_ids":["S3","S4","S7","S8"]},"annual_recurring":{"band_2026_usd":"250K_TO_1M","scope":"Data-access fees, residual-review labor, model monitoring, periodic outcome and subgroup analyses, independent validation, audit support, drift response, and reason-code maintenance for one product.","confidence":"LOW","assumptions":["Application volume is moderate and residual referral rates remain bounded.","No major redevelopment or multi-product rollout is included.","Existing compliance and data infrastructure can absorb part of the workload."],"source_ids":["S3","S5","S7","S8"]}},"verified_pipeline_gates":{"externally_supported_problem":{"status":"YES","reason":"Independent government and primary research support small-business documentation burden and disadvantages associated with short credit histories, although the hypothesized correlation/residual mechanism is not directly established.","source_ids":["S1","S2"]},"externally_credible_adopter_or_authorizer":{"status":"YES","reason":"Mission-based lenders and fintech platforms are credible adopters of adjacent low-paperwork underwriting, and interagency guidance provides a regulatory pathway for responsible alternative-data use.","source_ids":["S3","S4","S5"]},"distinct_testable_incremental_claim":{"status":"NO","reason":"The numerical claim is falsifiable, but as written it compares the package only with a full checklist and therefore cannot distinguish modal residual governance from already deployed electronic-data intake, embedded underwriting, or conventional supervised feature selection.","source_ids":["S3","S5","S6"]},"bounded_next_evidence_step":{"status":"YES","reason":"A retrospective, single-lender, three-way replay can test field reduction, predictive performance, subgroup false negatives, and residual-review yield without changing live decisions.","source_ids":["S2","S3","S6","S8"]},"no_unresolved_safety_or_authority_stop":{"status":"UNCERTAIN","reason":"There is no categorical prohibition, but unresolved issues include whether latent modes can generate accurate adverse-action reasons, whether omitted fields can be recovered safely, whether rare subgroup outcomes are estimable, and who validates human residual overrides.","source_ids":["S4","S7","S8"]},"credible_cost_scope_and_range":{"status":"YES","reason":"A credible stage-bounded scope can be specified from offline replay through one-product deployment, although ranges remain low-confidence because no partner architecture, application volume, residual rate, or vendor pricing is available.","source_ids":["S3","S5","S7","S8"]}},"next_evidence_step":"Secure one lender's historical application cohort and preregister a three-way offline comparison of the full checklist, the lender's strongest current digital or supervised-selection baseline, and the unchanged modal protocol. Require at least 25% fewer applicant-supplied fields, predefine the discrimination metric and two-percentage-point equivalence margins, report confidence intervals for each sufficiently powered subgroup, audit which omitted fields are still needed to compute each mode, and have blinded underwriters classify whether structured residual referrals recover genuinely decision-critical exceptions.","blocking_evidence":["No direct prevalence estimate for creditworthy small firms whose applications fail specifically because correlated document requirements or low-variance evidence are mishandled.","No evidence that PCA/SVD components can be computed after omitting at least 25% of applicant-supplied fields without requiring equivalent information through another channel.","No head-to-head evidence against current electronic cash-flow, embedded accounting-data, or supervised feature-selection underwriting.","No demonstrated sample size or subgroup-outcome coverage sufficient to establish a two-percentage-point false-negative equivalence margin.","No validated mapping from latent modes and residual referrals to accurate adverse-action reasons and auditable human-review decisions.","No direct adopter commitment to test the modal-residual layer rather than deploy established paperless underwriting tools."],"research_disposition":"PRIOR_ART_DIFFERENTIATION_STUDY","world_novelty_boundary":"Alternative-data underwriting, reuse of accounting or bank records to remove paperwork, automated credit decisions with human exception handling, and PCA-based SME credit-risk reduction are not world-novel. The only plausible novelty boundary is the combined use of explicit modal reconstruction residuals, human routing, and simultaneous field-count, predictive, and subgroup equivalence constraints. That combination was not found as an established practice, but its incremental value and technical coherence are unverified.","arm":"DISCARDED_AUDIT","candidate_version":0,"controller_recommendation":{"action":"STOP_EMPIRICAL_RESEARCH_NEEDED","repairable":true,"material_progress_observed":false,"progress_targets":["Differentiate the claim from mature paperless cash-flow and embedded-underwriting baselines.","Demonstrate how omitted applicant fields can be avoided rather than merely recombined into PCA/SVD components.","Establish adequately powered out-of-sample and subgroup equivalence tests.","Validate explainable adverse-action reasons and governance for residual human review."],"reason":"The hypothesis does not satisfy the preregistered strict differentiated-opportunity endpoint as written. The problem and adopter pathway are credible, but prior art substantially collides with the burden-reduction mechanism, the comparator does not isolate a distinct incremental advantage, and safety, explainability, and subgroup-power questions remain unresolved. A bounded retrospective study could repair the evidence gap, but the current record does not support pilot or adoption inquiry."}}