Calculation & validation methodology

AI performs document extraction and classification only: it reads a certification packet and proposes field values with a confidence score and a page reference. Deterministic code performs every calculation and every rule evaluation. No language model decides a compliance outcome.

• LIHTC (IRC §42)
• Project-Based Section 8 (HUD Handbook 4350.3)
• HOME (24 CFR Part 92)
• HOTMA income and asset provisions (24 CFR Part 5)
• Tax-exempt bond financed units

Accepted uploads are PDF and common image scans of tenant income certifications and their supporting verifications. Jurisdictional scope is the federal baseline nationwide; state-specific determinations require a validated state pack.

• Household composition and student status
• Employment, self-employment, benefit and other income sources
• Asset balances, imputed income and HOTMA asset thresholds
• Applicable income limit set, effective date and area
• Contract rent, tenant-paid utilities and the utility allowance schedule in force
• Set-aside election, applicable fraction and unit designation
• Third-party verification dates, signatures and expiration

Income and rent figures are computed in whole dollars using the rounding convention published by the controlling authority for that program. Every limit, allowance and rule is applied as of the effective date in force on the certification's effective date, not the date of review.

Any field required for the decision that falls below the configured confidence threshold must be human-verified before the engine will return Pass or Fail. Reviewer verification is recorded per field with the actor and timestamp.

Pass — every required input is present and verified to policy, and no evaluated rule failed.

Fail — every required input is present and verified to policy, and at least one evaluated rule failed. The finding cites the rule, version and correction path.

Unable to determine — required evidence is missing, below the confidence policy, unverified, or affected by an unresolved rule conflict, or the missing state rule could change the outcome. Final Pass/Fail and human sign-off are blocked, and the blocking reasons are logged with the review.

Rules are derived from archived copies of the controlling source document, stored with a SHA-256 content hash, publication date and effective range. A rule change creates a new version; prior versions are retained so any historical finding can be reproduced exactly. When a source becomes uncertain, the affected rules are suspended rather than guessed, and reviews that depend on them return Unable to determine.

• NCSHA state HFA directory — https://www.ncsha.org/housing-help/
• HUD state information — https://www.hud.gov/states
• Federal HOME regulations (24 CFR Part 92) — https://www.ecfr.gov/current/title-24/subtitle-A/part-92

1. Identify every controlling agency and program scope.
2. Archive the current QAP, compliance manual, forms, notices, limits and utility guidance with content hashes.
3. Extract candidate requirements with exact citations and effective dates.
4. Obtain independent compliance-expert review.
5. Build positive, negative, boundary, layered-program and supersession fixtures.
6. Require 100% expected results on deterministic release tests.
7. Record reviewer, approval, validation report and limitations.
8. Activate prospectively; retain previous versions.
9. Monitor official sources and suspend affected rules when uncertainty arises.

Release testing uses fixture sets covering positive, negative, boundary, layered-program and supersession cases. Ground truth is established by a credentialed affordable-housing compliance professional independent of the engineer who authored the rule. Deterministic release tests must return 100% of expected results before a pack is activated.

Benchmark in progress. No per-field precision or recall figure, end-to-end false-pass or false-fail rate, or time-savings estimate is published because no reproducible benchmark supports one yet. Figures will appear here with the benchmark date, engine build, dataset description and known limitations attached.

Rule and engine changes are reviewed, versioned and released prospectively. The full fixture suite runs on every change; a regression blocks release. Each completed review stores an evidence manifest containing the extracted inputs, the deterministic calculation trace, the rule versions and source hashes, the outcome, the engine and model versions, and every human action.

CertivoIQ does not replace your compliance staff, your state agency's determination, or legal advice. Regulatory and financial consequences vary by program, jurisdiction and deal structure, and CertivoIQ does not guarantee avoidance of findings, penalties or credit recapture.