On the morning of a fair-lending examination at a mid-sized regional bank in Nashville, the lead examiner asked a single opening question that stopped the compliance director mid-sentence: "Before we look at your disparity statistics, tell me about the less-discriminatory-alternative search your team conducted when you deployed this underwriting model." The compliance director, a twelve-year fair-lending veteran, knew exactly what the examiner was referring to. She also knew, with the kind of sinking certainty that only comes from institutional memory, that the bank had never documented one. The model had been purchased from a well-regarded vendor in 2024. It had been validated by a third-party consultant. It had passed a pre-deployment disparate-impact screen. But the documented search for a less-discriminatory alternative (LDA) that could have achieved the same credit risk accuracy with less adverse impact on protected classes? That document did not exist. What followed was not the exam the compliance team had expected. It was an exam organized around the legal and regulatory implications of that absence, and it changed how the bank governed its AI models for years afterward. This lesson is about building that documentation before the examiner arrives, understanding why the LDA search is the linchpin of the legal defense, and constructing the specific, evidence-based record that converts a disparate-impact finding from a liability into a defensible decision.

Why the LDA Search Is the Linchpin of the Disparate-Impact Defense

To understand why the LDA (less-discriminatory alternative) search is the linchpin of the disparate-impact defense, it helps to trace the legal argument from its origin through the specific obligation it creates for a lender using AI in credit decisioning.

Disparate impact in lending operates through a three-step burden-shifting framework applied to credit under ECOA (Equal Credit Opportunity Act) and the Fair Housing Act. For FHA claims, the Supreme Court confirmed the doctrine's availability in Texas Department of Housing and Community Affairs v. Inclusive Communities Project, 576 U.S. 519 (2015). Regulation B (Reg B, 12 CFR Part 1002) is the CFPB's implementing regulation for ECOA. In Step 1, the examiner or plaintiff establishes a prima facie case: a facially neutral lending practice produces a statistically significant adverse outcome for a protected class. In Step 2, the burden shifts to the lender to establish business necessity: the challenged practice is required to achieve a legitimate, substantial credit-risk objective. In Step 3, the examiner or plaintiff may defeat the business-necessity defense by showing that a less-discriminatory alternative exists that would serve the same objective with less adverse impact.

The legal power of Step 3 is asymmetric in a way that many compliance programs do not fully account for. The lender's Step 2 defense, business necessity, can be documented and quantified: the institution can show that a particular feature improves model accuracy by a measurable amount, and that removing it would reduce that accuracy by a specific, quantified degree. That is a defensible record. But the Step 3 counter-argument defeats that defense not by challenging the accuracy measurement but by showing that the objective could have been achieved differently, with less discriminatory impact. If a less-discriminatory alternative exists and the lender never looked for one, the lender loses the case not because its business-necessity argument was wrong but because the search was never done.

This asymmetry is the reason the LDA search must precede the exam, not respond to it. A search assembled after a finding is challenged on its face: the examiner or plaintiff can argue that the analysis was conducted to justify a decision already made rather than to genuinely evaluate alternatives. A search conducted before deployment, at model modification, and on a scheduled cycle, with contemporaneous documentation, is evidence that the institution made a genuine, good-faith inquiry into whether a less-discriminatory alternative was available and concluded, on the evidence, that none existed or that the most accurate available option represented the best balance of fairness and credit-risk performance.

The LDA search is not a compliance checkbox. It is the documented evidence that you looked, found what you found, and chose the most defensible option available. Without the search, the defense does not exist.

OCC Bulletin 2026-13, the April 2026 interagency model-risk guidance that superseded OCC 2011-12 and pulled AI and generative AI under model-risk, fair-lending, third-party, and board-governance expectations, makes the LDA obligation concrete. The bulletin's framework for ongoing model monitoring includes fair-lending risk as a dimension of model risk, which means the LDA analysis is not just a fair-lending compliance document but a component of the model-risk record that board and senior management are expected to oversee. An institution whose model-risk file contains a deployment validation but no LDA documentation is incomplete by the 2026 standard, and an examiner with access to both the model-risk file and the fair-lending examination file will note the gap.

What the LDA Search Requires: Constructing the Legal Defense

Building the LDA documentation requires understanding what it must accomplish legally before building the operational process around it. The documentation serves three legal functions: first, it establishes that a good-faith search was conducted (defeating the argument that the institution never evaluated alternatives); second, it establishes the standard of comparison (what alternatives were evaluated, and why); third, it establishes the conclusion and its basis (why the chosen model configuration represents the best available balance of accuracy and fairness). Each function requires specific content.

The scope of the search: what must be evaluated. The LDA search is not an unlimited obligation to test every conceivable model configuration. It is an obligation to evaluate alternatives that were reasonably available to the institution at the time of the decision. The relevant question is: given the state of credit-risk modeling at the time of model deployment or modification, what alternative configurations existed that might have reduced the measured disparate impact? This is a practical, not theoretical, inquiry. Common scope elements include:

Alternative feature sets: If a feature with high disparate-impact contribution can be replaced by a less-discriminatory feature with similar predictive power, that alternative must be tested. For example, if zip code (a high-correlation proxy for race and national origin) is used as a predictor and census-tract-level median income or unemployment rate could substitute with similar predictive accuracy but lower disparate impact, the substitution must be evaluated.

Threshold alternatives: If the model applies a binary approve/decline threshold, alternative threshold placements must be tested to determine whether a threshold adjustment would reduce the adverse-outcome disparity without a material reduction in predictive accuracy. A threshold that is set at the optimal accuracy point may produce more disparate impact than a threshold set slightly below or above that point; the LDA analysis must evaluate whether the accuracy trade-off at the lower-disparity threshold is acceptable.

Fairness-constrained model alternatives: In 2026, fairness-constrained machine learning is a mature methodological category, with algorithms that explicitly optimize for a combination of predictive accuracy and outcome equality. The LDA search should evaluate whether a fairness-constrained model achieves comparable accuracy to the baseline model, because the existence of a published fairness-constrained methodology that achieves comparable accuracy is the kind of readily available alternative that Step 3 was designed to surface.

Post-hoc correction alternatives: Some disparity can be addressed through post-hoc correction at the decision stage rather than model retraining, such as threshold calibration across demographic subgroups. The LDA search should evaluate whether post-hoc correction achieves disparity reduction comparable to model replacement, because post-hoc correction may be lower-cost and faster to implement, which affects the business-necessity analysis.

The standard of comparison: what counts as a "sufficiently accurate" alternative. The LDA framework does not require adopting any alternative that reduces disparity, regardless of accuracy cost. It requires adopting an alternative that is "equally effective." The question of what counts as equally effective is where the business-necessity analysis and the LDA analysis merge: an alternative that reduces disparity but also materially reduces the institution's ability to predict credit risk is not equally effective. The challenge is defining "material" in a way that is specific, quantified, and defensible.

The lesson here is to work from the institution's credit policy. If the policy requires a model that produces a minimum AUROC (area under the receiver operating characteristic curve) of X to support responsible lending decisions, then an alternative that produces an AUROC below X is not equally effective by policy definition. Document the policy requirement, the threshold, and the tested alternatives' performance against that threshold. This anchors the "equally effective" standard to an institutional policy rather than to a case-by-case judgment, which is more defensible under examination.

The conclusion and its basis: documenting the outcome. The LDA documentation must conclude with a specific finding: either (a) no less-discriminatory alternative was found that meets the equally-effective standard, with the tested alternatives, their disparity profiles, and their accuracy profiles documented; or (b) a less-discriminatory alternative was identified that meets the equally-effective standard, with the institution's plan to adopt it documented. An inconclusive finding, one that simply reports the testing without reaching a conclusion, is not a defensible LDA document because it does not demonstrate that the institution made a good-faith determination.

When to Conduct the LDA Search: The Timing Obligation

The timing of the LDA search is as important as its content, because the legal defensibility of the search depends significantly on when it was done. Three timing obligations apply to an institution using AI in credit decisioning:

At model deployment. Every AI model deployed in credit decisioning should have a completed LDA search in its model-risk file before it goes into production. Even if the pre-deployment disparate-impact test shows no material disparity, the proactive LDA search documents that the institution evaluated available alternatives and deployed the model configuration representing the best available balance of accuracy and fairness. This proactive documentation is stronger evidence of good-faith compliance than a search initiated only when a disparity is found.

The argument against proactive LDA testing is typically cost: testing multiple model configurations requires modeling resources, data science capacity, and time. This argument should be evaluated against the alternative cost, which is the cost of conducting the LDA search reactively after a finding, while under examination scrutiny, without the institutional credibility that comes from a contemporaneous pre-deployment analysis. At most well-resourced institutions, the cost of a proactive LDA search at deployment is a fraction of the cost of the remediation and enhanced examination that follows an LDA documentation gap.

At model modification. Any modification to a model that could affect its disparate-impact profile requires an updated LDA analysis. Model modifications include: changes to the feature set (adding or removing variables), changes to model thresholds or decision boundaries, changes to model weights resulting from retraining on updated data, and changes to the decision logic downstream of the model's output (for example, adding a human review step for a specific score range). The LDA documentation is model-version specific: a search conducted on Version 1.0 does not cover Version 1.1 if Version 1.1 includes a retrained model with different feature weights.

In practice, this means the institution's model change management process should include a required LDA review step. When a change request is submitted for an AI credit model, the review checklist should include: does this change potentially affect the model's disparate-impact profile? If yes, an LDA review is required before the change goes into production. This is a process design decision, not just a compliance decision, and it belongs in the institution's model risk management framework alongside the standard technical validation steps.

On a scheduled review cycle. The LDA analysis, unlike the disparate-impact test itself, is not solely dependent on the test results for its timing. An LDA analysis conducted at deployment in 2024 may be superseded by developments in credit-risk modeling by 2026: new techniques, new available data sources, new fairness-constrained algorithms. The institution should schedule periodic LDA reviews that ask: given the current state of credit-risk modeling, are there less-discriminatory alternatives available today that were not available when the last LDA search was conducted? Annual review is a defensible minimum; institutions with rapidly evolving model stacks may benefit from more frequent reviews.

The Five-Step LDA Process: Building the Documentation

The LDA search process translates the legal requirements described above into a structured, repeatable analytical workflow with specific documentation outputs at each step. Each step generates a document or evidence element that contributes to the LDA file the institution will produce under examination.

Step 1: Establish the baseline disparity measurement and identify candidate features.

Before the LDA search can identify alternatives, it must establish what it is trying to improve: the model's measured disparate impact on each protected class in the institution's market. The baseline measurement should use the same methodology as the institution's regular disparate-impact testing (univariate disparity analysis and multivariate regression), so that the LDA analysis is directly comparable to the testing cycle that triggered it (or to the pre-deployment test if the LDA search is proactive).

Alongside the baseline measurement, the step produces a feature-level attribution analysis using SHAP values (SHapley Additive exPlanations, a method adapted to machine learning from cooperative game theory), permutation importance, or conditional independence testing. The attribution analysis identifies which features contribute most to the measured disparity for each protected class: these are the candidate features for the LDA search. The baseline document should report: the adverse action rate ratio and disparity index for each protected class, the multivariate regression residual disparity, and the top-five (or top-ten for large models) features by disparate impact contribution, with their SHAP magnitude and the demographic subgroup breakdown of their effect.

This document is the LDA search's starting point and establishes its scope. An LDA search that does not begin with this analysis has no principled basis for deciding which alternatives to test.

Step 2: Identify and define the alternative model configurations to test.

Based on the candidate features identified in Step 1, the LDA search defines the specific alternative configurations to evaluate. Each alternative should have a clear hypothesis: what disparity reduction is expected from this change, and what accuracy cost is expected? Without a hypothesis, the testing is exploratory rather than purposeful, and exploratory testing is harder to document as a good-faith search.

Common alternative categories for credit model LDA searches include:

Feature removal alternatives: Remove the highest-disparity-contribution feature and retrain the model on the remaining features. Measure the resulting disparity profile and accuracy metrics. Repeat for each of the top-five candidate features, individually and in combinations.

Feature substitution alternatives: Replace the highest-disparity-contribution feature with a correlated-but-less-discriminatory proxy. For example, replace zip code with census-tract-level unemployment rate or median income if those features have lower correlation with protected class characteristics while retaining predictive power. This requires access to alternative data sources and a hypothesis about why the substitute is less discriminatory.

Threshold adjustment alternatives: Test the model at thresholds above and below the current decision boundary, measuring the change in disparity profile at each threshold point. Plot the Pareto frontier of accuracy versus disparity across the threshold range to identify whether a threshold adjustment achieves meaningfully lower disparity without material accuracy loss.

Fairness-constrained alternatives: Train a version of the model using a fairness-constrained algorithm that explicitly penalizes disparate-impact outcomes during training. Evaluate whether the fairness-constrained model achieves accuracy comparable to the baseline within the institution's minimum AUROC or similar threshold.

For each alternative, document the hypothesis, the specific configuration (which features were removed, substituted, or modified, or what algorithm was used), and the model version identifier that will be used in the testing.

Step 3: Execute the testing and record results for each alternative.

Each alternative model configuration is trained (or adjusted) and tested on the institution's loan application data using the same testing methodology as the baseline measurement. The testing produces a structured results table for each alternative showing:

Accuracy metrics relative to baseline: AUROC, Kolmogorov-Smirnov statistic, Gini coefficient, or the institution's primary credit performance metric. The comparison must be apples-to-apples, using the same sample, the same time period, and the same performance definition as the baseline measurement.

Disparity metrics relative to baseline: adverse action rate ratio and disparity index for each protected class tested. Again, apples-to-apples against the baseline measurement. A disparity reduction that is not measured using the same methodology as the baseline disparity finding is not a valid LDA result.

Statistical significance of both the accuracy and disparity changes: a change in AUROC from 0.85 to 0.84 may or may not be statistically significant depending on the sample size. A change in the adverse action rate ratio from 1.8 to 1.5 may or may not be a practically meaningful improvement. Both dimensions of the result need statistical context, not just point estimates.

The results table is the empirical core of the LDA documentation. It provides the evidence base for the business-necessity analysis in the next step.

Step 4: Conduct the business-necessity analysis for each alternative.

The business-necessity analysis evaluates each alternative's accuracy cost and asks: is this accuracy cost acceptable under the institution's credit policy and risk appetite? The analysis has three layers.

Layer 1 is the policy threshold check: does the alternative meet the institution's minimum model accuracy requirements as defined in the credit policy? If the policy requires a minimum AUROC of 0.80 and an alternative achieves 0.77, the alternative fails the policy threshold and the institution can document this as the reason for rejection without needing to conduct the full financial impact analysis.

Layer 2 is the financial impact quantification: for alternatives that meet the policy threshold, what is the expected financial cost of the accuracy reduction? This requires translating the AUROC or KS-statistic reduction into estimated additional default losses using the institution's historical performance data. If removing zip code reduces the model's AUROC from 0.83 to 0.81, and historical data shows that an AUROC reduction of this magnitude corresponds to an expected increase in annual default losses of approximately $X, that $X figure is the business-necessity argument. It must be specific, derived from the institution's data, and documented in the analysis.

Layer 3 is the proportionality assessment: even if the alternative imposes an accuracy cost, is that cost proportionate to the disparity benefit? An accuracy cost of $2 million per year against a disparity reduction that reduces the adverse action rate ratio for a protected class from 1.1 to 1.0 is a different proportionality calculus than the same $2 million cost against a disparity reduction from 1.9 to 1.1. The institution does not need to formally calculate a disparity-per-dollar figure, but the proportionality assessment should be documented qualitatively: "The accuracy cost of Alternative 3 is proportionately large relative to the marginal disparity reduction from a 1.1 to 1.05 ratio" is a documented judgment. "The accuracy cost is too high" is not.

Step 5: Document the selection decision and schedule the next review.

The LDA analysis concludes with a documented selection decision that specifies: the model configuration selected (baseline or alternative), the basis for the selection, and the specific alternatives rejected, with the documented reasons for their rejection. The document should be signed by the responsible compliance officer, the model-risk officer, and (where required by the governance policy) a senior management representative.

The timing of the next LDA review should be scheduled at the conclusion of the current search, and the review date should appear in the model inventory. If the institution retained the baseline model over an alternative on business-necessity grounds, the next LDA review should evaluate whether the business-necessity rationale still holds given any changes in modeling technology or available data since the last search. Technology evolves; a feature that was necessary for accuracy in 2024 may be replaceable in 2026.

The Role of AI Tools in the LDA Process

AI tools can accelerate several components of the LDA search, most importantly the model training and testing components (Step 3), where automated machine learning pipelines can train and evaluate dozens of alternative configurations in hours rather than days. This acceleration is genuinely valuable: it allows the institution to test a more comprehensive set of alternatives than would be feasible with manual model builds, which strengthens both the quality of the LDA analysis and the defensibility of the conclusion.

But AI tools introduce specific risks in the LDA context that must be managed:

The selection-bias risk in automated search. An automated LDA search that uses an AI-driven hyperparameter optimization algorithm to find the most accurate alternative model configuration may converge on configurations that optimize accuracy subject to the algorithm's constraints, but may not be designed to minimize disparate impact as a co-objective. The automated search must be explicitly designed to evaluate both accuracy and disparate impact across the configurations tested, not just to maximize accuracy among configurations that happen to reduce disparity.

The documentation incompleteness risk. Automated machine learning pipelines are excellent at generating model configurations and measuring their performance. They are typically less good at generating the narrative documentation required for the LDA file: the hypothesis for each alternative, the business-necessity analysis, the proportionality assessment, and the selection rationale. These narrative elements require human judgment and institutional knowledge that automated pipelines do not possess. The institution's LDA documentation must include these narrative elements, not just the quantitative results table produced by the pipeline. An LDA file that consists entirely of model performance tables without business-necessity analysis is incomplete by the legal standard.

The hallucination risk in narrative generation. If generative AI tools are used to draft the narrative portions of the LDA documentation, every factual claim in those drafts must be verified against the actual testing results before the document is finalized. A generative AI tool that drafts a business-necessity analysis based on hypothetical performance figures rather than the actual testing results produces a false LDA document, regardless of how professional the language sounds. The institution's compliance and model-risk officers must verify every specific claim (accuracy metrics, disparity measurements, financial impact estimates) against the underlying data before signing off on the LDA documentation.

The practical AI-assisted LDA workflow uses automated machine learning to train and evaluate alternative configurations (Step 3), uses AI-assisted attribution tools (SHAP, permutation importance) for the baseline feature analysis (Step 1), and uses generative AI to draft the narrative documentation. Human compliance and model-risk analysts verify the narrative against the quantitative results, conduct the business-necessity and proportionality analysis (which requires policy knowledge and institutional context that AI tools lack), and write the final selection decision and rationale.

What an Examiner Looks for in the LDA Documentation

Understanding what an examiner is looking for in an LDA documentation review helps institutions build documentation that is organized and structured for examination, rather than for internal convenience. Based on the legal framework and OCC 2026-13's governance expectations, examiners reviewing LDA documentation are typically looking for the following elements, in this order:

Evidence that the search was conducted before deployment, not in response to a finding. Contemporaneous dating is important. An LDA document dated two months before model deployment, with version-controlled components that show the testing was conducted on pre-deployment data, is more credible than a document assembled after a disparate-impact finding. The model-risk file should include the LDA search as a pre-deployment validation component, timestamped and version-controlled alongside the other validation materials.

A clear scope: what alternatives were tested and why. The examiner wants to understand how the institution decided which alternatives to evaluate. An LDA search that tested only one alternative and concluded that no less-discriminatory alternative exists is less convincing than one that tested five or ten alternatives using a principled selection process based on the feature attribution analysis. The scope section of the documentation should explain: we identified these features as the primary drivers of disparity, we evaluated these specific alternative configurations because they represent the available approaches to reducing that disparity, and we used this methodology for each.

Apples-to-apples comparison between baseline and alternatives. The examiner will compare the baseline disparity measurement to the alternative disparity measurements using the same metric, the same sample, and the same time period. Inconsistencies in methodology (different sample sizes, different time periods, different disparity metrics) raise questions about whether the search was designed to find alternatives or to exclude them.

A quantified business-necessity analysis, not an assertion. "Removing this feature would hurt our model" is not a business-necessity analysis. "Removing this feature reduces our AUROC from 0.84 to 0.81, which, based on our historical performance data, corresponds to approximately $1.8 million in additional annual default losses at our current origination volume" is a business-necessity analysis. The examiner expects numbers, derived from the institution's own data, not from vendor benchmarks or industry averages.

A documented selection decision with a named, responsible officer. The selection decision should be specific: "We are retaining the baseline model configuration for the following reasons, and will revisit this decision at the next scheduled LDA review on [date]." The decision should be signed by the officer responsible for fair-lending compliance and the model-risk officer, establishing that a human made the compliance determination based on the LDA evidence.

A scheduled next review date. An LDA analysis with no next-review date is an analysis that treats the current conclusion as permanent. The examiner will note the absence of a next-review date because it suggests the institution does not have an ongoing process for re-evaluating the LDA conclusion as modeling technology evolves.

When the examiner in Nashville asked about the LDA search and the compliance director could not produce one, the examination's trajectory changed because the absence of the search was itself the finding. The exam was no longer about whether the bank's model was discriminatory; it was about whether the bank's AI governance program met the standard that OCC Bulletin 2026-13 requires. The answer, without an LDA document, was plainly no. That one document gap, which would have taken perhaps two weeks of modeling work and two more weeks of documentation to produce before deployment, cost the bank far more than its absent cost. The lesson is straightforward: build the LDA search into the model deployment workflow, document it thoroughly, sign it before the model goes into production, and update it when the model changes. The exam will come eventually. The document should be ready before it does.

The LDA Search in the Model Risk Governance File

The LDA search belongs in the model-risk governance file alongside the model validation documentation, not in a separate compliance filing that is disconnected from the model's technical history. This structural placement matters because OCC Bulletin 2026-13 pulls fair-lending risk explicitly under the model-risk governance framework, meaning the model's fair-lending documentation and the model's technical validation documentation are part of a single model-risk record that board and senior management are expected to oversee together.

Practically, this means the model inventory entry for an AI credit model should include a reference to the LDA documentation, with the date of the most recent LDA search and its conclusion. The model change management process should include an LDA impact assessment as a required step whenever a change potentially affects the model's disparate-impact profile. The periodic model validation review should include a check on whether the LDA documentation needs to be updated. These integrations ensure that the LDA search is treated as a governance obligation, not as a one-time compliance task that can be deferred until an examination requires it.

The integration also has a practical examination benefit: when a fair-lending examiner and a model-risk examiner are reviewing the institution's AI credit governance program simultaneously (which OCC 2026-13's integrated examination approach produces), the LDA documentation that sits in the model-risk file is accessible to both examination teams. Its presence in the model-risk file signals that the institution treats fair lending as a dimension of model risk governance, not as a separate compliance obligation. Its absence from the model-risk file, conversely, signals the opposite, and that signal becomes the starting point for both examination teams' inquiry.

The institution that builds the LDA search into its model deployment workflow, documents it thoroughly, integrates it into the model-risk governance file, and updates it on a scheduled cycle has built what the fair-lending framework requires: evidence that the institution looked for a less-discriminatory alternative, found what it found, and made a defensible, documented decision. That evidence is the defense. The exam tests whether you have it.

Key Takeaways

• The LDA (less-discriminatory alternative) search is the linchpin of the disparate-impact defense because it addresses Step 3 of the legal framework: even if a lender demonstrates business necessity, an available less-discriminatory alternative that was never evaluated defeats that defense. An undocumented search is legally equivalent to no search.
• The LDA search must be proactive, conducted before deployment and at each model modification, not assembled in response to an examination finding. A contemporaneously documented search is stronger evidence of good-faith compliance than a retroactive one assembled under examination scrutiny.
• The five-step LDA process produces five documentation elements: the baseline disparity measurement with feature attribution, the defined alternative configurations with hypotheses, the quantitative testing results table, the business-necessity analysis with financial impact figures, and the selection decision with a named responsible officer and a scheduled next-review date.
• Business necessity must be quantified, not asserted. "Removing this feature would reduce our AUROC from 0.84 to 0.81, corresponding to approximately $X in estimated annual additional default losses at our current origination volume" is documentation. "It would hurt the model" is not.
• AI tools can accelerate LDA search components (automated model training, SHAP attribution, pipeline testing) but cannot substitute for human judgment in the business-necessity analysis, the proportionality assessment, and the selection decision. The LDA document must include human-authored narrative alongside AI-generated quantitative results, and all factual claims in AI-generated narrative must be verified against the actual testing data.
• The LDA documentation belongs in the model-risk governance file alongside the technical validation documentation, not in a separate compliance filing. OCC Bulletin 2026-13 treats fair-lending risk as a dimension of model risk, and the integrated examination approach means both model-risk and fair-lending examiners will look for LDA documentation in the model-risk record.
• Examiners look for six specific elements in an LDA document: evidence of pre-deployment timing, a clear scope showing which alternatives were tested and why, apples-to-apples comparison methodology, a quantified business-necessity analysis, a documented selection decision signed by a named responsible officer, and a scheduled next-review date.