Build a Trust-Safe Citation Verifier.

A citation can be malformed, point to no source, identify the wrong source, quote inaccurately, or fail to support the proposition. Those are different questions and require different evidence.

A structural parser can reject an impossible reporter series. It cannot establish that a well-formed citation exists, remains good law, or supports the sentence.

Practice: Classify ten supplied examples by failure layer and write what evidence would resolve each one.

Define accepted inputs, covered citation families, approved sources, possible verdicts, evidence returned, abstention rules, and the person authorized to release work.

If only existence was checked, the result must not imply proposition support or current precedential status.

Practice: Complete a one-page contract with an explicit non-goal for every promised capability.

Map the document from input through parsing, retrieval, model processing, human review, export, logs, and deletion. Identify every external service and every place confidential material could persist.

Use synthetic or public material until authority, contract, account settings, retention, and supervision are resolved.

Practice: Draw data-flow records for browser-local and connected builds; mark every boundary where information leaves the device.

Store document ID, claim text, citation string, normalized citation, source snapshot, located passage, verdict, confidence, limitations, reviewer, timestamps, and version.

Keep machine output separate from human disposition. Never overwrite the evidence that produced an earlier result.

Practice: Create one known-good and one known-bad record another reviewer can reproduce.

Start with a narrow citation family and deterministic patterns. Return the matched text and parsed fields so a learner can see exactly what the parser recognized.

Unknown formats should become unparsed review items, not disappear. Measure extraction recall on the document types you expect.

Practice: Test a sample plus five boundary cases and record misses without changing the labels.

Capture enough context for review while preserving the exact proposition being tested.

When one sentence contains multiple claims or citations, split the review units or mark the mapping ambiguous. Do not let proximity silently become support.

Practice: Annotate a paragraph with three claims and two citations; explain every mapping boundary.

Use official court or government material when the task requires an authoritative rule or filing. Databases and aggregators have different coverage, licensing, status, and citator boundaries.

Record which source answered which question. A source that establishes existence may not establish current validity.

Practice: Create an approved-source hierarchy for one jurisdiction and define the fallback.

Normalize the query, preserve the original citation, log the source and retrieval time, and return zero, one, or multiple candidates without forcing a match.

Store a permitted snapshot or stable reference. Respect authentication, rate limits, licensing, and redistribution restrictions.

Practice: Design responses for no match, one match, and ambiguous matches.

Retrieve candidate passages around meaningful terms, quotations, headings, and citations. Preserve page or paragraph locators.

Retrieval is candidate generation. The best-scoring passage is not proof that the proposition is supported.

Practice: Return the strongest candidate and a plausible distractor for three propositions.

Test who, what, when, jurisdiction, conditions, degree, and exceptions. A source may support a narrower proposition while the draft overstates it.

Require quoted source text, identify mismatches, and abstain when context is incomplete.

Practice: Rewrite three overstated claims to the strongest formulation their sources support.

Use labels such as structurally invalid, no source located, identity ambiguous, supported, supported with qualification, contradicted, unsupported, and not verified.

Confidence describes the basis for a provisional result; it does not convert missing evidence into support.

Practice: Assign verdicts to twelve cases and write the permitted next action.

Put the proposition, original citation, normalized record, source link, passage, verdict, reason, unresolved issue, reviewer status, and version in one place.

Separate machine suggestion, reviewer decision, and release decision. Make changes traceable.

Practice: Create a one-page report that lets a supervisor disagree without rerunning the system.

ABA Formal Opinion 512 applies existing competence, confidentiality, communication, supervision, candor, and fee duties to lawyers using generative AI. Rule 11 separately governs representations made by presenting a paper in federal civil litigation.

The verifier assists inquiry. It does not become the lawyer, editor, researcher, or accountable releaser.

Practice: Identify the approver, prohibited inputs, escalation triggers, and release evidence for a high-stakes filing.

Record the document hash or version, tool version, ruleset, sources checked, timestamps, results, overrides, reviewer, and release status.

Log enough to reconstruct the decision without retaining confidential content longer than approved.

Practice: Draft an audit record and retention rule.

Construct cases for valid citations, malformed forms, nonexistent authorities, wrong metadata, inaccurate quotations, overstatement, contradiction, ambiguous mapping, unavailable sources, and unsupported formats.

Keep test labels independent of the system and document the coverage boundary.

Practice: Create at least twenty labeled cases, including cases designed to expose false passes.

Report results by failure class. False passes create unwarranted assurance; false flags create review burden and may hide systematic gaps.

Do not publish one accuracy number without the dataset, denominator, labels, version, coverage, and error distribution.

Practice: Calculate a confusion matrix for each verdict family and justify a release decision.

A local demonstration, connected research pilot, governed workflow, and institutional system have different evidence, security, integration, and operational requirements.

Release only the capability level actually tested. Document exclusions and conditions required to advance.

Practice: Write a deployment memo with status, permitted users and data, limitations, and rollback trigger.

Present the contract, architecture, evidence, test results, limitations, human gate, and change plan. Demonstrate one success and one abstention.

A credible defense explains what the system cannot establish and shows what happens when evidence is missing.

Practice: Record a ten-minute capstone defense using only evidence in the project package.