Resolving conflicting evidence in AI-enabled work

Define the exact conflict

Write the competing propositions in comparable terms. Check whether they use the same population, period, unit, definition, version, geography, and decision context. Many apparent conflicts are scope differences: one figure may cover bookings while another covers recognized revenue, or one policy may apply only to a specific region.

Identify the decision that depends on resolution and the consequence of being wrong. Rank source authority for that question before evaluating which answer is more convenient. Preserve the original evidence, including dates and versions, rather than copying only AI-generated summaries.

The U.S. Government Accountability Office's AI Accountability Framework organizes practices around governance, data, performance, and monitoring and provides questions and procedures for assessors. Its emphasis on quality information and evidence-based evaluation supports tracing disagreements to data and method rather than accepting a system output.

Sources: [1]

Build the conflict-resolution docket

Create one row per disputed proposition. Include the exact claim, source, source role, publication date, effective date, population, unit, method, assumptions, dependencies, corroboration, contradiction, authority owner, test performed, finding, decision impact, disposition, and unresolved uncertainty.

Use a resolution ladder. First normalize definitions and periods. Then inspect the primary record, reproduce calculations, verify versions, and ask the accountable source owner. Seek an independent source where appropriate. If the conflict remains material, narrow the claim, present both interpretations, defer the decision, or escalate rather than manufacturing certainty.

Do not treat source authority as a fixed global ranking. A policy owner may be authoritative about the current rule, while a transaction ledger is authoritative about what occurred and a method note explains how a measure was calculated. Match authority to the exact proposition in dispute.

The NIST AI Risk Management Framework Core supports mapping context, documenting data and assumptions, measuring risks, obtaining review, recording limitations, and responding to findings. It provides risk-management guidance, not an automatic rule for deciding which source is correct.

Sources: [2]

Worked artifact: a synthetic service forecast

Imagine a synthetic operations team preparing a staffing forecast. A finance workbook reports 18,400 monthly cases, while the service dashboard reports 21,100. An assistant recommends staffing to the larger number and describes the dashboard as more current. The docket shows that the workbook counts completed cases and the dashboard counts all opened cases, including reopened tickets.

The analyst retrieves the underlying exports, aligns the month and timezone, and separates opened, completed, and reopened cases. A further discrepancy remains because one region changed its routing rule halfway through the period. The analyst asks the data owner to confirm the mapping and runs the forecast as a range rather than selecting one total.

The final brief states both definitions, the reconciled counts, the remaining regional uncertainty, and the staffing decisions each bound could support. It schedules a mapping correction before the next cycle. This example is synthetic and is not evidence of Two Prune customer outcomes.

Verify the resolution

Ask another reviewer to reproduce the normalization and calculation from the cited records. Check that the chosen disposition fits the decision consequence and that unresolved conflicts remain visible downstream. Reopen the docket when a source, definition, owner, system, period, or policy changes.

Retain superseded values with their status rather than deleting the disagreement from history. Downstream users should be able to see which figure was replaced, why, by whom, and on what evidence. This prevents a corrected presentation from obscuring the original source problem.

Two Prune can create a controlled opportunity to observe how a participant navigates contradictory sources, verifies calculations, manages uncertainty, and escalates. It does not establish organizational truth, certify evidence, or make employment decisions.