Lumen: AI-Assisted Research Library and Courtroom Validation System

The report sounded credible. Its evidence trail did not.

An early AI-generated report included fabricated quotations and citations that were not supported by the source material. I had already presented it to stakeholders, so I rejected the report, disclosed the failure, sent a source-checked replacement, and moved verification ahead of synthesis and strategy use.

Move verification before synthesis and strategy

The before-state moved from a strategy question to search, individual interpretation, a freeform AI report, and only then manual fact-checking. I redesigned that service around source readability, isolated extraction passes, structured evidence records, critique, and resumable checkpoints so verification traveled with the work.

1. Prove every source is readable
2. Extract candidate pain records
3. Normalize without erasing provenance
4. Challenge grounding and scope
5. Preserve rejection and disagreement
6. Compare evidence for decisions

A pain point became a record, not a sentence

The first version organized findings by analyst-defined domains. I shifted to a first-principles schema so the source taxonomy remained metadata rather than silently defining the strategy. Each record connected the person and situation to evidence strength and review state.

1. Actor experiencing the problem
2. Situation and context
3. Desired outcome
4. Friction and consequence
5. Source and location
6. Confidence and review state

AI handled repetitive, inspectable work. People retained judgment.

AI retrieved evidence, proposed candidate records, normalized language, compared patterns, and surfaced uncertainty. I framed the question, inspected behavior, defined critique gates, resolved ambiguous merges, approved strategic relationships, and retained the ability to reject, defer, or override.

1. AI retrieves and extracts
2. AI normalizes and compares
3. AI surfaces conflict and uncertainty
4. People frame and interpret
5. People approve relationships
6. People decide and override

Connect the visible decision journey to the backstage evidence system

The blueprint makes the service design contribution explicit: stakeholder decisions and reviewer actions depended on a backstage system of source gates, Cowork loops, structured matrices, decision logs, partner modeling, and a grounded agent experience.

Challenge plausible claims before they influence strategy

I turned the operating method into reusable prompt and context contracts for source scope, record structure, review behavior, and answer policy. Three isolated extraction passes fed a moderator sweep and five review lenses, followed by explicit admit, reject, contest, or defer states, rejection logging, resumable checkpoints, and a single-writer merge.

1. Run independent extraction passes
2. Test grounding and citation support
3. Check scope and domain fit
4. Inspect attribution and duplicates
5. Admit, reject, or defer
6. Preserve rationale and dissent

Connect customer evidence to partner strategy without hiding tradeoffs

Leadership supplied a universe of 106 prospective partners. I rebuilt the decision-support workbook around separate customer, solution, and business lenses, then used value, feasibility, quadrants, and build waves instead of forcing a false rank from one to 106.

1. Customer pain coverage
2. Solution feasibility and constraints
3. Mutual business value
4. Quick win
5. Strategic bet
6. Hold or revisit

Put the agent downstream of evidence and inside a larger service

The mature extraction and critique loops ran in Copilot Cowork. Separately, I configured a Copilot Studio agent using eight SharePoint-linked domain knowledge documents, tested grounded questions, published it to the immediate team, and connected it to a principal architect's technical-compatibility agent.

1. Cowork evidence loops
2. Structured pain matrices
3. 106-partner decision workbook
4. SharePoint knowledge set
5. Customer-evidence agent
6. Technical-compatibility agent

A working foundation reached limited release. Broader impact remains unproven.

The work produced runnable research loops, preserved adjudication records, structured pain matrices, a three-lens partner model, and a published team agent. A Senior Director in Core Strategy was the intended decision audience, but I was laid off before presenting the completed findings.

1. Completed: research loops and logs
2. Completed: partner decision support
3. Completed: agent QA and team sharing
4. Reconstructed: public workflow diagrams
5. Not completed: leadership presentation
6. Not measured: adoption or business impact

Evaluate decision quality and design the path to adoption

The next stage would pair a representative evaluation set with a staged rollout, reviewer guidance, feedback capture, governance review, failure recovery, and a named operational owner.

1. Citation correctness
2. Unsupported-claim rate
3. Retrieval relevance
4. Duplicate and contradiction handling
5. Reviewer corrections and deferrals
6. Impact on a real strategy decision

The core design problem was deciding where AI should stop

This project joined service design, research operations, AI workflow design, evidence governance, decision support, and adoption planning. The value was not a more fluent answer. It was a service in which people could inspect the source, challenge the reasoning, preserve uncertainty, and remain responsible for consequential decisions.