The Connecticut Feed Test: Why Law-Firm AI Must Preserve Source Provenance

Fusion Legal & Tax · October 9, 2026Thought Leadership6 min read

A state-specific legal-news page can be an excellent radar screen. It is not necessarily a reliable citation layer.

Consider the Law360 Connecticut page. It functions as a continuously updated collection of litigation, regulatory, transactional, and legal-industry developments. That makes it useful for discovering issues. Its very usefulness as a live feed, however, creates a source-provenance problem for law-firm AI: an item visible today may move as the page changes, while the index URL continues to resolve.

The resulting link can look healthy without continuing to support the proposition attached to it. For firms building AI-assisted research, document review, knowledge management, or client-alert workflows, that distinction—discovery source versus supporting authority—should become an explicit control.

AI can accelerate retrieval without completing the legal analysis

The adoption pressure is real. In its 2025 discussion of AI in tax and accounting, Thomson Reuters reported that the most common current uses of AI-powered automation included “drafting emails and correspondence, reviewing documents, summarizing information, and conducting basic research”. Those are also natural entry points for law firms because they involve large volumes of text and repeatable first-pass work.

Bloomberg Tax similarly explained on April 10, 2025, that generative AI can scan high volumes of content and produce concise summaries. In a vendor-specific survey of senior-level tax professionals who regularly used Bloomberg Tax, 84% of respondents agreed that Bloomberg’s AI tools helped them find answers more quickly. That result should not be generalized to every platform or workflow, but it captures the operational attraction: compress the search process so professionals can devote more attention to interpretation, judgment, and strategy.

The governance mistake is to treat that compression as completion.

A model may accurately identify a relevant headline, summarize the visible text, and return a functioning URL. None of those steps establishes that the linked page is the underlying opinion, order, statute, regulation, agency release, or specific article supporting the final proposition. A research system should therefore preserve not only what the model said, but also what document the lawyer actually reviewed.

The feed test

Before an AI-generated proposition enters a memorandum, pleading, diligence report, client alert, or internal precedent bank, the reviewing lawyer should be able to answer six questions:

1. What kind of page is this? Is it an underlying document or a changing index, topic hub, search-results page, tag archive, or newsroom feed?
2. What exact language supports the proposition? The workflow should retain the quotation or source passage, not merely the model’s paraphrase.
3. Who issued or authored it? A court opinion, regulator publication, news report, vendor analysis, and law-firm commentary carry different evidentiary and persuasive weight.
4. When was it published or effective? Time-sensitive propositions should retain both the publication date and, where applicable, the effective or decision date.
5. Which jurisdiction and procedural posture apply? A Connecticut trial-court filing, a Second Circuit decision, and general federal commentary are not interchangeable merely because they appear in the same state-focused stream.
6. Can another reviewer reproduce the result? The citation should lead a second lawyer to the same supporting material without requiring that person to reconstruct the model’s search path.

If the answer to the first question is “feed,” the item may still be useful. The workflow should simply continue to the deepest stable source available before treating it as substantiation.

Build provenance into the system, not into individual memory

Training lawyers to “check the links” is necessary but incomplete. Link checking can confirm that a page loads; it does not confirm that the page contains the proposition for which it was cited. A durable AI workflow should capture a compact provenance record at the moment research is performed:

This record does more than improve citation hygiene. It creates reusable institutional knowledge. A later lawyer can see why a source mattered, which qualifiers controlled the analysis, and where fresh research is required.

Benchmark provenance, not just answer quality

Many AI evaluations ask whether the output “looks right” or whether lawyers prefer it to a manual result. Firms should add provenance-specific tests.

A useful benchmark set can mix stable primary materials with deliberately unstable or incomplete sources: live news feeds, search pages, summaries, later-amended guidance, superseded documents, and authorities from the wrong jurisdiction. The system should then be evaluated on whether it:

• distinguishes an index from an underlying document;
• carries forward limiting words such as “generally,” “most,” “may,” and “unless”;
• separates allegations in a filing from findings in a decision;
• retains dates for time-bound guidance and reporting;
• identifies when a source supports only part of a compound proposition;
• declines to manufacture a pinpoint citation when none is available; and
• routes unresolved authority questions to human review.

The scorecard should measure more than the percentage of acceptable final answers. Firms should track unsupported propositions, mismatched links, omitted qualifiers, jurisdiction errors, stale sources, and reviewer time required to reach a publishable result. A tool that drafts quickly but repeatedly disconnects propositions from sources may shift work rather than reduce it.

The same control belongs in document review

The feed test also offers a useful principle for large-scale discovery: a classification is not a substitute for the underlying evidence.

AI can assist with responsiveness predictions, issue coding, chronology creation, and document summaries. The defensible unit of work remains the document and the review process around it. Firms should preserve the model version, instructions, document population, validation sample, reviewer decisions, overrides, and escalation criteria. They should also distinguish between a model’s prediction and a lawyer’s legal conclusion.

This framing is protective rather than restrictive. The objective is not to dissuade lawyers from using AI. It is to help teams move faster while making sure no one downstream has to guess which source, document, or reviewer judgment carried the analysis.

A practical adoption sequence

For firms moving from experimentation to routine use, provenance can be implemented incrementally:

1. Choose one bounded workflow, such as first-pass news monitoring or internal research summaries.
2. Define acceptable source types and identify which materials require primary-source confirmation.
3. Require proposition-level support, particularly for rules, deadlines, dollar amounts, eligibility conditions, and procedural steps.
4. Create a feed flag so dynamic indexes remain discovery tools rather than final citations.
5. Run a benchmark set containing both strong and misleading source patterns.
6. Record human review and overrides instead of silently editing the model’s answer.
7. Audit the published work product to determine whether links still support the propositions attached to them.

The strategic point is straightforward: law firms do not need to choose between AI-assisted speed and careful legal work. They do need an architecture that treats retrieval, summarization, verification, and professional judgment as separate stages.

A live Connecticut news page is valuable because it helps lawyers see what is developing. A mature AI workflow goes one step further: it preserves the specific material that allows another professional to understand, test, and responsibly use that development.