News · 2026-10-24

An AI-built Enron wiki tests whether citations actually support its claims

An AI-built Enron email wiki reports that a second verification pass increased fully supported facts from 392 to 466 in samples of 500. The project turns correspondence into dated entity pages with inspectable quotations, but its own final audit still found unstated details in 34 sampled facts. It is a useful demonstration of how citation checking can improve generated knowledge without making every cited statement automatically true.

Key facts

• The October 23 build used 45,047 deduplicated messages from eight Enron mailboxes.
• It produced 214 entity pages containing 49,401 facts.
• A second-model audit rated 466 of 500 final sampled facts fully supported by the cited source.
• Primary source: the project’s methodology and audit description.

The public wiki asks whether a company’s “brain,” as the authors put it, can be assembled from correspondence. It is more specific than a chatbot placed over an email archive. Entity pages collect facts and relationships, while a date slider changes what a reader sees as the underlying correspondence changes. The Dynegy page, for example, tracks a relationship through different phases instead of presenting one timeless company description.

Each fact links to a message and quotes the sentence used as evidence. That design gives a reader a short route back to the record. The harder question is what the record actually supports. A sentence mentioning a proposed deal can establish that someone discussed the proposal; it may not establish that the deal closed, that every executive approved it, or that the author’s account was accurate. This is the difference between a source being present and a claim being justified.

The workflow addresses both problems separately. First, a model reads messages associated with selected entities and proposes facts with quotations. An exact-text check rejects quoted sentences that cannot be found in the cited email. Repeated facts are merged. Then another model reads each fact with the cited message and keeps it, narrows it, or discards it. Page summaries and relationship phases are composed from the resulting facts.

Think of two editors reviewing a report. The first checks that every footnote points to a real sentence in a real document. The second asks whether that sentence supports the whole assertion. The first job is mechanical enough to automate reliably for exact text. The second requires understanding what the words do and do not imply. The project’s published audit verdicts make that distinction unusually inspectable.

The authors report that 37 of 62,205 candidate facts quoted text that could not be found and were removed. Those were rejected candidates, rather than 37 errors knowingly left in the final wiki. The more revealing result comes from the separate support audit: 466 of 500 final sampled facts were judged fully supported, while 34 included a detail the email did not state. The final sample contained no facts whose core was judged unsupported, but that is still a sample evaluated by a model, not a guarantee covering all 49,401 facts.

A human blind-checked 50 of the final judgments and agreed with 48. The two disagreements were later supported when the full emails, rather than shortened excerpts, were examined. This check provides some calibration for the judge while also exposing a methodological hazard: a verifier can make a mistake if its evidence window omits the relevant passage. Our guide to AI judging explains why the evaluator’s context and rubric are part of the measurement.

The creators disclose additional limits. The archive slice is small and role-selected, entity discovery sampled 3,200 messages, and long messages were truncated to windows around mentions. Message timestamps need not equal event dates. The verification process was refined after inspecting outputs; the authors identify a later fact audit as held out, while some page-level checks concern material the verifier had already seen. Those distinctions matter when assessing how far a favorable sample can generalize.

Reception in the Hacker News discussion combined interest in searchable institutional memory with concern about the cost of even one misleading fact in legal or business work. The founder said the commercial workflow normally proposes changes for human approval, whereas this demonstration was much closer to automatic. That is a substantive counterargument to treating the demo as ready-made automated reporting.

The lesson is useful beyond Enron: show evidence, verify quotations, separately assess whether evidence supports each claim, and retain uncertainty where it does not. The project reports meaningful improvement while preserving a visible residual error category. Readers can inspect the result today, but neither a matching quotation nor an email author’s assertion substitutes for independent proof of the underlying event. The related lesson on natural-language inference explains the semantic check that a citation list alone cannot perform.

Primary source, verified: read the paper →

Ground Truth. (2026, October 24). An AI-built Enron wiki tests whether citations actually support its claims. Ground Truth. https://groundtruth.day/news/enron-wiki-audit-citations-do-not-prove-support.html

@misc{groundtruth:enron-wiki-audit-citations-do-not-prove-support,
  title  = {An AI-built Enron wiki tests whether citations actually support its claims},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/enron-wiki-audit-citations-do-not-prove-support.html}
}

Topics: knowledge-management · grounding · citations · evaluation · tools