Enron Brain: this+that's AI-Written Wiki From 45,047 Enron Emails, Every Fact Cited
this+that used an AI pipeline to turn 45,047 Enron emails into a 214-page company wiki containing 49,401 facts, and every fact cites the specific email it came from. No person edited a page or a fact. A held-out audit run by a second AI found 93% of sampled facts fully supported by their source email, and zero of 500 sampled facts wrong at its core.
That is the headline claim of the project, published at thisandthat.chat/enron. The notable part is not that a model can summarize email at scale. It is that the authors attached a source and a lifespan to every claim, then invited anyone to check the output against a corpus they can also read.
What the Enron Brain actually is
The Enron Brain is a set of wiki pages about the organizations Enron dealt with, generated from Enron's internal email. The corpus is 45,047 emails from eight mailboxes. The output is 214 pages and 49,401 facts.
Two design choices separate this from a summarization demo:
- Every fact is a sentence with a source. A fact quotes the sentence from the email it came from rather than being a number pulled into a schema.
- Every fact has an expiry condition. When the mail shows a fact stopped being true, the page expires it on that date. Each entity page has a slider: drag it and the page shows what was known on that date, with facts that were only true for a while expiring in front of you.
The stated reason for facts-as-sentences is that fields like headcount or revenue belong in a CRM or another system of record. In this framing, a company brain sits beside the system of record and holds what changes too often or does not fit in a field.
The accuracy claim, stated precisely
The project reports two audit numbers from a held-out check performed by a second AI:
- 93% of sampled facts were fully supported by their email.
- 0 of 500 sampled facts were wrong at their core.
Those are different bars. "Fully supported" is a stricter test than "not wrong at its core," and the two figures should not be read as the same measure of quality. The methodology page linked from the project site describes how the check was run.
The honest caveat is built into the claim itself: the audit was performed by a second AI, not by human reviewers, and it sampled facts rather than verifying all 49,401.
How the pipeline works
The published pipeline has four named stages, in this order:
- Find the types
- Find the entities
- Build the page
- Verify
The pipeline proposed 57 narrow entity types, and a person kept 54 of them. Almost all of the surviving types were kinds of organization, so the site groups its lists by a coarser class and keeps the pipeline's own type as a tag on each entity.
Worked example from the site: an email sent from chairman.ken@enron.com on November 8, 2001, announcing that Enron and Dynegy were holding discussions about a possible merger and that the New York Stock Exchange had briefly suspended trading in both companies. On the generated Dynegy page, that message becomes a cited relationship entry: Dynegy described as both a competitor and a trading counterparty, with the relationship at that moment labeled "rescue talks," dated November 8, 2001, and marked as checked against the email.
Which entities earned a page, and why the rule is arbitrary
The pipeline found many kinds of named thing in Enron's mail. Only some got pages:
- 214 entity pages, primarily organizations.
- 78 products and platforms, listed without pages.
- 76 facilities and market points, listed without pages.
- 49 events, listed without pages.
The inclusion rule was a message-count threshold: organizations named in at least 25 messages received pages, plus the five most-discussed parts of Enron itself, which were three business units, a trading platform and a trading system.
The project is explicit that this cutoff is a proxy for human judgment, not a principle. A real company brain would not filter by count, because a customer signed today might have sent only one or two emails and its page would still be wanted. The authors describe entity-type selection as setup work a person does when configuring a company brain, and they made that choice before building any page.
The unfiltered lists also show how noisy entity extraction is on real corporate mail. Enron's facilities include trading hubs, Enron Field, a parking garage and two tankers. Its events mix business conferences with charity bike rides and Cirque du Soleil. Its products mix software with an exchange, a statute and trade associations. The site's own framing: every company's mail names things it does not need to track, in categories that are a bit of a mess, and Enron's mail is no different.
The hardest failure mode: the model already knows the story
In the Hacker News thread, co-founder Jeff (jreynar) identified the central technical difficulty:
The hardest part was that the model already knows Enron. It kept adding details from the public record that the email didn't contain.
This is a general problem for any model writing about a well-documented historical subject. The model's pretrained knowledge of Enron's collapse is a contamination source: it can produce plausible, true-sounding facts that never appear in the source emails, which defeats the purpose of a corpus-grounded wiki. Jeff states that the methodology page documents how those additions were caught, what the system still gets wrong, and the cost.
Cost
The project cost about $11,000 at list prices, according to the founder, though the team used a subscription rather than paying list. There was no attempt to bring that cost down.
Why Enron was chosen as the test corpus
Enron's email is one of the few large, free, real corporate mail archives available. FERC released it during its investigation, and Carnegie Mellon has maintained it since 2003. That makes it possible for an outside reader to check any generated page against the primary source.
A second reason is that Enron's corporate history is settled. A page claiming, for example, that Dynegy agreed to acquire Enron in November 2001 can be checked against the historical record instead of being taken on trust. The site recommends Bethany McLean and Peter Elkind's The Smartest Guys in the Room for the full narrative, noting that the mail shows the story only in pieces.
What the most-discussed entities reveal about the corpus
The ten organizations with the most messages in the corpus are:
| Entity | Role in the corpus | Messages |
|---|---|---|
| FERC | The federal regulator overseeing much of Enron's business, where Enron pushed for RTOs and open transmission access, opposed California price caps, and sought to narrow refunds. | 1,501 |
| General Electric | An Enron turbine supplier whose consent governed turbine contract assignments, a Dabhol co-investor, and in late 2001 a rumored rescuer and then bankruptcy creditor. | 1,499 |
| Pacific Gas and Electric | California's largest utility: a gas and power trading counterparty and retail-market incumbent whose 2001 Chapter 11 left Enron a $404 million creditor on its unsecured creditors' committee. | 773 |
| Dynegy | Houston rival and one of Enron's largest trading counterparties that agreed to acquire Enron in November 2001, then terminated the merger and claimed its Northern Natural Gas pipeline. | 593 |
| El Paso Corporation | A Houston competitor and top-ten Enron trading counterparty that bought Enron's Cogen interests, bid for other Enron assets, and restricted its trading with Enron in November 2001. | 478 |
| California Public Utilities Commission | California's electricity and gas regulator, which subpoenaed Enron's records from 2000, received its PX credit complaint, and suspended direct access in September 2001 over Enron's opposition. | 477 |
| Duke Energy | An Enron rival and one of its largest trading counterparties, a co-defendant in California price-manipulation suits, a buyer of Enron assets, and a creditor in its bankruptcy. | 471 |
| California ISO | California's grid and spot-market operator: Enron's trading venue and counterparty, then its adversary over price caps and refund claims, and a debtor owing Enron tens of millions. | 460 |
| Dow Jones | Supplied Enron's custom news clipping service, worked with it on power indexes, and through the Wall Street Journal published the October 2001 Fastow partnership stories. | 418 |
| Reliant Energy | Houston rival and top-ten trading counterparty that coordinated with Enron on California subpoenas and faced contempt alongside it, then limited and slowed its Enron trading in late 2001. | 400 |
This distribution is itself a finding about the corpus: it is dominated by regulators, counterparties and rivals, not by internal operations.
What the Hacker News discussion added
Reaction in the thread split between enthusiasm for the method and skepticism about the framing.
The strongest practical suggestion was that this pipeline should become standard practice for public investigations. One commenter wrote:
This is a neat trick. I would like to see a series of House investigations that do this and publish the wiki as a matter of course.
A more direct critique was that the project announced a wiki without clearly presenting it in the submission. One commenter asked simply:
where's the wiki?
The pages are published on the project's own domain, linked from the submission, rather than in a wiki platform.
The broader ambition was stated by another commenter:
Before you ridicule this, AI has the potential to perform proper journalism. Holding public agencies and corporations accountable.
A wry counterpoint about the state of the forum itself came from another commenter:
At this point I think it's just AIs using hacker news to demonstrate to other AIs.
One commenter argued about tooling taxonomy, that Hacker News needs a vibecoding tag as Lobsters has or had, reflecting unease about how much of a project like this is model output versus engineered system.
The most substantial thread drift was about whether corporate fraud still carries consequences, using Enron as the contrast case:
Ah, Enron. The good ole days when fraudulent execs still went to jail and corporations still saw consequences of fraud. 2002, Arthur Andersen gets indicted for deliberately aiding Enron in cooking the books. 2003, Bush DoJ's attorney general binds prosecutors to effectively always settle corporate cases without convictions. So nowadays you have companies like Boeing murdering 300+ without real consequences.
That argument is the commenter's, offered as context for why a searchable, citation-backed record of a corporate collapse might matter beyond the technical exercise.
What the project does and does not demonstrate
What it demonstrates: an LLM pipeline can extract entity-level facts from a large real email corpus, attach a source sentence to each fact, derive expiry dates from the mail itself, and survive an automated audit at a 93% fully-supported rate.
What it does not demonstrate: that the method generalizes to the entity-selection problem. The 25-message threshold is a stand-in for human judgment, and the site says so. The audit was performed by a second AI rather than by people. The corpus is a public, frozen, historically settled archive, which is exactly the condition that produced the project's hardest problem, the model's prior knowledge of Enron bleeding unsourced details into the pages. A live company's mail would not have that contamination problem in the same form, but it would also not have a settled public record against which every claim could be checked.