# Answer: where did potentially fraudulent ideas originate in the Enron e-mails, and how did they spread?

*Full analysis, audit trails and caveats: `REPORT.md`. Tables: `results/`. Figures: `figures/`.*

**Short answer.** In the surviving mail, the ideas later associated with Enron's fraud did not start as schemes. They started as ordinary work vocabulary inside small professional clusters, and the words themselves were not transmitted far:

- related-party SPE hedges in legal and structured-finance paperwork;
- discretionary "prudency" reserves in desk P&L notes;
- prepay deals in deal lists and booking requests;
- document retention in routine policy reminders.

They became contested only in autumn 2001, and that framing came in from outside: press coverage, congressional material and litigation notices, relayed inward by media-monitoring staff. The famous crisis phrases have outside or censored roots:

- "implode in a wave of accounting scandals" first appears in the corpus through press newsletters in January 2002, five months after it was written;
- "megawatt laundering" is a press coinage;
- the 401(k) "lockdown" frame arrives through clippings.

The single clean internal root is the 14 August 2001 "never felt better about the prospects for the company" memo. It is a reassurance claim that events showed to be misleading in hindsight. It has one echo, with the wording changed.

**How the paper's method was applied.**
- Idea families are presence/absence traits of 149 custodian mailboxes. They are placed on a tree of who communicates with whom.
- Fritz–Purvis D and an autologistic model separate within-cluster ("vertical") spread from office-level ("horizontal") spread, as in the folktale study.
- Because e-mails carry timestamps and addresses, roots were reconstructed on the time axis, and each descent link was given an evidence class: forward, receipt with re-use, prior tie, shared press source, or unobserved. The folktale study had to rely on tree-based ancestral-state inference instead.
- Duplicate mailbox copies (263,698 of 517,401 rows) and quoted text were never counted as independent transmissions.

**What is statistically supported.**
- *Raptor/LJM/SPE vocabulary (189 messages, 57 internal authors).*
  - People used it after being exposed to it more often than chance allows (0.36 against 0.21 under timestamp permutation, p = 0.004).
  - Wording is inherited along social links (p = 0.001).
  - It is strongly clustered on the communication tree (D = −0.91, p = 0.002); the cluster effect is significant and the office effect is not.
- *Mark-to-market reserve vocabulary.* Also clustered (D = −1.14), with inherited wording.
- **Crucially, the benign United Way campaign also shows exposure-driven spread (p = 0.048).** These tests detect that talk followed contact within work groups. They do not detect wrongdoing.
- *Prepay mail.* Written independently each time (no inheritance). This is convergence on a shared practice, not a meme.
- *The retirement family.* Mostly one external petition copied more than 1,100 times into one mailbox. This is shared-source convergence, not descent.

**Where on the ladder from business to deception.** The corpus contains:
- **ordinary business** (the large majority);
- **advocacy** (process assurances such as "The business units, RAC and Arthur Andersen all signed off");
- **one hindsight-misleading reassurance** (the August 2001 memo);
- **discretionary reserve use** ("I am allowing Rogers to take $4 million from Prudency C … will absorb most of the curve shift");
- **internal concern cues**, the strongest being a risk-research head's refusal in October 2001 to "support the valuations my group has produced so far for the LJM/Raptor related transactions" without legal due diligence.

**No message in the corpus shows a sender stating knowing deception.** No Enron sender orders relevant documents destroyed; the retention mail consists of routine policies followed by preservation orders. The code-named California trading strategies do not appear with their trading meaning. The corpus is a selection of 159 mailboxes with folders deleted, so these are absences in the record, not exonerations. Matching an idea family says nothing about any individual's conduct, and no one is accused here.

**Main limitations.**
- Families are defined by topic, with thresholds set from 371 hand labels (8 per score band), so the precision intervals are wide.
- Four family definitions were tightened after the first labelling round (disclosed in REPORT §2).
- One legal mailbox (shackleton-s) holds 53 of the 189 SPE expressions.
- "Roots" are the earliest surviving messages, not true origins.