The Notice/Rule Gap

A law that drew a measurable boundary

The Plain Writing Act of 2010 requires federal agencies to use plain language in "covered documents." The definition does two useful things at once — it names a document type the Federal Register publishes, and it exempts another:

Plain Writing Act of 2010 · Pub. L. 111-274 · § 3(2)

A covered document "includes (whether in paper or electronic form) a letter, publication, form, notice, or instruction" — but "does not include a regulation."

So the same publication, written by the same agencies in the same years, contains one class of document the Act covers and another it explicitly does not. That is a control group, and it is why this study measures a gap rather than a level.

No break at the boundary

If the Act changed behaviour, the gap between covered and exempt documents should move after 2011. It does not. The gap wanders between 0.00 and 0.08 for twenty-six years, and the Act's compliance deadline passes without leaving a mark.

Measured formally, no headline metric shows a detectable change. Every interval below crosses zero:

Why we do not report this as "the Act failed"

A difference-in-differences only carries a causal reading if the two arms moved in parallel before the intervention. On passive voice they did not: the gap was already widening through the 2000s, at 0.053 per decade before the Act against0.007 after. Something was moving these two document types apart long before Congress acted.

That failure of assumption is why the chart above is the finding and the table is only supporting evidence. The honest statement is narrower than a headline would like:there is no discontinuity at the Act's boundary, and the trends that might be mistaken for its effect predate it. We cannot tell you the Act did nothing — only that nothing it did is visible here.

One measure, the Throat-Clearing Score, does clear significance. We are not reporting it as a finding: the estimate is 0.016 points on a 0–100 scale, against arm means of 0.008–0.019. It is detectable and meaningless, and treating it as a result would be the kind of thing that gets a study taken apart.

Who writes plainly now

Separately from the question above — and making no claim about the Act — here is how plainly each agency writes today. This uses post-2011 notices only, and asks a simpler question with much better evidence behind it: not whether a statute moved anyone, but what the writing actually looks like.

Agriculture Department notices carry a passive construction in 34.5% of sentences. Postal Regulatory Commission notices: 10.8%. That is a 3.2× spread — though most agencies cluster tightly between 23% and 31%, which is its own quiet finding about how uniformly the federal government writes.

The 200-word floor is doing real work and is disclosed for that reason. Without it the Postal Service tops the table with 0.7% passive — not because it writes exceptionally plainly, but because its notices have a median of 96 words and there is barely anything there to be passive. Its one substantial notice is 27.5% passive, dead average. Unfloored, the spread inflates to 47×.

How this was measured

Every document came from the Federal Register's public API. Metadata was harvested month by month for 2000–2025; full text was sampled at 400 documents per type per year, selected by a deterministic hash of the document number so the sample is reproducible rather than random. Federal government works are public domain.

Raw documents are mostly not prose — mastheads, form fields, authority blocks, signature blocks. A whitelist filter keeps only text that reads as prose, and it is applied identically to both arms. That matters here: roughly 15–18% of notices are meeting announcements against 0.2% of rules, so measuring raw text would have compared genres rather than writing.

Every measure comes from an analyzer that already runs on this site — the same code behind the passive voice finder, the sentence length visualizer, and the readability explainer. Nothing was written specially to produce this result.

Confidence intervals come from an agency-clustered bootstrap (1,000 replicates, seeded so they reproduce exactly). This matters more than it sounds: treating all 20,710 documents as independent makes two of these results look significant. They are not. Documents from one agency share a house style, and pretending otherwise manufactures findings.

What this cannot tell you

• Not every Federal Register notice is a "covered document" under the Act. Document type is a proxy for coverage, not a match.
• Many documents the Act most cares about — benefit letters, application instructions, tax forms — never appear in the Federal Register at all.
• Readability formulas are crude, which is why the headline measures here are structural facts (passive voice, sentence length, nominalization) rather than composite grades. See how readability formulas fail.
• Year 2000 runs materially longer-sentenced than any later year in the Federal Register's own text, before any processing of ours. Every estimate is reported with and without it in the published data; no conclusion here depends on that year.

Check our work

Everything below is the actual output of the analysis, not a summary of it. The per-document file is the one to start with if you want to disagree with us.

• Per-document scores — 20,710 rows, one per document, every measure (3.4 MB CSV)
• Event study — per-year means and gaps with intervals
• Difference-in-differences — every metric, with and without 2000, plus pre-trend slopes
• Agency ranking — the table above, both arms
• Per-agency difference-in-differences — the 19 agencies with enough documents in all four cells

Published August 26, 2026. Data is public domain; the analysis is free to reuse with attribution. Corrections are welcome — get in touch.