When a joke didn’t get the laugh, he’d drift into the political content that’s his default: crime numbers, Iran, etc.

To see whether the structure of the speech backs that read up, I ran my linguistic matrix on the transcript and set it against the National Republican Congressional Committee (NRCC) Dinner from March.

The audiences, though, weren’t the same kind of room.

At the NRCC Dinner, T was speaking to supporters who know when a line is meant to land and respond accordingly.

Saturday’s audience was primarily the press. Members of his cabinet were there, but seated at tables surrounded by journalists rather than clustered as a bloc. He was performing comedy for a room largely made up of people he’s often at odds with, and that kind of audience doesn’t extend the same courtesy laugh.

I measure political speeches the same way every time, using a few structural yardsticks that have nothing to do with what a speaker says and everything to do with how the speech is built.

Sentence-to-sentence cohesion (CCS-L) measures how much each sentence borrows vocabulary from the sentence right before it. A high score means the speech moves in tight, connected steps. A low score means each sentence tends to start fresh.

Both dinners score low here, which is typical for T’s unscripted style.

They WHCA came in at 0.046, NRCC at 0.059. Neither speech leans on the sentence right before it to carry it forward.

Topical cohesion (CCS-T) measures whether a subject the speech raises early comes back later, versus getting raised once and dropped. A high score means the speech keeps circling back to the same handful of themes. A low score means topics get introduced and abandoned.

WHCA scored 0.952, near the top of the range. Almost every subject he raised came back around at least once, primarily due to his prepared script.

NRCC scored notably lower, 0.766, meaning more of its material was one-and-done.

The Interruption and Fragmentation Index (IFI) measures how often the speaker breaks off mid-thought: cutting a sentence short, inserting an aside, correcting himself, or interrupting himself with a question, all counted per sentence. A higher number means more self-interruption.

NRCC scored 0.091.
WHCA scored 0.153, nearly 70% higher.

WHCA shows meaningfully more of these self-interruptions than the comparable dinner speech from four months earlier.

Flesch-Kincaid (FK) Grade is a readability score. It estimates the grade level a piece of writing was pitched at, based on sentence length and word complexity.

WHCA came in at grade 6.3, mainly due to his use of the prepared remarks.
NRCC came in at grade 4.4, consistent with the ad-libbing rhythm of his rally-type speeches.

Put together, WHCA is a speech that keeps returning to the same handful of subjects, but does so less awkwardly, with many restarts and asides along the way.

The topic chart adds something the numbers alone can’t. Laid out beat by beat, the sharpest, meanest lines of the night (celebrity insults, mock-boxing-card bit, nicknames) aren’t ad-libbed cruelty. They’re the material written specifically for this event.

Between those bits are his ordinary rally content: Iran, immigration, crime statistics. These are the same subjects that show up in all of his speeches, surfacing in short, scattered pockets rather than staying in the background.

When the audience didn’t laugh along even when the material called for it, he reached for the content that works everywhere else. He did that not because it was funnier, but because it was familiar.

Transcript: https://docs.google.com/document/d/1lC5ZhxoxBmPDpP2yKIr-ue4hr1At-7DgiRMRmGv4QqA/edit?usp=sharing

The framework tracks how much each sentence echoes the one before it, whether the speech circles back to topics it already covered, and how often the speaker breaks off or branches away from a sentence before finishing it.

For a reference, I compared it to his NATO summit press conference from July 8, just over a week earlier. That was an unscripted setting, with reporters pushing back, while the prime time address had just a camera and a set of prepared remarks.

Some of this gap is exactly what you’d expect. Press conferences produce longer, looser sentences and more self-interruption. A solo address to camera produces shorter sentences and less of it. The self-interruption score dropping from 0.267 to 0.044 fits that pattern cleanly.

The topic return score is the number that breaks the pattern. Nearly every other T speech I’ve measured, scripted or unscripted, lands somewhere between roughly 0.79 and 1.0. This address comes in at 0.370, and it didn’t get there by moving with the other numbers. It moved on its own.

This is worth separating out. Normally, a low topic return score paired with a high self-interruption score points to the delivery; a speaker interrupting himself and drifting off topic in the moment. Here the two numbers split apart. Self-interruption was about as low as I’ve measured, meaning the delivery itself was controlled. The topic return score is what’s low. If this address felt hard to follow to some viewers, that isn’t because he was going off script and losing his thread. It’s because the content itself moves through five dense, mostly separate blocks of information with almost no recap or return in between.

The likely explanation isn’t the room, it’s the writing. This address had almost no em dashes, the punctuation mark that usually signals a sentence bending off in a new direction mid-thought. Only thirteen showed up across nearly three thousand words, and most of those were brief reactions layered onto a fixed script rather than the sentence actually going somewhere new. Lines like “This was just given out” or “The media reported it” read as comments added at the seams, not as the sentence changing direction the way it does in his less scripted appearances.

That matters because the mid-sentence pivoting this project was built to measure depends on the speaker leaving the page. Structurally, this address stayed on it. The speech moves through five numbered disclosures in order and mostly doesn’t circle back, which is exactly what a tightly prepared script produces, and not what his looser, associative style produces.

So the takeaway here is less about Trump’s usual habits and more about this particular script. When the remarks are tightly written and mostly followed as written, what gets measured is closer to the writer’s structure than the speaker’s usual patterns. If this speech read as confusing or hard to follow at points, the numbers suggest that’s a content problem, not a delivery problem.

https://docs.google.com/document/d/1jKjRuOcRg1oi-yU-bGLuLbWnwB-PcUCPus0S-s6C6jY/edit?usp=sharing

This speech was almost entirely off script, and it shows. Rather than circling back through a handful of themes, the way most of his recent speeches have done, this one moves in a long, mostly one-directional chain. Here are his topics, in order:

Commonwealth-not-a-state joke, McCormick thanks and campaign story, Iran, Secretary of War Hegseth, then a long stretch introducing roughly twenty guests by name and title one after another. Then Pennsylvania history, the investment announcement, military spending, Venezuela, the election, and merit.

From there: crime statistics for three cities, tariffs, the CHIPS Act, energy and the power grid, tax policy, a tribute to the late Lindsey Graham, coal, a tribute to a journalist, tax policy again, Iran again, manufacturing jobs, NATO and inflation, Navy shipbuilding spending.

Then an extended digression into aircraft carrier elevators, magnets, and steam catapults. He flags this one himself mid-speech: “This is not exactly part of my speech.” From there: submarines, the renaming of the Department of Defense to the Department of War, missile defense, smaller defense contractors, trucking policy, and unproven election fraud claims, before closing.

That chain is the actual finding here. Thirty-some topics in sequence, most of them touched once and left behind, is a different structural signature than the recurrence pattern this project has documented in his more scripted appearances.

Now the numbers. This speech ran 8,329 words across 479 sentences, averaging 17.4 words per sentence, the longest average sentence length in this project’s corpus so far. Flesch-Kincaid grade level: 7.6.

Consecutive Cohesion Score, Lexical (CCS-L), which measures word overlap between each sentence and the one right after it, came in at 0.036. That’s low, consistent with most of his unscripted appearances in this project.

Consecutive Cohesion Score, Topical (CCS-T), which measures whether the speech returns to earlier subjects after moving away from them, came in at 0.494. That’s a real departure from his typical range in this corpus, which usually runs 0.79 to 0.95. It matches one other entry closely: his NRCC Dinner remarks in March, at 0.500. Two data points isn’t a pattern yet, but it’s a second data point.

Integrated Fragmentation Index (IFI), which measures self-interruption, truncation, and parenthetical asides per sentence, came in at 0.230. That sits right where you’d expect for an extemporaneous press-conference-style speech, well above his scripted, teleprompter-read appearances and in line with his other off-the-cuff remarks.

Transcript: https://docs.google.com/document/d/1UyJ9oJMtnjRu9v2SfC0Q4S0enXwXAo9waETK1RjbMPg/edit?usp=sharing

Based on this approval, I began to plan the trip. Because the museum is out of town, we would meet at the school and take a bus, and each student who wanted to go would have to contribute $20 toward the overall cost.

The museum directors wanted to make sure that my students were sufficiently prepared for what they would see, and they required documentation that they had received lessons in the classroom prior. I had taught these lessons during elective time, and about 25% of my students are in band, so they did not have the lessons. Those students would have to complete the work on their own time, turning in documentation that they did so. The students already enrolled in my elective class did not need to do this supplemental work.

Ten percent of my students attend the school on a needs-based scholarship program, and the $20 fee would be a hardship for their families. I asked the school administration if they would cover the cost, and they said no. If these students could not contribute the money, they simply would not be able to go.

Twenty percent of my students ride the bus to school on weekdays, and since there’s no school bus service on Saturdays, they would have to find another way to get to campus for our early departure. I know that some of these students will not be able to find that transportation, and that means they will not be able to go on the trip either.

When I initially asked my parents if they thought the trip was a good idea, 83% said it was. They didn’t know that about 25% would have to do extra work, that 10% would not be able to afford the fees, and that another 20% would face a hardship getting to campus on time.

I never went back and asked the parents again, this time telling them all of it — the paperwork, the fee, the early morning bus. I don’t know what number I would have gotten. I know it wouldn’t have been 83%.

A good idea that doesn’t work for everyone impacted stops being a good idea. That’s true of a field trip, and it’s true of election reform.

83% of Americans support voter ID, which is genuine, broad support for an idea. But supporting voter ID isn’t the same as supporting the SAVE America Act. Nearly a quarter of young voters lack the required documentation. 10% live below the poverty line, and fees for documents are an undue burden. 20% live in rural areas and would have transportation requirements to register in person. The bill is more than voter ID, and agreeing with the idea doesn’t automatically mean agreeing with everything the bill actually does to meet it.

A good field trip includes every student who wants to go. Likewise, election reform should not leave anyone out.

Ankara ran 6,340 words across 333 sentences, averaging just over 19 words per sentence. That is long for a Trump talk and contributed to a Flesch-Kincaid Grade level of 7.9.

G7was much longer, but it was full of short, choppy sentences. Ankara is shorter overall but runs in longer sentences.

Structural Comparison

The Interruption and Fragmentation Index (IFI) measures how often a sentence gets interrupted, either by a truncated clause, a parenthetical aside, or a self-generated question.

Consecutive Cohesion Score, Lexical (CCS-L) measures how much vocabulary carries over from one sentence to the next.

Consecutive Cohesion Score, Topical (CCS-T) measures whether the speech returns to earlier subject matter after wandering away from it.

Ankara is roughly two and a half times more fragmented than G7 by IFI, and its local, sentence-to-sentence cohesion is lower too. The two events are nearly identical on CCS-T.

Both speeches keep circling back to the same handful of subjects. Whatever is different between them is happening at the individual sentence level, not at the overall structure.

The opening statement, before any reporter spoke, ran a tight loop of three subjects: NATO and its member leaders, defense spending targets, and the American defense industry. It rotated through those three every few sentences for the full eighteen-sentence-block statement. He broke away only twice, once about communism and TikTok, and once about AI and the power grid. Neither tangent lasted more than a block or two before he returned to the main topics.

The question-and-answer portion behaves differently, and had to be measured differently. Ten of his eighteen answers were substantive enough to check for internal drift. Six stayed on the subject asked. Four did not.

Asked whether the Iran war was a strategic dead end, he moved from Iran’s destroyed military into Venezuela and then into border policy, none of it prompted by the question. Asked about the danger of communism, he pivoted partway through into a boast about the American economy.

Asked about oil markets and the ceasefire, he moved from oil prices into the security of a nuclear site, then into Israel and the old Obama-era nuclear deal. Asked about Turkey’s possible return to the F-35 program, he closed the answer with an unprompted compliment to China.

The pattern in all four is the same. Nothing in the question pulled him there. Something in his own habits did.

This press conference did not happen in isolation. Earlier the same day, in a separate joint appearance alongside Ukrainian President Zelenskyy, T gestured at Zelenskyy and asked reporters if they had a question for “President Putin.”

He also referred to the “Islamic Republic of Japan” while describing an attack on a U.S. aircraft carrier. Afterward he posted on Truth Social: “President Zelenskyy and I just had a News Conference with the Fake News. It went very well. Everybody is looking for a solution. Very positive!”

Those two errors happened before this press conference, but they might explain something. By the time he stood at the podium for the NATO wrap-up, he was already two days into travel and back-to-back meetings, and he had already misspoken twice that day in ways serious enough to make headlines. Maybe he was just exhausted.

But there might be a different explanation. Every other high-fragmentation entry in his speeches was delivered to a friendly room, a rally crowd or a party dinner. His lowest fragmentation scores came at press conferences in front of reporters. When he is standing in front of the press, rather than in front of supporters, he seems to fill his remarks with all the talking points he wants reporters to cover, causing him to switch from topic to topic.

T came to the podium with prepared remarks, and he eventually got through them all. He could still hold a subject across a whole answer, but he could not get through a sentence without stumbling or losing focus.

I am not trying to make a claim about intent or state of mind. It is an observation about where the structure of a speech breaks down and where it does not, on a day when the person delivering it had already shown, twice, that something was off.

Transcripts:

G7: https://docs.google.com/document/d/1qC4yEb3Zb8eC_-VdRjiko_gNd_rISFC2oSINQxgI1Hw/edit?usp=sharing

Nato: https://docs.google.com/document/d/1a_-UXpUkTKv4o4sEa84IM3i2FodCdIXQH8ip5nLANtA/edit?usp=sharing

A note on methodology: the G7 entry above supersedes the version locked on 6/17/26 (CCS-L 0.057, CCS-T 0.087, IFI 0.127). Two things changed. First, em-dash classification in this framework now follows a thematic-continuity standard rather than a strict grammatical-completion standard, meaning a dash counts as a paired parenthetical if the material on either side is topically connected, even if the clause before the dash is never technically finished. The original G7 entry remains on record as historical, but is no longer used in cross-speech comparison. IFI continues to count truncations, paired parentheticals, and self-generated interruptive questions, excluding continuations and externally-caused truncations.

The Mall address had a different pressure behind it. Weather had pushed the start past 11 PM, right before the fireworks, and a long line of guests was waiting to be introduced. There was less room to wander, and the guest structure changed the shape of the speech itself.

Instead of a rally built around one narrative thread, this felt closer to a State of the Union, where the speaker periodically pauses to recognize people seated in the crowd. Trump had called this event a rally beforehand. It didn’t read like one.

To check that impression against the text, I ran both speeches through the same structural framework I use on every entry in this project. It doesn’t evaluate what a politician says or whether it’s true. It measures how a speech is built: how often a sentence breaks off mid thought, how tightly one sentence connects to the next, and how often the speech circles back to a topic after leaving it.

One of these measures is the Interruption and Fragmentation Index, or IFI. It counts how often a sentence gets cut off, self corrected, or interrupted by a question, then divides that by the total number of sentences. A higher score means more of these breaks. Most of Trump’s speeches in this project land somewhere between 0.2 and 0.6.

Both of these addresses came in well under that range. Rushmore scored 0.057. The Mall address scored 0.079. Both confirm that these were unusually script anchored speeches for him. Between the two, though, Rushmore was the tighter of the pair, with roughly half the rate of these self interruptions per word.

Two other measures track how a speech holds together sentence to sentence and topic to topic. Consecutive Cohesion Score, Lexical, or CCS-L, checks how many words repeat between one sentence and the next. Consecutive Cohesion Score, Topical, or CCS-T, checks something broader: whether the speech returns to earlier subjects after moving away from them, instead of abandoning them outright.

Rushmore scored 0.047 on CCS-L and 0.952 on CCS-T. The Mall address scored 0.068 on CCS-L and 0.954 on CCS-T. The lexical scores are low in both cases, which is normal. Exact word repetition between neighboring sentences is rare in fluent speech, so this number stays small even in tightly built remarks. The topical score is where the real signal sits, and both speeches land close to the top of that scale.

Almost every topic introduced in either address got picked back up again later on. Neither one lost its thread, even with the Mall address constantly shifting attention to a new guest on stage.

Most of the gap between the two speeches traces to one specific kind of break, the kind used to identify or explain something before finishing a sentence, like naming a flag or a guest mid clause. Because the Mall speech introduced so many individual people, this kind of interruption showed up far more often there than at Rushmore. That’s a byproduct of the guest structure, not necessarily a sign of a less prepared speech.

Genuine self correction, the kind where a thought gets abandoned mid sentence and never picked back up, was also somewhat more common at the Mall address. So while both speeches held close to their scripts overall, Rushmore was the more disciplined of the two by this specific measure.

None of these numbers say anything about what a speech is about. But something stood out reading both back to back that this framework was never built to catch: neither speech felt celebratory the way past Fourth of July remarks usually do. Both carried a harder edge, built more around conflict and defense than commemoration.

That showed up clearly in word choice. Military and combat language, war, veterans, medals, weapons, battle, appeared close to five times more often in the Mall address than at Rushmore, driven partly by the run of Medal of Honor recipients introduced on stage. Communism came up often in both speeches, referenced about a dozen times each, consistently framed as a live threat to be defeated.

One word barely showed up at all. Across both speeches combined, more than 7,500 words spent honoring the country’s founding and its 250th year of independence, the word democracy appeared exactly once.

Whether that absence means anything is up to the reader. Structurally, it’s simply notable: two speeches built around the anniversary of the nation’s founding, running past an hour combined, circling back again and again to threat and defense far more than to the word most commonly used for the system being celebrated.

Transcripts:

Mt. Rushmore: https://docs.google.com/document/d/1oZ6bevmYBp4dV2P2fuwxB5BnWlSAAlVQUWmt5_JpJr8/edit?usp=sharing

July 4th on the Mall: https://docs.google.com/document/d/1Fq1vNOKc9BShHPinoFMwwjkiFe1z6ID9l9-nQMI5Fas/edit?usp=sharing

What he delivered is unlike anything I have measured from him this term. He held to his script. And by the proportions of the text, celebration was a minor use of the occasion. The four founders carved into the mountain behind him received 4.1 percent of the words, a single line each. The anniversary frame and the catalogue of national achievements together account for about a quarter of the speech.

The largest share, just over 30 percent, went to defining what an American is. Character, culture, identity, language. Another 22.6 percent went to naming an enemy, communism, nearly double the founders and the achievements combined. Where the speech does celebrate, late in the text, it mostly celebrates the present: investment figures and international standing. I measure structure, and the structure of this speech is a definition wrapped in a ceremony.

The Discourse Cohesion and Fragmentation Matrix, the framework I have applied to more than a dozen of his speeches this year, puts numbers to the difference.

The headline number is the Interruption/Fragmentation Index, or IFI. It counts the times a speaker breaks off a sentence or wedges an aside into the middle of one. Questions count too. The total is divided by the number of sentences. Mount Rushmore scored 0.057. The Theodore Roosevelt Library dedication two days earlier, also a ceremonial speech, scored 0.165. The Mack Trucks rally in June scored 0.207, though a rally is a different occasion type and that comparison carries a flag.

To put 0.057 in context, the scripted addresses in my file from other speakers sit in the same low band. Zohran Mamdani’s America 250 address scored 0.078. Michelle Obama’s remarks at a June press event scored 0.075. This is the first Trump speech I have measured that lands among the scripted entries.

The em dash record tells the same story. The transcript contains 36 dashes. Eighteen of them form nine pairs that open an aside and close it again, returning to the sentence they interrupted: “we come to this beautiful mountain—and it is beautiful—to express our gratitude.” A dash pair that closes is a writer’s habit. In the rally transcripts, asides open and rarely find their way back.

Only three self-interruptions counted, and two of them land on familiar ground. He broke a list of national achievements to note that the Nobel committee “haven’t given me one.” He restarted a sentence about international standing: “Like no nation—remember this: we are respected like no nation in the world.” The third was a stumble on scripted material about envy. When the sentence broke on his own terms, the restart came from his standing repertoire.

A fourth interruption did not count, and the decision deserves explanation. Early on, he abandoned a sentence naming the lieutenant governor when fighter jets passed overhead. The index is meant to measure a speaker interrupting himself. An interruption forced by an F-35 is a different kind of event. I excluded it, and the rule now applies going forward: breaks caused by outside events do not count, provided the transcript itself shows the cause. Here it does. His next words were about the flyover.

The speech also reads differently on paper. It scores at a 7.6 grade level on the Flesch-Kincaid scale, which estimates the years of schooling needed to read a text comfortably. His speeches this year have held remarkably steady between grade 4.3 and 5.0 across occasions as different as rallies and commencements. Sentences here average 13.9 words against the usual 10 to 11.

One number needs a warning label. The type-token ratio, which measures how much of a text’s vocabulary goes unrepeated, came in at 0.273 against a corpus norm near 0.17. This speech is half the length of the others, though, and shorter texts score higher on that measure automatically. I report the number and lean on the length-neutral scores instead.

The two cohesion scores complete the picture. The Consecutive Cohesion Score, Lexical, or CCS-L, measures how much vocabulary neighboring sentences share. It came in at 0.047, the lowest of any Trump entry. That may sound like fragmentation. In a scripted text it signals composition. Written prose moves forward by swapping vocabulary from sentence to sentence, while spoken improvisation recycles words, which nudges the score upward. The lowest CCS-L in my entire file belongs to Mamdani’s scripted address.

The Consecutive Cohesion Score, Topical, or CCS-T, measures whether a speech’s vocabulary keeps returning as the text moves along. Mount Rushmore scored 0.952, the highest in the file, just above the TR Library dedication at 0.918. One caution applies: a shorter speech has fewer segments to compare, which makes a perfect return rate easier to reach. And as always, this score captures recurrence of vocabulary. It cannot tell purposeful development from repetition.

One more pattern shows up in the text itself. The topics this speaker carries from speech to speech, the ones my running tallies track across the corpus, sit almost entirely in the final sixth of the transcript, with one short passage near the three-quarter mark. The long middle, the founders and the argument about national character, is nearly free of that standing vocabulary.

That split may explain the low interruption count better than the script alone does. In his other speeches, a self-interruption is usually a branch into familiar territory, a stored line or a comfortable tangent. Here, where the material was unfamiliar, there was nothing to branch into. The two self-interruptions that did occur both land on his standing repertoire, the Nobel grievance and the line about international respect, in the one stretch of the speech built from that repertoire. Where the text was new to him, it stayed intact. I read this as a speech written for him rather than by him. That’s an interpretation the numbers support but cannot prove on their own, since the matrix measures text, not process.

Transcript: https://docs.google.com/document/d/1oZ6bevmYBp4dV2P2fuwxB5BnWlSAAlVQUWmt5_JpJr8/edit?usp=sharing

The speech was 2,296 words across 129 sentences, for an average sentence length of 17.8 words and a Flesch-Kincaid grade level of 9.2. That is not surprising, as he was working from a prepared script and did not deviate or ad-lib.

The first cohesion measure is CCS-L, the Consecutive Cohesion Score, Lexical. It asks how much vocabulary each sentence shares with the sentence that follows it, using a word-overlap calculation averaged across the full speech. The score here is 0.037, which is very low. Of 128 consecutive sentence pairs, 70 share no substantive words at all.

A low CCS-L means each sentence tends to introduce fresh language rather than repeat the language of the last one. In extemporaneous speech, low lexical cohesion can accompany drift. In a drafted text, it reflects a writer varying word choice sentence by sentence. This speech belongs to the second category.

The second measure is CCS-T, the Consecutive Cohesion Score, Topical. Where CCS-L works sentence by sentence, CCS-T works at the level of the whole speech, dividing it into 300-word segments and asking whether the subjects raised in each segment return later or disappear. The score is 0.793, which is high. Five of the six measurable segments return later in the speech, and no segment transition drops below the abandonment threshold. Nothing raised was dropped.

The shape of those returns is worth describing. The harbor imagery, the recurring question of what one sees when looking at America, the figure of George Washington, and the 250-year count all appear in the opening fifth of the speech, recede through the middle, and reappear in the closing fifth. The speech is built as a bookend. Its ending is assembled from the materials of its beginning.

The third measure is IFI, the Interruption and Fragmentation Index. It counts three things per sentence: thoughts that break off unfinished, parenthetical asides, and questions. The score is 0.078, which is very low. Across 129 sentences there are zero truncations. The speech contains six parenthetical asides, every one of them completed, with the sentence resuming and closing after the insertion. There are four questions, and all four are versions of the same question: what do we see.

Two notes to consider regarding the IFI figure. First, an address recorded without an audience carries none of the pressures that shape live delivery. There were no interruptions to recover from and no applause to wait out. Second, the recording itself. I have no way of knowing if his speech was a single continuous take, or several takes edited together. Restarts and flubbed lines can be removed before the released version exists. The IFI here describes the video as released.

A linguistic analysis can only capture the structure of the words. Please take the time to read his remarks for the full impact of his thoughts.

Transcript: https://docs.google.com/document/d/19OOn5OHQIm9DEjkI8PDF0kgDqKPyLEIFP0wc-08uJRE/edit?usp=sharing

Yesterday’s speech had the widest gap I have measured between the two cohesion scores. At the sentence level, this speech barely holds together, but across its full length, its vocabulary returns consistently.

The speech was 8,151 words in 738 sentences, for an average of 11.0 words per sentence. For comparison, a typical declarative sentence in written English runs 15 to 20 words. This speech had a Flesch-Kincaid grade level of 4.6. These numbers describe a speech of short, simple sentences using common vocabulary.

The first metric is Consecutive Cohesion Score, Lexical, or CCS-L. It measures how much vocabulary adjacent sentences share. If a sentence grows out of the one before it, the two will share words. If the speaker has moved to a new thought entirely, they will share almost none. For every pair of adjacent sentences, I calculate the proportion of content words they have in common relative to all the content words across both, excluding filler words. I then average across all 737 consecutive pairs.

The CCS-L score is 0.071. A score of 1.0 would mean every sentence shares all its vocabulary with its neighbor, and 0.0 would mean none do. His most recent speech at a comparable event, the 2026 Coast Guard commencement, scored 0.092. This dedication comes in below that. In this speech, 527 of the 737 consecutive sentence pairs, just over 71 percent, share no meaningful vocabulary at all. Each of those sentences arrives with no lexical thread back to the sentence before it.

The second cohesion metric, Consecutive Cohesion Score, Topical (CCS-T), asks a different question. Instead of looking at neighboring sentences, it divides the speech into 300-word segments and measures how often those segments share vocabulary with distant segments. It is a rough measure of whether the speech keeps returning to its own material across its full length.

The score is 0.918. That is the highest CCS-T for Trump that I have on record.

The mechanism is visible in the word counts. The Roosevelt name family, counting Roosevelt, Theodore, Teddy, and TR together, appears 61 times and is the single most frequent meaningful vocabulary cluster in the speech. Behind it sit the words that anchor nearly every Trump speech: great and its variants at 59 occurrences, the America family at 40, country at 34.

Of the 351 non-adjacent segment pairs in this speech, 339 showed meaningful vocabulary overlap. There is no 300-word stretch of this speech where that core vocabulary disappears. Whatever the speaker is talking about at any given moment, he is never more than a few sentences from saying Roosevelt, great, country, or people again.

The standing limitation of this metric matters here. CCS-T detects recurrence. It cannot distinguish a topic being developed from a topic simply being mentioned again. A score of 0.918 tells you the speech kept circling back. It does not tell you what the circling accomplished.

The fragmentation metric, IFI, counts mid-clause interruptions as a proportion of total sentences. I classify every em-dash in the transcript, sorting them into genuine truncations, where a clause is abandoned before it resolves; parenthetical asides, where the speaker interrupts himself and then returns to the main clause; and continuations, where the dash simply connects two complete thoughts. Truncations and parentheticals count as interruption events, while continuations do not.

This speech contains 186 em-dashes. I found 31 truncations, 52 parenthetical asides (31 of them full double-dash insertions, 21 single-dash), and 72 continuations. Add 39 question-mark sentences and the interruption count is 122 across 738 sentences.

The IFI is 0.165.

Some examples of what these look like in practice. A truncation: “I agreed to — well, when we do it, you are going to be president.” The original clause never gets its object. A parenthetical: “They lost 38,000 people — our people — building the Panama Canal.” The insertion is complete and the main clause resumes. A continuation, which does not count: “He had a dream — a complex dream, because of the way he traveled.” One thought extending itself.

One number worth pausing on is the 72 continuations. Nearly forty percent of this speech’s dashes connect complete thoughts rather than break them. The IFI of 0.165 lands almost exactly on the Coast Guard commencement’s 0.161. Both are ceremonial addresses, both scored under the current formula. The two occasions produced the same interruption rate.

The speech has a genuine architecture. It is organized around five stated lessons from Roosevelt’s life: Americans never give up, a great nation requires courage, America must work, America thinks big, and Americans are one people. All five lessons are announced, delivered in order, and completed. This is the structural difference between this speech and the G7 press conference I analyzed in June. That address lost its subject and never recovered. This one never loses its subject for long. Roosevelt is the most frequent name in the speech, and the frame he anchors survives from the opening to the close.

What happens inside that frame is another matter.

The speech does not reach the first lesson until roughly a third of the way in. Before that come acknowledgments, the Eisenhower train, the transferred federal land, an extended Panama Canal passage, and a long story about a UFC fight that was forecast to be rained out and was not.

Once the lessons begin, each one follows the same pattern: it opens on Roosevelt and drifts. The courage lesson moves from a Roosevelt quotation to a riff on whether the speaker could have passed the line off as his own. The lesson on an America that works opens on civil service reform and ends on the border czar. The lesson on thinking big opens on an 1886 Roosevelt speech in Dickinson and closes on a policeman whose 401(k) gains repaired his marriage. But every drift ends the same way, with a return to Roosevelt and the next lesson.

The frame always comes back. That is what the CCS-T of 0.918 is measuring. The detours are what the CCS-L of 0.071 is measuring: at the moment-to-moment level, one sentence gives almost no warning of what the next will contain.

Read together, all of the numbers describe a speech with a prepared skeleton that held for eight thousand words, and connective tissue that did not. The five-lesson structure was real and it was completed. Nearly everything between its joints was arrival-from-nowhere material, sentence after sentence with no lexical link to the sentence before.

Structurally, this was two speeches at the same time: a ceremonial address with five numbered lessons that reached all five, and, threaded through its gaps, the loosest sentence-to-sentence performance my analysis has found.

Transcript: https://docs.google.com/document/d/1XxpYkGcBAAIRADL8Zm0HLM-cRmQQKT5nDothnI8Gses/edit?usp=sharing

The Flesch-Kincaid grade level is 5.1. That means the speech reads at roughly a fifth-grade level.

The cohesion scores for this speech fall in a range consistent with other Trump speeches I have analyzed.
CCS-L is 0.082, meaning that sentences-to-sentences do not share much vocabulary.
CCS-T is 0.351, meaning core themes recur across the speech even as specific vocabulary shifts. Neither number is surprising.

The IFI — Interruption/Fragmentation Index — is 0.307. That means roughly one in three sentences contains a measurable interruption: a truncation, a parenthetical aside, or a question.

Of 263 classified em-dash instances, 72 were genuine truncations (the thought breaks off), 111 were parenthetical asides (an insertion mid-sentence), and 80 were continuations.

The truncations tend to cluster when the speech loses its footing, especially around the middle where he goes most off-script. “The part about the — Joe Biden’s — well, questions.” “He was not going to — another chance.” “Under — and he just — a nasty one.” These are not rhetorical pauses. The thought genuinely does not complete.

This was a speech about faith. Blocking the transcript into 37 content segments based on word counts:
17 were on-topic (faith, religion, religious liberty)
10 were transitional (secular content with a religious connection at entry or exit)
10 were fully off-topic (voter ID, California ballots, Iran, immigration, economy, rent control). The off-topic blocks are concentrated in the middle third of the speech.

The longest off-topic stretch runs from voter ID through the California ballot digression, into Colombia, then no men in women’s sports, with no religious framing at any point.
The speech opens and closes on faith. The midsection wanders into secular political territory and stays there for an extended run before the closing brings it back.
That’s fairly typical for him; he often uses the teleprompter more at the beginning and the end.

The most repeated content word in this speech about faith is “country” at 41 times.

“Faith” and “God” each appear 12 times. The entire faith word family combined (every form of god, faith, religion, church, prayer, Christian, and related terms) totals 75 instances, or 1.2% of the speech.

“Country” alone appears more than half that total by itself.

The occasion was faith, but the vocabulary was nation. That gap between stated topic and dominant vocabulary is visible in the data.

It is also visible in the block analysis: 27% of the speech has no religious framing at all, and another 27% uses religious framing only as a wrapper around secular content. The word that best captures what this speech is about, at a lexical level, is “country.”

In other words, it was a typical Trump speech that recycled the same content.

Transcript: https://docs.google.com/document/d/1lvx60TMt0CwqPf5c1w9-GB0F_wUDOUfEXE50AoDAgMM/edit?usp=sharing

For years, California has run something called the MCO tax, a fee charged to health insurance companies that funds Medi-Cal, our state’s health program for low-income residents. The federal government matched those funds dollar for dollar, letting California support Medi-Cal without draining the state budget. It has been working this way for over 20 years.

In November 2024, California voters moved to strengthen this system. Nearly 68% passed Proposition 35, which required MCO tax money to go directly toward Medi-Cal improvements, specifically paying doctors more so they would actually accept Medi-Cal patients. It also capped the tax on private insurance plans at $2.50 per member per month to prevent premium hikes. Voters were clear: use this money to improve care, and don’t make private insurance more expensive to do it.

Then Congress stepped in and dismantled it.

Last year, Fong voted yes on H.R. 1, the “One Big Beautiful Bill.” Buried inside it was a new federal rule requiring states to charge Medi-Cal and commercial insurance plans the same tax rate. That single change made Prop 35’s structure legally impossible to maintain. Under the old system, Medi-Cal plans were taxed at $182.50 per member per month while commercial plans paid just $1.75.

To keep federal matching funds flowing, California now has to charge everyone the same rate, which the state has set at $8.85 per month. That blows past the $2.50 cap voters approved, and the state is now structuring the new tax in a way that explicitly bypasses Prop 35 entirely. For families with private insurance, the likely result is premium increases of $100 or more per person per year, and up to $400 for a family of four.

Fong’s letter complains about all of this at length. It never once mentions that his vote for H.R. 1 is what set it in motion.

Newsom is not without blame. He quietly opposed Prop 35 before it passed because he wanted the flexibility to use MCO revenue for general budget needs, which is exactly what his administration is now doing. Structuring the new tax to sidestep a voter-approved initiative is a legitimate complaint. Fong raising that concern would carry more weight, however, if he hadn’t cast a vote that created the underlying crisis. Without H.R. 1, Prop 35 functions as voters intended and none of this is necessary.

The premium increase is also worth keeping in perspective relative to the other damage H.R. 1 does to CA-20. That same bill is projected to disenroll 1.3 million Californians from Medi-Cal by 2030, and the Central Valley has one of the highest concentrations of Medi-Cal enrollees in the state. The constituents who stand to lose their coverage entirely are not mentioned anywhere in Fong’s letter.

Fong voted for the bill that created this problem, and he is now demanding answers about the problem he created. He is counting on most people not knowing enough about the history to notice.

As I did with the Obamas’, I ran Momdani’s through the linguistic analysis matrix I use to evaluate the structure of a speech. His numbers are interesting, but they require more explanation than usual. The speech uses deliberate repetition in ways that the matrix picks up but cannot interpret.

Remember that this analysis is purely structural. It does not address what the speech was trying to accomplish or how effectively Mamdani delivered it. A transcript is not a performance, and much of the power of this speech came from the electricity of the event. Those things are real and meaningful, but they are outside the reach of this framework.

STRUCTURAL CONTEXT

The readability score is a low 6.5, below a typical formal prepared address. For comparison, Barack Obama’s OPC dedication address from the same date scored FK 9.4 and Michelle Obama’s scored FK 7.5.

This is not a surprise and it is not a criticism. Genre matters here. Mamdani was speaking at an outdoor championship rally to a jubilant crowd, not delivering a formal address to a seated audience. The occasion selects for short sentences and plain language. The readability score shows that the speech was prepared appropriately for this occasion.

A second structural pattern runs throughout the speech: the name-plus-contribution. Mamdani consistently pairs a specific player’s name with a specific act (”Karl-Anthony Towns finds the strength to mourn his mother and still pull in rebound after rebound.”) The name and the act do all the work, and the sentence ends. This construction directly produces the short average sentence length which is a feature of the structure, not a symptom of simple thinking.

CCS-L (LOCAL COHESION)

CCS-L: 0.117

CCS-L measures how much vocabulary each sentence shares with the one immediately before it. A higher score means more word-to-word continuity between adjacent sentences.

Mamdani’s 0.117 is notably higher than the OPC addresses from the same date: Michelle Obama scored 0.027 and Barack Obama scored 0.024. The primary driver is the speech’s central structural device: the phrase “It is in that 0.4%” begins eight consecutive sentences, each attaching a different player’s name and contribution to the same opening frame.

When the same phrase anchors eight consecutive sentences, the matrix registers strong adjacency overlap, but it cannot tell you whether that repetition is intentional or not. In this case it is, building rhythm toward a payoff. The high score reflects a deliberate rhetorical choice, not conventional sentence-to-sentence flow.

CCS-T (THEMATIC COHESION)

CCS-T: 0.947

CCS-T measures vocabulary overlap across 300-word segments of the speech, using a formula that accounts for topic return and abandonment rates. A score near 1.0 indicates that the same language and themes recur consistently across all sections.

A methodological note is required here. At 1,107 words, this speech produces only four segments and three non-adjacent segment pairs, which is the minimum needed to compute the metric. The 0.947 result is technically valid, but with so few data points it carries limited interpretive weight. What it does confirm is that the speech never strays from its organizing themes. The themes of the Knicks, New York, the 0.4%, the 53 years run from the first sentence to the last.

IFI (INTERRUPTION AND FRAGMENTATION INDEX)

IFI: 0.070

The IFI measures how often a speaker interrupts their own sentence before completing it, counting genuine truncations (T), parenthetical asides (P), and question marks (Q), divided by total sentence count. Continuations (C) are excluded.

The speech contains six em-dashes total. Four are continuations, with the dash standing in for a comma or colon, and the sentence flowing straight through. Two are parenthetical asides, including the paired construction around “the same guy that so many said was too small” in the Brunson passage. There are no genuine truncations. The three question marks come from the closing rhetorical sequence (”What is New York if not...”).

An IFI of 0.070 is low. It indicates a speaker who completes his sentences. For reference, Michelle Obama’s IFI at the OPC dedication was 0.075; nearly identical, and driven similarly by deliberate audience asides rather than cognitive fragmentation. Barack Obama’s was 0.028, among the lowest I have studied. Trump’s speech at the G7 was 0.127, the highest I have seen.

COMPOSITE SCORES

The CCS-L gap between Mamdani and the Obamas is large, but as noted above, it is driven by the repetitive structure of the 0.4% section rather than by conventional sentence-to-sentence cohesion. The IFI scores for Mamdani and Michelle Obama are nearly the same, which is structurally interesting given how different the two speeches are in length, occasion, and register. Both speakers interrupt themselves sparingly and deliberately. The CCS-T scores are not directly comparable to the Obamas’ addresses because of the short-speech segment limitation noted above.

Finally, the numbers only tell part of the story. What they cannot tell you is what it must have felt like to be in that crowd. Mamdani’s delivery, with his timing and pacing, is real and meaningful and entirely outside the reach of a textual analysis. A transcript is a record of what was said. The feeling that came through was a combination of the power of the written word and the charisma of the delivery.

When I watched the OPC dedication on June 18, I knew that I would run Barack Obama’s address through the matrix. It was a formal prepared speech for a historic occasion, and those are exactly the kind of texts I usually analyze. But Michelle’s speech felt different. So much of it was addressed directly to her husband, and I did not want to study what felt like a love letter.

I changed my mind, though, because she is too good at this for the numbers not to be interesting. Michelle Obama is a skilled public speaker, and her speech was a masterclass. I wondered what that structure would look like.

This post looks at both speeches together, but not to compare them as better or worse. The data is here to explain why they worked, and to show that two structurally coherent speeches can be successful in completely different ways.

These two speeches had different purposes and different primary audiences.

Michelle was addressing Barack. The crowd was present, and she spoke to them directly at key moments, but the organizing center of her speech was a person. Her speech began centered on that person and widened outward into a civic appeal. Barack was addressing the moment. The Center itself and the democratic values it represents were his frame. His speech moved outward from the personal into history.

Barack’s speech is longer and more complex; nearly two full grade levels above Michelle’s on the Flesch-Kincaid scale. Barack’s address moves through a historical argument: the founding era, the unfinished promise of democracy, the present crisis of faith in institutions, a call to civic action. That kind of architecture requires longer, more subordinated sentences. Michelle’s speech is more direct and more immediate. The ideas are not simpler, but the register is more intimate, and intimacy tends toward shorter sentences.

REPETITION PATTERNS

The vocabulary each speaker returns to most often reflects what each speech was actually about. Michelle’s most repeated words are personal and emotional (hope, choice, folks, love). ‘Choice’ appears seven times in her closing section alone, each time applied to a different civic act: hoping, voting, speaking up, showing up. The repetition is intentional and it builds.

Barack’s most repeated words are institutional (people, center, democracy, faith, story). ‘This center’ appears nine times. Democracy appears ten. These are not accidental returns; they are the load-bearing words of a speech organized around a place and an argument.

COHESION INPUTS (CCS-L)

CCS-L measures sentence-to-sentence cohesion: how much vocabulary each consecutive sentence pair shares. Both scores are below 0.10, which usually reflects a formal prepared speech rather than casual address. Neither speaker relies on repetition to carry ideas from one sentence to the next.

The small gap between them is consistent with what each speech was doing structurally. Michelle’s slightly higher CCS-L reflects the demands of managing two audiences simultaneously. When a speaker is moving between addressing a specific person and addressing a crowd, the local connective tissue of the speech has to work harder to keep the listener oriented across those transitions. Her repeated “you never forgot” constructions and her use of direct address as a device are deliberately creating that cohesion.

Barack’s slightly lower CCS-L is the structure of a speech that builds an argument rather than manages a relationship. He moves forward through phases (personal origin, policy record, historical precedent, present challenge, call to action), and consecutive sentences within those phases do not need to echo each other because the larger architecture is doing the work.

COHESION INPUTS (CCS-T)

This is the most striking pair of numbers in the analysis. CCS-T measures whether a speech stays on topic across its full length, dividing the text into 300-word blocks and checking whether each block shares vocabulary with what came before. Both speeches score just above 0.95. Neither one wanders. Every block in each speech connects back to something said earlier.

That is not a given. A score this high means both speakers held their focus from the first sentence to the last, despite the fact that they were doing completely different things. The CCS-T tells you both speakers were disciplined.

FRAGMENTATION INPUTS (IFI)

IFI measures mid-clause self-interruption: genuine truncations (T), parenthetical asides (P), and question marks (Q), divided by sentence count. When I analyze a Trump speech, a high IFI usually shows thoughts that break off, and parentheticals that derail and never return. Neither score here is doing that.

Barack’s 0.028 is among the lowest I have found, with only four parentheticals across 180 sentences. He never breaks frame, never steps outside his argument, never addresses the crowd as a separate audience from the one he is building his case for. That kind of unbroken forward motion is the signature of a speaker who knows exactly where every sentence is going before he says it. Credit must go to the speech writers (he collaborated with longtime chief speechwriter and former White House Director of Speechwriting, Cody Keenan) and Barack for the preparation.

Michelle’s 0.075 is higher, but the mechanism is entirely different. All nine of her parentheticals are controlled audience asides. She deliberately breaks the fourth wall: “I’m not done, y’all.” “Oh, there is truly no higher calling than that.” These are not interruptions of her own thought process, but rather moments where she chooses to step outside the speech and speak directly to the room.

The result is counterintuitive: I believe Michelle’s higher IFI is evidence of more rhetorical control, not less. She is managing two audiences simultaneously, and the parentheticals are the way she does it. Barack does not need that mechanism because he has one audience and one frame from beginning to end.

COMPOSITE SCORES

The scores are reflections of different rhetorical strategies. Both speeches are coherent at the macro level, which is shown by the identical CCS-T scores. At the local level, Michelle’s speech is slightly more active, with more surface movement and more deliberate register switching. Barack’s speech is more uniform, with less variation at the sentence level and more sustained forward momentum.

The data reflects what the listener could sense. Michelle’s speech was intimate and warm, built around a person and widening outward. Barack’s speech was ambitious, built around an idea and expanding it from beginning to end. Both are structurally excellent in the way that fits each speaker’s purpose.

The basics

The speech ran to 3,509 words across 187 sentences, for an average sentence at 18.8 words. This is closer to the norms of formal written English than to improvised speech. Obama definitely adhered to his prepared remarks.

The Flesch-Kincaid grade level is 9.2, which means the speech is more complex than most political speeches I have analyzed. It is not academic, but it was composed of complete sentences rather than short bursts.

For reference: Trump’s Coast Guard Academy commencement address from May 20, 2026 (the closest occasion type I could find of a formal prepared speech) ran 8,282 words at an average sentence length of 11.0 words and a a Flesch-Kincaid grade of 5.0.

Obama’s speech is less than half the length, at nearly twice the sentence complexity, with much greater vocabulary variety.

How the speech repeats itself

Before getting to the cohesion scores, it is worth looking at what the speech actually repeats, because repetition in this address is almost entirely intentional.

Self-reference (I, me, my, mine, myself) appeared 88 times. Collective reference (we, our, us) appeared 131 times. That ratio, 0.67, means the speech uses personal language, but subordinates it consistently to the collective. The opening sections are anchored in “I” (the personal history, Chicago, the South Side) and the speech shifts progressively toward “we” as it moves into civic argument.

The main thing holding this speech together is deliberate repetition, not filler. ‘A belief in’ appears five times, each time introducing another democratic value. ‘The same spirit’ appears four times in the closing passage. ‘I have seen it’ appears three times as the speech turns toward examples of people doing civic work. ‘This is where’ appears three times in the Chicago section. These phrases function as signposts, tellngl the listener where they are in the argument.

Sentence-to-sentence cohesion (CCS-L): 0.024

CCS-L measures the word-level overlap between consecutive sentences. Two adjacent sentences that share content vocabulary score high. Two sentences that each introduce entirely new vocabulary score near zero.

Obama’s CCS-L is 0.024.

That is the lowest cohesion score I have found in the speeches I have analyzed. It is lower than Trump’s G7 press conference from yesterday, which scored 0.057, which was the worst I have analyzed. It was also lower than the T’s Coast Guard commencement at 0.092.

I want to be precise about what that means, because in this case the interpretation is the opposite of what a low score usually signals.

In most of the speech I analyze, low CCS-L reflects associative drift; the speaker moves from sentence to sentence without building on what came before, not because the prose is advancing but because the thread has been lost.

In Obama’s speech, the low score reflects the opposite condition. Each sentence introduces new vocabulary because the speech is moving forward deliberately. The cohesion is carried not by repeated words but by the parallel constructions, recurring frames, and place anchors, and by the progression from personal narrative to civic argument to historical frame to closing call.

A speech can be formally unified while scoring low on lexical adjacency. This is an example of that.

Topic-level cohesion (CCS-T): 0.091

CCS-T measures whether the same thematic vocabulary recurs across non-adjacent segments of the speech. It works at the level of 300-word blocks rather than individual sentences, and it asks whether topics that appear early also appear late, and how often.

The score from Obama’s speech is 0.091. Of 55 non-adjacent segment pairs evaluated, five showed overlap. That is a low recurrence rate.

The topic zone map helps explain why. The speech moves through a recognizable arc: Chicago and personal history dominate the first three segments. Democracy and American nationhood take over in segments four through eight. Faith, spirit, and civic action rise in segments nine and ten. Chicago and faith return together in the final two segments. It is a classic rhetorical structure; the speech closes where it opened, but the middle has done work to justify the return.

The low score is not a sign of incoherence. It reflects the fact that the speech actually moves. The section on democratic values does not keep recycling the vocabulary from the Chicago opening. It introduces new material while building on the same foundation. Every topic thread that gets opened does come back: Chicago, democracy, America, faith, and people all reappear across the full length of the speech, giving a topic return ratio of 1.000. But when they come back, they come back in new language rather than the same words. That is why CCS-T reads low even though the speech holds together.

Fragmentation (IFI): 0.005

IFI measures how often a speaker interrupts their own clause before completing it. The numerator counts genuine truncations, parenthetical self-interruptions, and questions. That total is divided by the sentence count to produce a rate.

Obama’s IFI is 0.005.

In 187 sentences, the analysis found one qualifying interruption event and one question mark. Every other sentence in this speech resolves. There are 33 em-dashes in the transcript, but nearly all of them are continuations or paired parenthetical asides. These are the kind that resolve cleanly and are normal in careful written prose. Zero genuine truncations were found in the two-pass classification.

For reference: Trump’s Coast Guard commencement scored 0.161; 121 interruption events across 752 sentences. Trump’s G7 press conference (notably, a different format) scored 0.127. The press conference is not a fair comparison for occasion type, but it is useful for scale. Obama’s IFI of 0.005 is, in practical terms, the floor.

This is a fully controlled speech, and Obama does not interrupt himself.

What the scores mean together

I have not seen this combination before in my analyses. Usually a low IFI and a low CCS-L don’t show up together. A speech that is this controlled tends to rely on repeated vocabulary to hold itself together. This one doesn’t. The structure is there, but it runs through the parallel phrases and the topic arc rather than through word repetition at the sentence level.

What we heard is what the data shows. Obama delivered a prepared address that knew where it was going, moved toward that destination deliberately, and arrived there without losing the thread. That is not as common as it sounds, and it was refreshing.

Transcripts:

Obama: https://docs.google.com/document/d/1i6aHWExWFg0Q3MfGx8-SfbaYCDyGB3O-RuWX-i5uvkU/edit?usp=sharing

Trump: https://docs.google.com/document/d/1FLR2RgINAPUwjUXpJsIoL9mPGfgANIIgefTAb4LA8LY/edit?usp=sharing

After running the full analysis on Tuesday’s G7 press conference, I want to say something I have not said before about any speech in this corpus: this is the worst I have seen. Not the most fragmented by a single metric. The worst overall, across the full picture — the numbers, the zone map, and the distance between what the speech was nominally about and what it actually contained.

Here is what the data shows.

The basics

The press conference ran to just under 10,000 words of Trump’s speech across the full Q&A session. That produced 999 sentences at an average of 9.9 words each. For context, a typical declarative sentence in written English runs 15 to 20 words. At 9.9 words, you are looking at a speech built almost entirely out of short bursts — declarations, self-corrections, and sentence fragments that parse as complete thoughts because they end with a period but often began without a subject or finished without a predicate.

The readability scores confirm the register. The Flesch-Kincaid grade level is 4.0, meaning the vocabulary and sentence complexity are calibrated to a fourth-grade reading level. The Flesch Reading Ease score is 84.1 out of 100, where higher numbers mean easier reading. For reference, a score in the low 60s is considered plain English accessible to most adults. This speech scores 20 points above that. The Gunning Fog Index, which penalizes long sentences and complex words, is 7.1.

These are not inherently damning numbers. Simple language is not bad language. But they establish the baseline: this is a speech with very short sentences, very common words, and very little syntactic complexity. The question the remaining metrics answer is whether that simplicity serves clarity or disguises its absence.

How well do consecutive sentences connect?

The first metric I use is called Consecutive Cohesion Score — Lexical, or CCS-L. It measures how much vocabulary consecutive sentences share. If sentence two grows out of sentence one, they will share words. If the speaker has moved to a new thought entirely, they will share almost none.

I measure this using a standard similarity calculation: for every pair of adjacent sentences, I calculate the proportion of content words they have in common relative to all the content words across both sentences. I exclude common filler words like “the,” “and,” “is.” I then average that figure across all 998 consecutive sentence pairs in the speech.

The score for this press conference is 0.057.

To understand what that means: a score of 1.0 would mean every sentence shares all its vocabulary with the one before it. A score of 0.0 would mean no sentence shares any vocabulary with the one before it. In the NPB speeches I analyzed in 2017 and 2018, the scores were 0.65 and 0.75 — meaning the speaker’s sentences were substantially connected to each other at the local level. By 2025 that had fallen to 0.45, and by 2026 to 0.33.

This press conference scores 0.057. That is the lowest CCS-L in the entire corpus. It is lower than any rally, any ceremonial address, any prior press conference I have analyzed. Sentence to sentence, there is almost no lexical thread. Each sentence is arriving essentially cold.

How well do topics return and develop?

The second cohesion metric, CCS-T, measures something different. Rather than asking whether adjacent sentences connect, it asks whether the speech returns to its own topics across its full length. I divide the speech into 300-word segments and measure how often those segments share vocabulary with non-adjacent segments — a rough measure of topical recurrence.

The score is 0.087.

For comparison, a speech in which topics were being actively developed and returned to would produce a score well above 0.30. The 2017 NPB speech, which I have used as a longitudinal baseline, has not yet had CCS-T computed under the current formula, but the trajectory from other corpus entries makes clear that 0.087 is very low. It means that across the 33 segments of this speech, only 46 of 528 possible segment pairs showed meaningful vocabulary overlap. Topics that were introduced were largely not picked up again in any sustained way.

There is a known limitation worth naming here. CCS-T can detect whether a word like “nuclear” or “deal” recurs across segments. It cannot detect how far the speaker traveled away from those words before returning to them, or whether the return represents development or simply another mention. That limitation matters for this speech in ways I will return to below.

How often does the speaker interrupt himself?

The fragmentation metric, IFI, counts mid-clause interruptions as a proportion of total sentences. I use a two-pass manual classification of every em-dash in the transcript, sorting them into three categories: genuine truncations, where a clause is abandoned before it resolves; parenthetical asides, where the speaker interrupts himself but returns to the main clause; and continuations, where the dash simply connects two complete thoughts. Under my current formula, both truncations and parentheticals count as interruption events. Continuations do not.

This speech contains 214 em-dashes across Trump’s turns. After manual classification, 33 are genuine truncations and 91 are parenthetical asides, for a total of 124 interruption events across 999 sentences.

The IFI is 0.127.

Some examples of what these look like in practice. A truncation: “We are using — they could stop us if they wanted to.” The first clause has no object. It simply stops. A parenthetical: “Those B-2 bombers — who would have thought they could handle it? — they handled three of the heaviest bombs.” The aside is complete, but the speaker has interrupted his own subject before stating the predicate. A continuation, which does not count: “The strait is closed — bad things will happen.” Two connected clauses.

At 0.127, the IFI reflects a speech with frequent self-interruption. What it does not capture, and what the zone map below makes visible, is the cumulative effect of those interruptions across a 10,000-word performance.

What the speech was actually about

I mapped the full speech into 101 ten-sentence blocks and labeled each by its primary topic. The nominal subject of the press conference was the Iran nuclear deal reached on Sunday, announced at the G7. That is what the reporter questions were about. That is what the opening statement was nominally structured around.

Here is a partial account of what the blocks actually contain, in order.

The speech opens with the Iran deal and stock market response, pivots immediately to Soleimani, moves to G7 allies, returns to Iran military losses, shifts to regime change framing, detours to the stock market’s brilliance and Herbert Hoover, returns briefly to G7 logistics, moves to Soleimani for a sustained four-block run, transitions to the Obama/JCPOA narrative for six blocks, then enters a four-block tangent on granite. Not the geology of Iranian nuclear sites. Granite as a building material, with specific reference to the new granite on the White House stairs, rated one million years. Then B-2 bombers, then Space Force facial recognition cameras, then back to nuclear stockpiles, then denuclearization broadly, then deal signing logistics, then Israel, then Lebanon’s cultural history, then reconstruction investment, then a fake news story about JD Vance, then the breakfast bombing of the first leadership group, then Herbert Hoover again, then the tax cut, then Gulf states ballistic missiles, then Pakistan and Qatar, then UAE’s Mohammed dropping bombs, then back to Soleimani one more time, then the Obama deal versus the Trump deal as road versus wall, then F-22s, then B-2s again, then Gaza and Hamas, then Lebanon and Hezbollah, then Syria’s president, then Abraham Accords, then the rigged election, then five million unvetted immigrants, then Ukraine, then Ebola, then artificial intelligence, then electric plants, then California’s power grid, then drug trafficking, then cartels in Mexico, then bilateral meetings, then Israel’s nuclear exposure again, then Biden’s ice cream order, then the Palace of Versailles and its gold.

That is the opening monologue, before the Q&A begins.

The longest sustained run on the Iran deal — the stated subject of the press conference — is six consecutive blocks in the Obama/JCPOA section. After that, the deal appears and disappears in single-block mentions, surrounded each time by material that has nothing structural to do with it. Herbert Hoover appears twice with no logical bridge between his appearances and the surrounding content either time. The granite tangent runs 40 sentences.

The Palm Beach speech I analyzed in April also showed this kind of episodic structure. But at Palm Beach, the zones were long and the detours were short. Here, there are no long zones. There is a nominal subject that the speech keeps losing and briefly recovering, interrupted by detours that are sometimes as long as the returns.

What this adds up to

CCS-L of 0.057 is the lowest sentence-to-sentence cohesion in the corpus. CCS-T of 0.087 is among the lowest topic-return scores. IFI of 0.127 reflects sustained self-interruption. And the zone map shows a speech in which the announced subject — a historic diplomatic agreement reached four days earlier — cannot hold the speaker’s attention for more than six consecutive blocks before something else takes over.

I want to be careful about what I am and am not claiming. I am not making a psychological inference. I am not evaluating the policy. I am measuring structure, and structurally, this is the most disorganized speech I have analyzed in nine years of running this framework.

The numbers do not explain why. They only describe what happened. What happened is that a speaker held a press conference about a single major event and was unable to sustain focus on that event for more than a few minutes at a stretch across a ten-thousand-word performance.

Transcript: https://docs.google.com/document/d/1qC4yEb3Zb8eC_-VdRjiko_gNd_rISFC2oSINQxgI1Hw/edit?usp=sharing

When RFK Jr. became HHS secretary, he gained the power to go looking for proof that has eluded him for three decades. He is using that power to pursue access to your personal, private medical records. He doesn’t want anonymous data and statistics; he wants your records that have your name on them.

Kennedy didn’t start this gathering of data. He’s the latest chapter in a story that has been building since January 2025.

This administration has already moved to centralize IRS, Social Security, immigration, and Medicaid information. In most cases, they are doing this without legal authorization or any announced framework for how the data would be used or protected. Medicaid records have been transferred to immigration enforcement. Social Security files were accessed by DOGE engineers who had no formal authorization to be in those systems. A megadatabase has been pulling in financial data from the IRS.

The chronic disease mission is the excuse that makes this particular door easier to open because it sounds like a public health action. Nobody is arguing against researching childhood illness, but the infrastructure being built around that mission goes far beyond the stated purpose.

To help with legitimate research, a long-time CDC veteran offered Kennedy access to massive databases of deidentified medical data, meaning names and identifying information had been removed. Kennedy turned him down. Instead, Kennedy sent advisers to CDC headquarters to download millions of identifiable patient records directly. He wanted names attached.

Kennedy’s team then approached the organizations that manage state health information exchanges, which are the systems hospitals use to securely share records with each other. Most states said no. Nebraska said yes, and federal money followed immediately.

Nebraska’s health department received $18.7 million, more than any other state, despite Nebraska ranking 38th in population. The nonprofit that cooperated received $13.6 million shortly after.

The plan Nebraska helped develop would give the federal government real-time, around-the-clock access to the medical records of 90% of Americans by 2028. Your prescriptions, your diagnoses, your doctors’ notes would all be shared with the government in real time.

The company managing it is Palantir, a surveillance technology firm founded by Peter Thiel, one of Trump’s and Vance’s most consequential political backers.

Democrats in Congress have been trying to get answers about how the government will use and protect your private data. They have written letters, introduced legislation, and forced subpoena votes. Republicans have blocked every attempt on party-line votes. When Democrats started winning some of those votes, Republican leadership rewrote committee rules to eliminate the forum where it was possible.

Democrats need to pick up three House seats in order to flip control. If that happens, Democrats will chair the committees that can subpoena documents, compel testimony under oath, and demand public answers about what is being done with your information.

House control does not guarantee that the collection of your medical records gets stopped. However, without a majority, Congress can write all the letters it wants, and nobody has to answer them.

“Prevents the largest tax increase in American history... protecting the average Central Valley taxpayer from an 18% tax hike or the equivalent of a $3,108 average annual tax increase per family.”

The median household income in CA-20 is $90,892. For households in the middle of that range, the tax savings from making the 2017 cuts permanent are real. But they sit alongside equally real cuts to Medicaid, food assistance, and school funding that fall hardest on the same working families Fong claims to be protecting. For the 11.6% of CA-20 residents living below the poverty line, and the far larger share living just above it, the math in this bill does not work out in their favor.

The 18% figure itself reflects what would have happened if the 2017 cuts had simply expired, a scenario that was never certain and that a Democratic Congress would almost certainly have addressed through targeted relief.

“Makes permanent and increases the doubled Death Tax exemption, helping ensure family farms, ranches, and other small businesses can be passed down to the next generation.”

The data on this has always been clear. According to the USDA, approximately 0.3% of farm estates nationally owed federal estate tax under the previous exemption level, roughly 87 to 89 estates per year across the entire country. In 2025, the exemption was already $13.99 million per individual, or nearly $28 million per married couple. The One Big Beautiful Bill raised it to $15 million per individual.

The farms affected by this change have average net worths exceeding $32 million. There is no meaningful population of working family farms in CA-20, or anywhere in the Central Valley, that was in any danger under the old threshold. The beneficiaries are the wealthiest agricultural landowners in the country.

“Expands and makes permanent the 199A small business deduction to 23%, protecting nearly 42,000 local small businesses from being hit with a 43.4% tax rate...”

The claim about 42,000 local small businesses deserves scrutiny. The 199A deduction applies to pass-through entities, businesses whose income flows directly to the owner’s personal tax return. In CA-20, most small businesses are sole proprietorships, family operations, and service workers, many of whom earn well below the threshold where this deduction provides meaningful benefit.

The Joint Committee on Taxation, Congress’s own nonpartisan scorekeeper, estimates that 61% of the benefit from this deduction goes to the top 1% of income earners nationally. A Tax Policy Center analysis found that 55% of the 2019 benefits went to the top 1%, with more than 26% flowing to just the top 0.1%. A recent bipartisan analysis confirmed that nearly two-thirds of the expanded deduction will go to taxpayers earning $500,000 or more, with nearly half going to those earning over $1 million.

The primary beneficiaries of 199A are law firm partners, real estate investors, and large LLC owners.

“Invests $60 billion in support for our farmers and ranchers.”

Federal farm support programs have always been built around covered commodity crops: corn, wheat, soybeans, cotton, rice, and a handful of others grown primarily in the Midwest and South. The Central Valley grows almonds, grapes, citrus, dairy, and other specialty crops that have historically received a small fraction of federal agricultural payments. When the USDA distributed $12 billion in bridge assistance payments in late 2025, $11 billion went to row-crop producers and just $1 billion was set aside for specialty crops. Corn payments alone accounted for nearly 42% of disbursements. California farmers, who primarily grow specialty crops, expressed public skepticism about the local impact, and the data bore them out.

Retaliatory tariffs from China, India, and the EU have meanwhile specifically targeted the crops Central Valley farmers depend on. China imposed a 35% tariff on U.S. almonds, a crop grown almost entirely in California. Dairy exports face significant exposure to retaliatory trade measures affecting China and the EU. The $60 billion Fong celebrated was designed for a different region’s agriculture, and our growers are absorbing the consequences.

“Doubles funding for trade promotion programs to $400 million annually for the Market Access Program...”

The Market Access Program has long drawn criticism for directing most of its support to large national commodity groups promoting corn, wheat, soybeans, and cotton, rather than to the specialty crop producers who make up CA-20’s agricultural economy. Doubling MAP funding without changing who controls those dollars does not change who benefits.

Beyond that structural problem, MAP and the Foreign Market Development program are marketing tools. They fund trade shows, export promotion campaigns, and brand recognition abroad. They cannot compensate a Kern County almond grower for the 35% Chinese tariff now applied to their product, or a dairy farmer for the collapse of export channels to China and the EU. The trade promotion investments Fong cited cannot undo the trade damage his party’s policies have caused.

“Provides long-term funding for the National Animal Disease Preparedness and Response Program, which is extremely important for controlling outbreaks like Highly Pathogenic Avian Influenza...”

Continued investment in disease preparedness is genuinely worthwhile, and this is one of the few provisions in the bill with direct relevance to CA-20’s agricultural community given the ongoing HPAI crisis affecting Central Valley dairy and poultry operations.

The honest limitation is that NADPRP provides planning grants and resources, not direct producer relief. Its benefits flow primarily to large, integrated operations with the administrative capacity to navigate federal applications. The smaller dairies and poultry operations hit hardest by the outbreak are the least likely to access meaningful support. The bill does something useful, but it falls short of what our farmers need.

“Makes a generational $144 billion investment in America’s defense, including $9 billion to improve servicemember quality of life...”

Fong cited the Pentagon’s original $144 billion request, but the enacted law provided $152 billion. Of that total, $7.5 billion is designated for military personnel and servicemember quality of life. The remaining $144-plus billion goes to weapons systems, munitions stockpiles, nuclear modernization, shipbuilding, and advanced aircraft.

This matters in CA-20 because our district is home to NAS Lemoore and thousands of active-duty military families and veterans who rely on the federal safety net. The same bill that claims to invest in servicemember quality of life cuts Medicaid by $1 trillion, cuts food assistance by $186 billion, and has already triggered the withholding of over $800 million in California education funding.

That last point has direct consequences for military families. Medicaid is the fourth-largest federal funding source for K-12 schools nationally, providing roughly $7.5 billion in school-based health services annually. The OBBBA’s Medicaid cuts will reduce school reimbursements for vision and hearing screenings, nursing services, counseling, and special education supports. The children of servicemembers at Lemoore attend those schools. Cutting their school health services while calling it an investment in servicemember quality of life is a contradiction the numbers make plain.

“Restoring integrity to Medicaid, SNAP, and other essential safety net programs by rooting out waste and strengthening them for those they are intended to serve...”

This is the section where the gap between Fong’s rhetoric and reality is most measurable, because the cuts are already in effect.

Start with the premise. The Kaiser Family Foundation’s analysis of who actually uses Medicaid found that 64% of enrollees under 65 are working full- or part-time. Another 12% are caregivers for family members. Ten percent are ill or disabled but not formally classified by Social Security. Seven percent are in school. The “able-bodied adults who won’t work” this bill claims to target represent, at most, 7% of enrollees, a group whose removal would account for roughly 13.5% of the $1 trillion in cuts. More than 85 cents of every dollar cut from Medicaid falls on working adults, children, seniors, caregivers, and people with disabilities. That is not waste reduction.

The consequences are documented. The Congressional Budget Office estimates 11.8 million Americans will lose health coverage under the bill’s health provisions by 2034, a figure combining direct Medicaid losses with related marketplace coverage losses. California’s own projections put up to 3 million Californians at risk of losing Medi-Cal by 2028. Work requirements, which take effect for Medicaid in December 2026, have already been shown in multiple state experiments to cause massive coverage losses driven primarily by paperwork failures among people who are fully eligible.

On food assistance, the harm has come in two distinct waves.

The first hit on April 1, 2026, when the OBBBA’s eligibility restrictions for lawfully present noncitizens took effect in California. Tens of thousands of humanitarian immigrants, including refugees, asylees, and survivors of domestic violence and human trafficking, lost CalFresh eligibility entirely. These are people who came here legally, many fleeing persecution, who had been part of our communities and our workforce.

The second wave began June 1, 2026.. Pandemic-era work requirement waivers expired February 28, 2026. Starting March 1, able-bodied adults without dependents ages 18 to 64, a category the bill expanded from the previous ceiling of 54, have been required to document 80 hours per month of work, training, or volunteering to keep SNAP benefits. The three-month compliance clock that began in March means the first benefit terminations for that group took effect June 1, 2026.

The cumulative result is that more than 3.5 million Americans lost SNAP benefits between the law’s July 2025 enactment and February 2026, before the work requirement losses even began, a drop of nearly 9% nationally across every single state. California saw a more than 6% decline in SNAP participation over that same period. California food banks were already serving a record 6 million people per month before the bill passed, up from 4.5 million at the height of the pandemic. They are now bracing to serve hundreds of thousands more as the eligibility cuts continue rolling out.

“Providing over $140 billion — the largest border security investment in history — to safeguard our borders and keep Americans safe.”

More than a third of CA-20 residents are Hispanic, and mixed-status families are woven throughout the agricultural economy, the service sector, and the broader community. The workers who harvest the crops Fong celebrates in his agricultural talking points are disproportionately from immigrant families.

A $140 billion enforcement-only package with no corresponding investment in immigration reform, farmworker protections, or legal pathways for essential agricultural labor sends a clear message about whose safety this bill prioritizes. It simultaneously cuts Medicaid and food assistance that U.S. citizen children in immigrant households depend on. Those children were born here. They are enrolled in CA-20 schools. They are eligible for every benefit by right of birth. The bill criminalizes their families and cuts the programs they rely on in the same legislation.

“Investing $12.5 billion as a down payment on efforts to overhaul and modernize failing air traffic control systems...”

Air traffic control infrastructure is a legitimate federal need. It is not the most urgent infrastructure need of a district where rural roads are deteriorating, emergency medical transportation is underfunded, and safe drinking water remains inaccessible in some communities.

Kings County, at the heart of CA-20, has a per capita income of approximately $45,400 (2023 figure, inflation-adjusted), among the lowest in California, and an unemployment rate that averaged 8.9% in 2024 and climbed above 10% in early 2026. Most residents here do not fly regularly. The people who benefit most from a $12.5 billion aviation investment are concentrated in urban coastal regions and the aviation industry itself. For families in CA-20 who drive 45 minutes to see a doctor, who have no reliable public transit, and who are uncertain about their tap water, this investment was never aimed at them.

“Implementing a $250 registration fee on electric vehicles and a $100 fee on hybrids...”

The San Joaquin Valley has some of the worst air quality in the nation, a consequence of traffic, agricultural emissions, fossil fuel infrastructure, and geography. The health burden falls heaviest on lower-income residents, who experience higher rates of asthma, heart disease, and respiratory illness.

Penalizing the shift to electric vehicles in a district with this air quality profile trades long-term public health for a regressive annual fee. Teachers, farmworkers, and young families who switched to cleaner vehicles to save money on gas are now paying $250 a year for that choice. The fee does nothing to improve roads. It discourages precisely the transition that would most benefit the air CA-20 residents breathe.

“Expanding 529 education savings accounts to empower American families and students...”

Tax-advantaged savings accounts benefit families who have disposable income left over after meeting basic expenses. In a district where per capita income in Kings County is approximately $45,400 (2023, inflation-adjusted) and unemployment has exceeded 10% in early 2026, the number of households in a position to fund a 529 account is limited. This provision was designed for a different income bracket.

It arrives in a bill that simultaneously cuts Pell Grant eligibility, reduces school-based health services, and shifts billions in costs from the federal government to states and school districts. Families who needed those programs will receive nothing from the 529 expansion. Families who can afford to save get a new tax shelter. The gap widens.

“Renewing 100% immediate expensing, incentive for research and development, and deduction for interest expenses...”

One hundred percent immediate expensing benefits companies making large capital investments in factories, equipment, and technology infrastructure. That describes the defense industry, the energy sector, and large-scale manufacturing. It does not describe the economic base of CA-20, which runs on small farms, food processing, logistics, and service work.

These provisions carry no requirements to create local jobs, hire American workers, or invest in rural areas. Any qualifying business can take the deduction and expand overseas. Central Valley communities watching federal revenue eroded by corporate tax preferences while their roads, schools, and public health systems go underfunded are not the intended constituency of this legislation.

“Delivering on President Trump’s priorities of no tax on tips, overtime pay, and car loan interest, and providing $4,000 of additional tax relief for seniors.”

The no-tax-on-tips provision primarily benefits tipped workers who earn enough to carry meaningful federal income tax liability, those at the higher end of tipped income. Workers near minimum wage already owe little or nothing on their tips and gain almost nothing from this change. The same pattern holds for the overtime provision: the benefit scales with income, not need, and part-time and seasonal workers who make up much of the CA-20 workforce see minimal impact.

The $4,000 senior deduction assumes seniors owe federal income tax in the first place. Many in CA-20 do not, because their income comes from Social Security and modest pensions. What they depend on is Medicaid for long-term care, Medicare Part B subsidies, and nutrition assistance. Those are the programs this bill cuts. A deduction that reduces a tax bill you do not have does not make up for healthcare coverage you lose.

One Year Later

Congressman Fong’s May 2025 statement described a bill that would protect Central Valley families, support local farmers, strengthen the safety net, and invest in the future. Eleven months later, the record looks different.

More than 3.5 million Americans lost food assistance before the SNAP work requirement cuts even began. In California, refugees and domestic violence survivors lost CalFresh eligibility on April 1. This week, June 1, 2026, the first SNAP benefit terminations under the new work requirements are taking effect. Up to 3 million Californians are projected to lose Medi-Cal by 2028. California food banks, already serving a record 6 million people per month before the bill passed, are now bracing to serve hundreds of thousands more. The almond and dairy farmers who form the backbone of our agricultural economy are navigating retaliatory tariffs that no trade promotion program can fix. The children of servicemembers at NAS Lemoore attend schools facing reduced Medicaid reimbursements for health services.

None of this was hidden. The CBO scores were public. The KFF analyses were published. The distributional tables from the Joint Committee on Taxation were available to every member of Congress before the vote.

Vince Fong read them and voted yes anyway.

The voters of CA-20 will have a chance to respond in November.

Sources: Congressional Budget Office; Joint Committee on Taxation; Kaiser Family Foundation; USDA Economic Research Service; UC Berkeley Labor Center; California Department of Health Care Services; California Association of Food Banks; Tax Policy Center; Bipartisan Policy Center; Center on Budget and Policy Priorities; CalMatters; KQED; CNBC; Center for American Progress.

The video features more interactions with statues than with constituents. They appear over and over as backdrops, while actual voters are largely absent as participants. There is no sustained conversation. No one is shown asking a question, offering a concern, or reacting to what Rep. Fong says. The few people who do appear are not positioned as constituents with importance, but rather as scenery.

This pattern repeats across the video’s job-site visits. We see aerospace, energy, military, and infrastructure settings. These are capital-intensive, tightly controlled environments where access is staged. The people in these scenes are hosts, managers, or institutional representatives guiding the congressman through facilities. They are not shown as voters seeking representation. We never see a conversation.

Agriculture, the economic backbone of California’s 20th Congressional District, appears only at a distance. There is a brief walk through an orchard, but there are no workers present. We see overhead shots of a tractor and cattle, and there is a dam scene focused on infrastructure. Throughout the video, agriculture is presented only as land and machinery. The farmworkers who make the district function never appear. There are no field crews, no packing sheds, no processing lines, no conversations about working conditions, housing, heat, or health. Agriculture exists as an abstraction, not reality.

The same distancing occurs in the video’s treatment of the military. There are more shots of veterans’ graves than of active-duty service members. The graves can’t ask questions about VA benefits or military housing.

Even the moments of visible diversity follow a careful rule. The only scenes in which people of color are apparent are a shot with a football team and a scene of teenagers walking together. In each case, the people shown are in silent groups. They are not speaking or presenting. There are no adult people of color with expectations of their representative. In Fong’s video, diversity is aesthetic instead of participatory.

None of this is accidental. Across the entire video, institutions, infrastructure, machines, monuments, and symbols are meant to do the work of representation. Actual constituents are not. When people appear, they are either authority figures in controlled settings or background figures without voice. Labor is erased at ground level.

This all matters because true representation is not simply showing up for a photo opportunity. It requires a willingness to listen to people who do not control the setting, people who may disagree, and people whose demands may be uncomfortable.

A representative confident in his connection to the district would show conversations. Instead, his video shows careful selection of spaces where conversation cannot take place.

California’s 20th District is young, multilingual, labor-heavy, and economically complex. It depends on farmworkers, caregivers, service employees, educators, logistics workers, and healthcare staff. None of that reality is visible here. What remains is a sanitized version of the district that is orderly and silent.

That is not what a functioning democracy looks like.

Rep. Fong’s video tells us who he cares about, and who he is choosing to ignore.

The Class of 2021 and the Class of 2026 received measurably different accounts of what they were being sent to do, and why.

Both men identified the threats that the cadets will face.

Austin focuses entirely outward: near-peer competitors, pandemic, cyber weapons, terrorism, the painful aftermath of a long war. Even his most critical domestic observation of democracy under strain is framed as a headwind to navigate rather than an enemy to defeat. There is no internal villain anywhere in his speech.

Hegseth identifies threats on two fronts simultaneously. The foreign adversary is present but backgrounded, while the domestic institutional enemy receives substantially more attention: “foolish and feckless leaders,” professors advocating “anti-American ideologies right here in these halls,” and generals repeating diversity slogans “with a straight face on national television.”

The graduating cadets in Austin’s speech are focused on external challenges, while the graduating cadets in Hegseth’s speech are sent to correct internal ones.

Austin’s speech contains zero religious language; no scripture, no invocation of God, no devotional register at any point. His moral framework is entirely civic, and his closing line reflects that consistently: “For democracy. For liberty. For the Constitution.”

Hegseth opens with Isaiah 6:8: “Then I heard the voice of the Lord saying, ‘Whom shall I send, and who will go for us?’” Faith functions as the primary moral foundation of the speech. His closing follows the same logic: “May Almighty God bless the United States Army. And may Almighty God continue to bless this great republic.”

Both closings are similar in structure. Austin’s three terms are civic; Hegseth’s are devotional. Austin also returns to constitutional framing repeatedly across the body of his speech; “I still believe in the tenets of our democracy and the words of our Constitution,” while Hegseth references the Constitution once, in a list, as one unifying principle among several.

Both men addressed the ideas of diversity and unity.

Austin frames the elimination of discrimination as a national security matter, stating explicitly: “That is not just a matter of national principle. It is a matter of national security.” He singles out the fact that all three service academy senior cadets that year are women and presents it as a point of institutional pride.

Hegseth calls the diversity-as-strength formulation “the single dumbest phrase in military history,” adding that “we had generals saying this with a straight face on national television.” He adds, “Diversity is not our strength. Unity is our strength.” That unity is defined as oath-based and mission-based rather than demographic, and he instructs the graduating cadets directly: “You will not see color. You will not try to meet arbitrary quotas.”

Both speakers are answering what makes a military cohesive and effective, and the answers they provide are mutually exclusive.

Austin directs no vilifying language at any domestic target across the entirety of his speech. His one critical external passage, about foreign competitors claiming the future belongs to a model that stamps out freedom, names no one and uses no dehumanizing modifiers.

Hegseth uses multiple clusters of vilifying language directed specifically at domestic institutional actors: “foolish and feckless leaders,” “woke and weak leaders,” “political leaders with ideological agendas and weak military leaders who were just looking to curry favor,” and professors advocating “anti-American ideologies right here in these halls.”

Austin’s praise is intrinsic throughout. The cadets are capable and prepared because of what they built in themselves across four years at West Point, and the emotional engine of the speech is straightforward affirmation: “You are ready.” The greatness is located inside the graduates, with the institution serving as the mechanism that formed it.

Hegseth’s praise is substantially oppositional. The cadets are elite compared to peers who chose “normal colleges,” disciplined compared to the “woke and weak” leaders who preceded them, and the corrective force against an institutional slide that weak and foolish leaders allowed. “You will restore our Army” is a motivational claim that only works if the Army has been degraded.

Hegseth does include genuine intrinsic praise, particularly in the practical platoon-leadership advice section. But his oppositional motivational architecture is prominent in a way that Austin’s never approaches, and Austin contains no oppositional current at all.

A useful grammatical marker appears in both speeches in their use of the word “different.” For Austin it is a brief observation, while for Hegseth it is load-bearing, with the cadets defined repeatedly by what they have chosen to be set apart from.

Austin locates the soldier's moral foundation in civic institutions and personal virtue: "Leadership demands character. I still believe in telling the truth," and "the values that we uphold are the values that will hold us up." The Constitution and democratic ideals serve as the focus, and the framework is entirely secular from opening to close.

In summary:

Austin faces the threat landscape outward; Hegseth faces it both outward and inward.

Austin closes on the Constitution; Hegseth closes on God.

Austin calls diversity a national security asset; Hegseth calls it the dumbest phrase in military history.

Austin directs no vilifying language at any domestic target; Hegseth directs it at the military and academic establishment.

Austin motivates through intrinsic praise; Hegseth motivates through contrast with a domestic opponent.

Austin grounds moral conduct in civic virtue and constitutional values; Hegseth grounds it in faith.

These are not differences of tone or rhetorical style. They are different answers to the questions of what the military is for, who the enemy is, what holds soldiers together, and what a soldier ultimately answers to.

Transcripts:

Austin, 2021: https://docs.google.com/document/d/1LbPaX0PBoEYLEdmNx40sXur51t9nq6vCz43F6fAIXUA/edit?usp=sharing

Hegseth, 2026: https://docs.google.com/document/d/178zDvo9oRbzTmxztISGNdjiyOUIyrFnLfy9Z0JEQDcQ/edit?usp=sharing

The numbers tell a clear story about how his speeches have changed. This is a structural analysis only. It measures how speeches are built, not what they say.

All three transcripts come from the same source. They are formatted consistently, so the comparisons are clean.

STRUCTURAL CONTEXT

The first thing I measure is basic architecture: how long is the speech, how long are the sentences, and how easy is it to read?

The sentence length and readability scores here are the most dramatic numbers in this entire analysis.

Length:
2020: 2,908 words
2025: 8,158 words
2026: 8,396 words

The 2020 speech is less than a third the length of either later speech because in 2020, he usually stayed on script.

Average sentence length:
2020: 17.0 words
2025: 11.6 words
2026: 11.6 words

Shorter sentences aren’t automatically worse, but they usually reflect ad-libbing.

Flesch-Kincaid Grade Level:
2020: 9.3
2025: 5.0
2026: 5.1
The gap is nearly 4 full grade levels, and it appears completely between 2020 and 2025.

REPETITION PATTERNS
This measures which words and phrases repeat most across the speech, and how often. The pattern of what repeats tells you something about how the speech is organized.

Most repeated 3-5 word phrase in each speech:
2020: “the United States” — 7 times
2025: “you have to” — 19 times
2026: “the Coast Guard” — 39 times

COHESION — SENTENCE LEVEL (CCS-L)
This measures whether consecutive sentences share vocabulary. When you finish a sentence and start the next one, do you carry any words forward?
A score of 0 means sentences share nothing. A score of 1 means identical vocabulary. In natural speech, scores typically fall between 0.05 and 0.15.

CCS-L scores:
2020: 0.089
2025: 0.107
2026: 0.103

Sentence-level cohesion has not changed meaningfully across six years. That’s important context for what comes next.

COHESION — THEMATIC LEVEL (CCS-T)
This measures something different: not whether consecutive sentences share words, but whether topics return across the whole speech.

Think of it as a map. Does the speech revisit its own territory, or does it keep moving into new ground without coming back?

CCS-T scores:
2020: 0.556
2025: 0.704
2026: 0.852

These numbers rise sharply. A rising CCS-T score here doesn’t mean the speeches became more coherently structured. It means they loop back more.

In 2020, the speech moves through distinct topics and doesn’t return to them much. The score is moderate because the structure is more linear.

In 2025 and especially 2026, the same vocabulary clusters cycle back constantly throughout the speech. This is the repetitiveness you notice.

FRAGMENTATION (IFI)

This is where the change is sharpest.

Fragmentation measures mid-sentence pivots, self-corrections, incomplete thoughts, and the density of breaks in the flow.

A score near 0 means the speech moves cleanly. Higher scores mean more interruption.

IFI scores:
2020: 0.015
2025: 0.045
2026: 0.048

The 2020 speech scores 0.015. Both later speeches score roughly three times higher.

Mid-sentence pivots:
2020: 6
2025: 78
2026: 98

Short abandoned fragments (<5 words):
2020: 27
2025: 153
2026: 130

Em-dash interruptions:
2020: 11
2025: 132
2026: 169

Every component rises from 2020 to the later speeches. The pivot count goes from 6 to 78 to 98. The abandoned fragments go from 27 to 153.

A mid-sentence pivot is when a thought starts in one direction and redirects mid-stream before finishing. An abandoned fragment is when the sentence simply stops.

In 2020: 6 pivots in a 2,900-word speech. In 2026: 98 pivots in an 8,400-word speech.

Even accounting for length, the 2026 speech has roughly four times the pivot density of 2020.

What the numbers say together:

Sentence-level cohesion: stable across all three speeches.
Thematic return: rising sharply — but driven by looping repetition, not linear development.
Fragmentation: tripled between 2020 and 2025, then held.
Readability: dropped nearly 4 grade levels between 2020 and 2025, then held.

The 2020 West Point address is a structurally different speech from the two that follow it. It is shorter, more complex, less repetitive, less fragmented, and more linearly organized.

The 2025 and 2026 speeches are nearly identical to each other across every metric. The structural change happened between 2020 and 2025. It has not reversed.

Transcripts:

2020: https://docs.google.com/document/d/1lrELmQh5K1bxQBuID3sVLouNLFi_q-AFmijPEpJlwG4/edit?usp=sharing

2025: https://docs.google.com/document/d/1pkanR3euXBA_Yfcx0VKqn6bk22yfNuigiUTNb4RKqZ8/edit?usp=sharing

2026: https://docs.google.com/document/d/1FLR2RgINAPUwjUXpJsIoL9mPGfgANIIgefTAb4LA8LY/edit?usp=sharing