Can Math Prove Corruption? Turning Tariff Favoritism Into a Bayesian Number

Two senators say tariff breaks are going to the politically connected. Everyone can see the pattern. The harder question: can you turn that suspicion into a number a hearing room has to take seriously?

Short answer: you can't prove intent. But you can measure how much better connected companies do, holding merit constant. That measurement has a name — Bayesian inference — and this film builds it, piece by piece, and then says plainly what it cannot say.

WHAT'S IN IT

• The correction the whole film turns on: not "can math prove corruption?" but "can math measure favoritism?"
• The February 2026 Wyden–Van Hollen letter on an opaque exemption process — and the catch it names: no formal public process was ever opened, so you work backward from the outputs.
• Measuring "connected" in three channels: campaign contributions (normalized by firm size), lobbying disclosures (down to the named issue and the revolving-door hires), and public gifts — the Rolex desk clock and the engraved gold bar of November 4, 2025, and Apple's gold-based plaque.
• The equation: a binary outcome, a logit link, two teams of variables (economic merit as the control, political connection as the allegation), and one coefficient — beta-pol — that carries the whole case. Plus the hierarchical layer that asks whether a firm beat its own industry's odds.
• Ten days in November: a 39% tariff, an icy relationship, the gifts, then a framework deal at 15% — and how a changepoint model tests whether the series really broke there.
• Priors from published research (Lopatin et al. 2024 on the first-term exclusion process), the posterior update, and the Bayes Factor as referee.
• The outside check the market provides: an event study where a zero abnormal return is the loudest result.
• What the math cannot say: motive. The model measures advantage, never intent — and the same research that found politics in one tariff program found none in steel and aluminum. A method that can exonerate is a method that can accuse.

HONEST LABEL The events are real. The method is real. The decisive numbers in this story are a worked scenario from the source essay, shown to demonstrate what the analysis would return if the data cooperate — nobody has run it on the real dockets yet. Skepticism means keeping those three sentences separate.

CHAPTERS 0:00 The question a hearing room can take seriously 0:20 Prove → measure: the correction 0:35 Act I — The Senate Finance letter and the missing process 1:31 Act II — Measuring "connected": money, lobbying, gifts 2:59 Act III — The equation: logit, beta-pol, hierarchy 4:13 Act IV — Ten days in November 5:13 Act V — Priors, posteriors, and the Bayes Factor 6:19 Act VI — The market already knew 6:57 Act VII — What the math cannot say 8:07 Verdict 8:34 Your Turn

YOUR TURN — paste this into Claude: I have a dataset of regulatory decisions: outcomes, applicant characteristics, and political-connection proxies. Walk me through building a hierarchical Bayesian logistic model to test favoritism against merit — (1) which priors and why; (2) how to partition industry effects from firm effects; (3) what result would falsify the favoritism hypothesis. Run it on any policy domain you care about. The machinery transfers.

SOURCES Wyden–Van Hollen letter and press release (finance.senate.gov, 4 Feb 2026) · Forbes and Axios (14 Nov 2025) on the Swiss deal and the gifts · Bloomberg (28 Nov 2025) and swissinfo on the Swiss complaint · Fortune / AppleInsider (Aug 2025) on the Apple pledge and plaque · CNBC and Fortune on the ballroom donor list · Lopatin et al. 2024, Economics & Politics, on political contributions and Section 301 exclusions (and the steel/aluminum null). Source essay: skepticism.ai, "When Democracy Becomes a Math Equation." Every claim was checked against its source before rendering; the essay's illustrative numbers are labelled as such on screen and aloud.

AI DISCLOSURE Four short scene-setting clips (the hearing room, the gold clock, the plaque, the Swiss bank facade) are AI-generated illustrations, not footage of real events. Narration is a synthetic voice (Liam, in for Bear).

Series: Computational Skepticism · Conducting AI @NikBearBrown

What this video covers

• 0:00 — The question a hearing room can take seriously
• 0:20 — Prove → measure: the correction
• 0:35 — Act I — The Senate Finance letter and the missing process
• 1:31 — Act II — Measuring "connected": money, lobbying, gifts
• 2:59 — Act III — The equation: logit, beta-pol, hierarchy
• 4:13 — Act IV — Ten days in November
• 5:13 — Act V — Priors, posteriors, and the Bayes Factor
• 6:19 — Act VI — The market already knew
• 6:57 — Act VII — What the math cannot say
• 8:07 — Verdict
• 8:34 — Your Turn

Transcript

A lightly formatted transcript of the video, with timestamps so you can jump to any section on YouTube.

The question a hearing room can take seriously

[0:00] Bonjour Liam in for bear, to senators say tariff breaks are going to the politically connected. Everyone can see the pattern. Today's question is harder. Can you turn that suspicion into a number of hearing room has to take seriously? Short answer, you can't prove intent, but you can measure how much better connected companies do holding merit constant. That measurement has a name, Bayesian inference. Let's build it. Someone typed the question the

Prove → measure: the correction

[0:21] Obvious way, can math prove corruption? Then stopped, crossed out prove, crossed out corruption. What's left is the question statistics can actually answer. Can math measure favoritism? That correction is this whole film. Picture the Senate Finance hearing room. A

Act I — The Senate Finance letter and the missing process

[0:36] Single sheet slides across the table. Companies that got tariff exemptions on the left and next to each their contributions, their lobbying, their donations. The correlation looks fabricated. It isn't. February 4th, 2026. Senators Wyden and Van Hollen put it in writing. An opaque process that appears to favor the politically connected. One that, quote, opened the door to corruption. But appears to favor is rhetoric. The room needs math. Here's the catch, the letter itself names no formal public process was ever opened for these exemptions. No portal, no published criteria. Small businesses had no door to knock on. The machinery you'd want to audit is the machinery that was never built. Which is exactly the situation Bayesian inference was built for. You can't see inside the black box, so you work backward from its outputs. Every exemption granted is a data point. Enough data points and the hidden decision rule starts to show its shape. Act two, measuring connected. Political

Act II — Measuring "connected": money, lobbying, gifts

[1:32] Influence is fuzzy, relationships, favors, implied understandings. A model can't eat fuzz, so you compress it into three measurable channels. Money given, lobbying bought, and gifts made in public. Channel one, campaign contributions. Every itemized receipt is a public record you can download in bulk. And you don't count raw dollars, you normalize by firm size. So a million from a giant and 10 grand from a machine shop are comparable signals. Channel two, lobbying. Disclosure forms name the exact issues bought. Firms literally file section 301 tariffs in quarterly reports, and they name the hires, former trade officials, ex-commerce staff. The revolving door leaves a paper trail. Channel three is the strange one. November

4th, 2025, a Swiss delegation visits the White House. The CEO of Rolex presents a custom gold desk clock. The head of a gold refiner presents a 1 kg bar engraved and 47 worth $130,000. It has company. In August, Apple's chief executive presented a glass plaque on a 24-karat gold base alongside a $600 billion investment pledge. These aren't bribes in any legal sense. They're signals, public demonstrations of alignment, so the model codes them honestly. Gift equals one. Now, the question is operational. Holding a company's economic merit constant, the legitimate case for relief, does turning these three dials up raise the probability of getting an exemption? That's testable. >> Act three, the equation.

Act III — The equation: logit, beta-pol, hierarchy

[3:00] >> The outcome is binary. A company gets the exemption or it doesn't, one or zero. The probability of the one, that's what the whole model exists to explain. But, probability isn't linear. Doubling a lobbying budget doesn't double your odds. The relationship curves, and it saturates. The logit link bends a straight line of predictors onto the zero-to-one scale. This one curve is the workhorse of every yes or no question in statistics. Inside the model, two teams of variables, X econ, the legitimate reasons, no alternative suppliers, severe harm, strategic products. X pol, contributions, lobbying, ballroom donations, gifts. The econ side is the control. The pol side is the allegation. And one coefficient carries the whole case, beta pol. If it's posterior distribution sits clearly above zero after merit is controlled, political connection is buying probability. The crony capitalism hypothesis compressed into a single Greek letter. One more layer because companies don't float free, they nest inside industries. If all consumer electronics got relief, Apple's win might be a sector tied, not a favor. A hierarchical model gives each industry its own baseline, so you can ask, "Did this firm beat its own industry's odds?" Act four, 10 days in November.

Act IV — Ten days in November

[4:14] The timeline, August, US-Swiss trade talks stall and a 39% tariff lands on Swiss imports. September, diplomats call the relationship icy. November 4th, the delegation, the clock, the gold bar. November 14th, 10 days after the gifts, the framework deal lands. The Swiss tariff drops from 39% to 15. Swiss lawmakers were unamused enough to ask their own prosecutor to investigate the gifting. Model it honestly, an intervention variable flips on at the meeting. A change point search asks where the series structurally breaks. Run the essay's version of this analysis and the break lands on the deal date with probability near 98%. The essay's worked numbers, remember, not a published study, but the shape of the test is exactly right. Notice what the Bayesian frame buys you here. Frequentist statistics asks, "Could this be random?" Bayesian statistics asks the question the hearing room actually cares about, "Given everything we know, how probable is it that the gifts moved the policy?" Act five, priors, posteriors, and the

Act V — Priors, posteriors, and the Bayes Factor

[5:14] Weight of evidence. Bayesian models start from what's already known, and something is known, peer reviewed research on the first Trump administration's exclusion process found political contributions raised approval odds roughly four percentage points per standard deviation. That published effect becomes the prior. Then the data speaks. If the new evidence shows a stronger effect, the posterior shifts right. If the effect vanished, the data drags the belief back toward zero. That's the whole Bayesian bargain. Your starting belief is explicit, and the evidence is is to overrule it. And you don't just estimate you referee two hypotheses on the bench, favoritism drives outcomes or merit does. The Bayes factor weighs which one the data prefer. Above 10, strong evidence. Above 100, decisive. The essay runs the full scenario hypothetically. Contributions worth about six points of probability. Ballroom donor status worth nearly 15 on its own. A Bayes factor of 247. Decisive if real data land the way the essay sketches. Nobody has run this on the real docket yet. That's the point of showing the machinery. Act six, the market already knew.

Act VI — The market already knew

[6:20] There's an elegant outside check, traders. An event study measures the abnormal return, how much a stock jumps beyond what the market predicted the day an exemption is announced. In the essay's telling, an unconnected company wins relief and its stock pops about 2% genuine news. A connected company wins and the price barely moves, zero surprise. The market had already priced the favor in. A zero is loud. No abnormal return means the announcement carried no information. The outcome was predictable from access alone. When the trading floor and the posterior agree, each one is validating the other. Act seven, what the math cannot say.

Act VII — What the math cannot say

[6:58] The model detects patterns, it cannot read minds. Maybe relief is traded for loyalty or maybe connected companies simply argue economic harm better with better lawyers on better paper. Beta Paul can't split those. The math is agnostic about motive, an honest analysis says so. Still in the other side because the data does, unconnected companies do win exemptions. The Swiss deal came with a 200 billion dollar investment pledge, a legitimate confound the gold bar story skips. And the same research that found politics in one tariff program found none in steel and aluminum. The method can exonerate, that's what makes it credible when it accuses. So what would falsify favoritism? Run the model on the real dockets and watch Beta Paul's posterior collapse onto zero. Watch connected and unconnected approval curves overlap. The hypothesis sticks its neck out, which is precisely what makes it worth testing. And the honest label one last time, the decisive numbers in this story are a worked scenario. The essay demonstrating what the analysis would return if the data cooperate. The events are real. The method is real. The verdict is waiting on the data. Skepticism means keeping those three sentences separate. Let's recap with Claude. Favoritism is

Verdict

[8:07] Measurable in principle. Define connections contributions, lobbying and gifts, then test whether they move approval odds with merit held constant. The machinery is a hierarchical Bayesian logic. Published research as the prior, Bayes factors as the referee, market reactions as the outside check, and the limits are part of the result. The model measures advantage, never intent. And the essay's decisive numbers are an illustration awaiting real data. Your turn. Here's the prompt. Read it

Your Turn

[8:34] With me. I have a data set of regulatory decisions, outcomes, applicant characteristics, and political connection proxies. Walk me through building a hierarchical Bayesian logistic model to test favoritism against merit. One, which priors and why. Two, how to partition industry effects from firm effects. And three, what result would falsify the favoritism hypothesis. Run that on any policy domain you care about. The machinery transfers. When democracy becomes a math equation, lay them in for bear.