# Registered Predictions — written 2026-08-22, before the October data exists **Everything below is a falsifiable statement made in advance. Each one carries a threshold, a stated failure condition, a confidence, and how it will be resolved. The outcome rows get APPENDED in October, whatever they show — nothing here is edited after the fact.** The point of this document is that it can only be written now. A prediction made before the data arrives and then met is worth more than any amount of post-hoc explanation, and a prediction that misses is worth more still, because it is the only thing that can actually correct us. **Ground rule adopted with this document:** confidences are stated as percentages and are meant literally. If the 60%-confidence predictions come in at 90% right, the confidences were badly calibrated and that is itself a finding to report. --- ## A. The tenure–money gradient **A1 — already registered 2026-08-21 (Source Log row 106), restated here unchanged.** At the ~15 October FEC Q3 refresh, the gradient holds: overall Pearson **r within 0.70–0.90**; incumbents-only **r ≥ 0.60**; positive sign within both parties. *Basis:* r = 0.82 all / 0.70 incumbents-only on receipts through 2026-06-30. *Falsifier:* overall r < 0.5, or a sign flip in either party. **Confidence: 80%.** **A2 — NEW. The gradient is a CHAMBER phenomenon, not a battleground one.** When the join is extended to all 435 districts (test E1-c), the all-435 correlation will land **within ±0.15 of the battleground figure**. *Reasoning:* the E3 stock-conduct test already pointed this way — battleground members were not an excess over the chamber (2/38 comparable = 5.3% vs chamber 14.6%), and the honest read there was "Congress is the story." Corporate PAC money tracking seniority is not plausibly a swing-district behaviour. *Falsifier:* all-435 r differs from battleground r by more than 0.15 in either direction. **Confidence: 70%.** *Why this matters even if it holds:* it changes the headline. "Battleground incumbents draw corporate money with tenure" and "Congress draws corporate money with tenure" are different stories, and publishing the first while the second is true is exactly the narrative error this site exists to catch. **A3 — NEW. Tenure will survive the committee-gavel control, weakened but standing.** Controlling for a committee chair / ranking-member flag (test E1-d), tenure's partial correlation with corporate-PAC share stays **≥ 0.50**. *Reasoning:* gavel proximity is real, but tenure also proxies network age, donor familiarity and name recognition — things that accumulate without a gavel. A full collapse would mean seniority buys nothing except committee power, which seems too clean. *Falsifier:* partial r < 0.50, i.e. the gradient largely collapses into the committee flag. **Confidence: 65%.** **A4 — NEW. The 2024 replication holds directionally but weaker.** The 2024-cycle rebuild (test E1-b) produces a positive correlation **between 0.5 and 0.85** — same sign, plausibly softer, because 2024's competitive-district set had a different incumbent mix. *Falsifier:* 2024 r < 0.3 (the plan's stated falsifier) or a negative sign. **Confidence: 70%.** --- ## B. Communication style **B1 — The broadcaster null HOLDS at scale. This is the big one.** In the 59-race sweep, candidates with a professional media background will **not** differ significantly from those without on social-register reading grade — **p > 0.05**, two-sided. *The test set already exists:* Flint (17 years statewide radio), Coleman (talk radio and TV meteorology), Feely (NFL then CBS), Crank, plus Stelson. *Reasoning:* we killed this once already with a pre-registered control, and the seven-subject benchmark scattered from 6.9 to 11.7 — wider than the candidate field's own spread. A signal that scattered at n=7 does not usually tighten at n=100. *Falsifier:* p < 0.05 with the media group reading lower. **Confidence: 85%.** **B2 — The three-voice architecture holds at 100%. An absolute claim, deliberately.** Every written attributed quotation in the sweep measures **exactly 0.00** third-person self-reference. Not "almost all" — every one. *Basis:* 23 for 23 so far, across House candidates, senators, three Cabinet departments, an intelligence agency and both parties. *Falsifier:* **a single** quote cell above 0.30 per hundred words. **Confidence: 80%** — and stated as an absolute precisely because a single counterexample kills it. That is the most useful shape a prediction can have. **B3 — The under-55 / 55+ register-gap split does NOT survive. A prediction against our own finding.** At n > 30 the age split fails to reach significance, or fails outlier removal again. *Reasoning:* it has already been downgraded once from a "gradient" to a coarse split; it fails when Kaptur is removed; it is operationalization-sensitive (the release→social metric shows r = +0.05, no effect at all); and Merrin already sits on it as a standing exception. *Falsifier:* the split holds at p < 0.05 **and** survives removal of the oldest subject. **Confidence: 65% that it dies.** **B4 — Executives code-switch harder than candidates, and it holds.** Adding more national officeholders, the quote→social gap for national-office holders stays **≥ 3.5 grades** against a candidate median near 2.2. *Basis:* Kelly +6.1, McMahon +5.7, Gabbard +5.5, Duffy +4.4 — but n = 4, which is why this is the shakiest style prediction here. *Falsifier:* the national-officeholder median falls below 3.5, or the difference from candidates fails significance. **Confidence: 55%.** Barely better than a coin flip, and stated that way. --- ## C. Platform and collection **C1 — The Facebook bound bites hard.** In the 59-race sweep, **at least 15%** of candidates will have an uncollectable social register. *Falsifier:* under 15%. **Confidence: 70%.** **C2 — The Bluesky asymmetry is real and will show up as a partisan skew in what we CAN collect.** Among candidates whose social cells meet the minimum, Democrats will outnumber Republicans by **at least 1.3 : 1**. *Reasoning:* Bluesky's open API is the single most productive collection route we have (Dunlap's 29 posts, the richest cell in the project) and Bluesky carries a documented heavy left tilt. This is the limitation we elevated above the Facebook problem, and it deserves a number attached rather than a caveat. *Falsifier:* the ratio is below 1.3 : 1, or it runs the other way. **Confidence: 60%.** *Note:* if this lands, it is a finding **against our own dataset's fairness**, and it gets published as prominently as anything else here. --- ## D. Roster volatility and the index itself **D1 — At least two more of the 59 races materially change before Election Day.** A withdrawal, replacement nominee, party switch, suspension or death, between 2026-08-22 and 2026-11-03. *Basis:* three races changed in roughly ten weeks after the roster froze — about 5% of districts per quarter. Ten weeks remain. *Falsifier:* fewer than two. **Confidence: 75%.** **D2 — Alaska's suspended candidate does not appear as the effective nominee.** Matt Schultz suspended his AK-00 campaign; I predict the race is not contested by him in any meaningful sense at Election Day. *Falsifier:* he actively contests it. **Confidence: 70%.** --- ## E. The age work **E1 — Between two and four of the seven blanks close by Election Day.** General-election coverage prints ages that primary coverage did not. *The seven:* Regnitz, Ollivant, Desmond, Powell, Harding, Berman, Flores. *Falsifier:* fewer than two, or more than four. **Confidence: 55%.** *Sub-prediction:* **Powell and Harding are the likeliest to close**, because both need only one outlet to state in its own voice a number that already exists on a questionnaire. **E2 — Coleman's asterisk resolves to 67, not 66.** **Confidence: 57%** — stated to the point, because that is literally what the birth-window arithmetic gives (182 days versus 139). This is the most honest confidence figure in the document and the one I would most like to see checked, because it tests whether a stated probability means anything. **E3 — The other four asterisks do NOT resolve.** Schultz, Vo, Knickerbocker and Flint still carry a ±1 at Election Day. *Reasoning:* resolving them needs a date of birth, a birthday mention, or an age printed after Election Day itself. General-election coverage restates the same age it already printed. *Falsifier:* two or more resolve. **Confidence: 70%.** **E4 — The Ohio voter file, if reached, makes Coleman and Knickerbocker exact.** *Falsifier:* the file is reached and does not carry usable dates of birth for them. **Confidence: 85%** conditional on access. --- ## F. The prediction about the predictions **F1 — At least one currently-published finding is retracted or downgraded by the October pass.** *Reasoning:* this project has retracted or corrected something in every single round it has run — the age gradient, the Nez uniqueness claim, the spoken-simpler rationale, the occupational asymmetry, the broadcaster convergence, the executive-quote-register claim, two published all-caps figures, and two age entries. The base rate is not low. It is 100%. *Falsifier:* nothing gets retracted or downgraded. **Confidence: 80%.** **F2 — My confidences in this document are overconfident on average.** Across the roughly sixteen predictions above, the share that come true will be **lower** than the mean stated confidence (about 70%). *Reasoning:* this is the direction almost everyone errs, and I have no particular reason to think I am the exception. Registering it makes the calibration itself checkable. *Falsifier:* the hit rate meets or exceeds the mean stated confidence. **Confidence: 60%.** --- ## How these resolve Each prediction gets an **outcome row appended** to the Source Log in October — never an edit to the row above it. The resolution for each is already specified: a threshold and a falsifier, both written before the data exists. A prediction that misses is not deleted, softened, or reinterpreted. It is marked MISSED, with the actual value beside it, and the reason it missed becomes part of the record. That is the entire point of writing them down today. *Registered 2026-08-22. Resolution due after the Q3 FEC refresh (~15 October) and the 59-race style sweep. Nothing in this document may be revised — corrections go in the October outcome rows.*