2026 U.S. Senate: Pre-Registered Predictions
July 30, 2026 · five races · two panel tiers · frozen before fielding
Pre-registered before fielding, July 29, 2026: Tier 1 battery frozen in git at 6:08 PM EDT (commit b7c50f3, panel amendment ad9e2c0), fielded 6:11\u20136:14 PM. Tier 2 frozen at 6:42 PM (979867c), fielded 6:42\u20136:44 PM. Resolves against certified results November 3, 2026.
Five Senate races that will decide the majority, fielded in two tiers. Tier 1 (GA, OH, TX): statewide live panels of ~20,600 simulated residents per state with persistent memory, daily news injection, and regional training corpus, n=1,000 per state. Tier 2 (NC, ME): census-weighted agents generated from ACS regional frames and relabeled to the state \u2014 no memory, no news layer, party mix from regional rather than state priors. Tier 2 is the weaker instrument and is labeled directional; publishing both tiers against the same resolution date is a deliberate test of how much the live-panel infrastructure adds.
Same scoring everywhere: first-token logprobs and Dirichlet electoral calibration, as in Issues 01\u201302. Raw and calibrated toplines published together with 95% bootstrap intervals. Michigan is excluded until its August 4 primary resolves \u2014 pre-registering a hypothetical matchup would weaken the exercise.
Synthetic calibrated panels \u2014 not probability samples. No humans were interviewed.
Measured defect in Tier 2, disclosed at publication
The ACS regional frames over-represent high earners and under-represent independents, and we can show you the size of it from the crosstabs below. Per 1,000 agents our Tier 1 live panels put 162 to 172 in the $125k-or-more bracket, in line with the roughly 16 to 20 percent these states actually report. Tier 2 puts 348 in North Carolina and 420 in Maine, close to double. Independents are also under-represented: 17 percent of the Maine sample against roughly a third of its real registered voters. Income and party both predict vote choice, so read NC and ME as output from a known-misspecified instrument, not as forecasts. Tier 1 (GA, OH, TX) is unaffected; those run on real statewide panels whose income distribution matches their states.
Maine has a second and larger problem that the frame does not explain. It is the only state here whose panel holds more Democrats than Republicans, 372 against 342, and yet it returns Collins by 31.9. Every other state's direction tracks its party composition. Maine inverts by roughly 62 points, because Democratic loyalty collapses to 74 percent when it runs 98 to 99 percent everywhere else, and independents break 88 percent to Collins when they break 5 to 38 percent for Republicans elsewhere. That pattern looks like the model reproducing Susan Collins's long-standing reputation for winning independents and crossover Democrats as a stored fact about her, rather than simulating 2026 vote intention. Refielding on a state frame will not fix it, so we are treating Maine as a diagnostic case rather than a forecast and reporting whatever November returns.
Georgia
Tier 1 · Live panelGeorgia statewide live panel · n=1,000 of 20,638
Jon Ossoff (D) vs Mike Collins (R)
Ossoff +11.5 · Dem share 55.7% calibrated (95% CI 53.3–58.1) · 57.3% raw
- Jon Ossoff55.7% (raw 57.3%)
- Mike Collins44.2% (raw 40.6%)
- Undecided0.1% (raw 2.2%)
Recent public polls: Wick (Jun 27–30) Ossoff +4; Beacon/Shaw (Jun 23–27) Ossoff +13; Echelon (Apr) Ossoff +7. Our read sits inside the recent poll range, at the high end.
Ohio
Tier 1 · Live panelOhio statewide live panel · n=1,000 of 20,638
Sherrod Brown (D) vs Jon Husted (R) — special election
Brown +18.4 · Dem share 59.1% calibrated (95% CI 57.0–61.8) · 60.3% raw
- Sherrod Brown (D)59.1% (raw 60.3%)
- Jon Husted (R)40.7% (raw 36.4%)
- Undecided0.1% (raw 3.3%)
Polling averages (early Jul): Brown +1.3 (270toWin), +1.5 (DDHQ). Our panel is far more bullish on Brown than human polls — driven by independents breaking 86/5 for Brown. Pre-registered as-is; November decides.
Texas
Tier 1 · Live panelTexas statewide live panel · n=1,000 of 20,637
James Talarico (D) vs Ken Paxton (R)
Paxton +10.7 · Dem share 44.6% calibrated (95% CI 42.3–46.6) · 43.7% raw
- James Talarico44.6% (raw 43.7%)
- Ken Paxton55.3% (raw 54.8%)
- Undecided0.1% (raw 1.6%)
Polling averages (Jul): Talarico +0.4 — a tossup (NYT/Siena Jun: tied 47/47). Our panel is far more bullish on Paxton than human polls. Pre-registered as-is; November decides.
North Carolina
Tier 2 · DirectionalCensus-generated agents (ACS South frame, state-relabeled) · n=1,000 generated
Roy Cooper (D) vs Michael Whatley (R) — open seat
Cooper +3.3 · Dem share 51.6% calibrated (95% CI 49.2–54.1) · 52.3% raw
- Roy Cooper51.6% (raw 52.3%)
- Michael Whatley48.3% (raw 45.7%)
- Undecided0.1% (raw 2%)
Polling averages (mid-Jul): Cooper +8.4; PPP (Jul 10–11) Cooper +4. Our directional read is tighter than the polling average.
Maine
Tier 2 · DirectionalCensus-generated agents (ACS Northeast frame, state-relabeled) · n=1,000 generated
Troy Jackson (D) vs Susan Collins (R)
Collins +31.9 · Dem share 34.0% calibrated (95% CI 32.1–35.9) · 29.9% raw
- Troy Jackson34% (raw 29.9%)
- Susan Collins65.9% (raw 68.3%)
- Undecided0.1% (raw 1.8%)
No public poll of this matchup existed at pre-registration — Jackson became the nominee July 25 after Platner withdrew. The panel gives Collins heavy crossover: 25% of panel Democrats and 88% of independents pick her. Collins outperformed her polls by ~9 points in 2020; November decides whether this read or the fundamentals-based skepticism is right.
Where we differ from human polls
Our Ohio read is well to the left of the polling average (Brown +18.4 vs +1.3), our Texas read is well to the right of it (Paxton +10.7 vs a Talarico +0.4 tossup), and our Maine read gives Collins a landslide no fundamentals model would print. North Carolina and Georgia sit near the polls. The divergences run in both directions, so this is not a uniform partisan lean \u2014 it is a different signal. Partisans vote their party at 92\u201399% in four of the five states; Maine is the lone break, with 26% of panel Democrats crossing to Collins, and we read that as an artifact rather than a finding. Elsewhere the differences come from how independents break. If the panel is wrong, the scorecard will say so in November.
Known instrument effect: the panel allocates far less mass to \u201cUndecided\u201d (0.1\u20133% raw) than human polls do (4\u201310%). Two-way vote share is the primary pre-registered metric.
Data downloads
Per-respondent probability distributions (n=5,000) available to researchers on request.