# Scores v5 — FLAGS (known limits and open questions, 2026-10-18)

1. **The Maine baseline is an estimate, not a measurement.** D+4.8 is the mean of three elections with a standard error of about 2.4
   points; the two-, four-, five- and seven-election means run from D+6.5 to D+7.6. Everything built on the Partisan EAI inherits that
   uncertainty as a shift of the whole scale: rankings do not move, the side a town "favors" in a neutral year can for towns within
   a few points of EVEN. The three elections all had the same Republican nominee; the rule is kept for its long-run properties and
   the value is re-estimated after each presidential election.
2. **"Neutral" means a tied national two-party presidential vote with the same candidates everywhere.** Midterm neutrality cannot be
   tested separately in Maine (every Maine midterm race carries a large candidate effect). Town-specific candidate appeal, local spending
   and hometown effects are not removed by any aggregate measure; they sit in the two-cycle noise (`cycle_change_sd`).
3. **House add-back** uses the district's lean in the same cycle's statewide races; House-only cycles (1992, 1998, 2010) borrow the mean
   of the adjacent cycles. Maine's own relative lean is 0.0 rolling but ±0.2 per cycle for this reason.
4. **Two-cycle window** (grid search): one strange local year leaves a mark; eight towns have one cycle (`n_cycles`).
5. **Small towns.** Under 100 two-party votes the raw vote moves ~20 points between cycles; ranks among small towns are noise.
6. **U.S. House candidate effects** remain at one third of a cycle; the add-back restores the district's partisanship but not a
   candidate's town-by-town variation.
7. **RCV races** enter as first choices. **Shared results lines**: votes divided by registration (`shared_results_line`).
8. **Two registration files on one page, by design**: Partisan Edge and the registration section use the latest file (2026-03-07);
   the Hidden Lean regression uses the Nov 2024 file in force at the vote. Both dates are printed. The 2025 list maintenance moved
   ~136,000 voters to inactive status unevenly between the two files.
9. **Hidden Lean** is a residual from one statewide line (two regressors): unenrolled lean, crossover and differential turnout together.
   Maine's own value is −0.7, not 0 (weights are 2024 presidential votes of rated towns).
10. **Voter-file headroom** uses 2026 registrant counts against the election's turnout; coverage 96% (2022) / 98% (2024); ~1,230 voters
    unmatched to a town. Turnout gap stability between elections r ≈ 0.5: headroom is a scale, not a forecast. No town meets the
    two-election turnout rule under v5 (one did under v4's zero).
11. **Implied unenrolled lean** (data only) is assumption-driven (93% loyalty); impossible values where enrolled partisans cross over.
12. **Registration history** before 2007 is a single "registered voters" count; the baseline for 1990–1994 history rows rests on the
    1980–1988 elections (Reagan twice), which is far from today's Maine.
13. **Thresholds (7 / 15 / within-5 / two-election headroom rule), the baseline window and the ⅓ House weight are fixed until the
    Nov 2026 results are scored** (README parking list).
14. **15% rule edges**: 1994 CD 2 (13.6%), 2002 Governor (11.4%), 2014 CD 2 (11.2%) qualify; 1996 President (17.6%) excluded.
15. **AVR** (2022) enrolls new voters as unenrolled by default: the unenrolled share rises for administrative reasons too.
16. **Forecast file** `forecast_2026-11-03/`: the Oct 17 freeze used v4's relative lean (district-referenced); the Oct 30 refreeze uses
    v5's. Both are kept and both will be scored.
17. **National totals** are typed constants (FEC official results, 1980–2024); verify against the FEC publications before the 2028 update.

18. **Battleground classes merged** (2026-10-19, Matt): `campaign_class` battleground_turnout / battleground_persuasion merged into `battleground` in the script and the tables (no town met the turnout rule; two State House districts did). `campaign_reading` for those rows is now "Battleground: a neutral year reads X here." Headroom columns unchanged. Site shows the class as a colored badge (safe D dark blue, lean D light blue, battleground gold, lean R light red, safe R dark red).