# VERDICT - Vugol Tensor (as of 2026-10-15; election 2026-11-03)

**Is there a defensible edge? Not in the point estimate. Yes, in honest uncertainty. Edge size: about 0 to 0.1 pts of average absolute error; the real value is calibrated intervals (sampling-only intervals cover 42% of outcomes at nominal 90%; total-error intervals cover about 89%).**

## Supported by evidence

- Polls err by about 4-5 pts per race (MAE 4.2 for polled races 2002-2024, [RUN]) and share a cycle-level error with SD about 2.4-2.7 pts (14 cycles). AAPOR 2016/2020/2024 and Shirani-Mehr et al. agree in kind (see FINDINGS.md).
- The recent direction (polls too favourable to Democrats in 2014, 2016, 2020, 2024) is real, but the pooled average over 14 cycles is +0.75 (t=1.06), and 2012 and 2018 went the other way. The miss in one cycle did not predict the next (corr about 0).
- Shrinking toward the state's last presidential margin gives a small gain (-0.11 pts MAE; 10 of 12 cycles; CI includes 0).

## Not supported

- Any correction that shifts margins toward Republicans (or Democrats) because of recent history: no better MAE out-of-sample, and it hurts in other cycles.
- A signed late-news (GDELT) term: unidentifiable with about 5 usable cycles; Comey-type effects remain disputed. Poll trajectories show late moves partly reverse.
- Deep learning, naive Bayes, gradient boosting on this problem: 14 independent cycles cannot support them.
- Mode, house-effect, regional covariance, and fundamentals components: plausible, **not tested** (no suitable data pulled).

## Recommended v0 for 2026 (midterm)

1. Point estimate = plain, de-duplicated, one-vote-per-pollster average of non-partisan polls in the last 21 days (min 2 pollsters), optionally shrunk toward state presidential lean with weight about 0.9 on polls. **No partisan bias adjustment.**
2. Show uncertainty as total-error intervals (about +/-8.5 pts at 90% for a single race margin) rather than margins of error. Present "polls can miss in either direction by this much, and miss together."
3. Joint outputs (chamber control, number of races flipping) from a simulation with a shared national error draw (SD about 2.5) plus race-level noise (SD about 4.6). Label as an experiment.
4. Pre-register now, before Nov 3: parameters, windows, the formula in MODEL.md section 8.2, and the scoring plan (compare against plain average and an external aggregator on 2026 results). Freeze it.
5. Track after election: 2026 cycle bias will be one more cycle in the table; the sign is unknown in advance.

## Risks

- Implied precision: a model that looks sophisticated but has no demonstrated point-estimate advantage; the sophistication can be mistaken for accuracy.
- Partisan perception: any shift, in either direction, will be read politically. Zero-mean design is both the honest choice and the defensible one.
- Overfitting to 2016-2024; survivorship (polled races are competitive); data provenance (Silver Bulletin file URL is opaque; 538 archive is being removed); `margin_actual` not independently verified.
- 2026 fundamentals (generic ballot, approval) were not tested here.
- A 19-day horizon: no time for new data work before election day.

## How to present it

- Publicly: "Vugol Tensor is an experimental uncertainty model. It does not predict that polls will favour either party. Over 14 elections from 1998 to 2024 polls overstated Democrats in some cycles and Republicans in others, and the typical miss for one race was about 4 points. We show ranges that reflect that." Include coverage statement and what was not tested. No letter grades or accuracy claims. Avoid the word "correction" for the centre.
- To analysts: publish the formula, code (`backtest.py`, `bootstrap.py`, `late_shift.py`), data sources and hashes, the negative results (global shift and persistence rules do not beat the plain average), the one weakly positive result (state-lean shrinkage), and the list of untested components. Invite replication.