PoPBaseline v1
PoPBaseline v1
PoPBaseline-v1.0 is a separate, opt-in opponent for the frozen ElectionSimulator v1.0-rc1 (Candidate A). It is not a poll aggregator and it does not change Candidate A or any certified RC1 artifact.
Reconstructed method
The baseline starts with exactly one stored Poll of Polls composition at the requested origin. It then samples historical joint changes over 21-, 28-, and 35-day windows, reverses each sampled change with equal probability, and combines equal-sized batches from the three windows. Changes are represented in CLR coordinates, so the perturbation preserves a positive nine-category composition and sums to 100 percent. No current-state uncertainty, election residual, geography, or seats are added.
The first-party 2018 page also specifies a simple support-vote rule:
s_support = s_sim + (-0.6 * abs(s_sim - 4) + 1.2) * X, X ~ Normal(1, 0.7)
for 2 < s_sim < 5, otherwise unchanged. Positive transfers are taken proportionally from same-block parties’ support above 4 percent. The canonical data model folds FI into REST, so FI is not invented as a separate recipient; the representable default recipients are L, C, KD, MP, and V. Support voting can be disabled in the configuration for the dynamics-only benchmark. Transfers are applied sequentially in the declared target order (L, C, KD, MP, V); that order is part of the baseline version’s deterministic configuration because the source does not specify a simultaneous multi-party reconciliation rule.
For exact 7/14/28/56/84/112-day benchmark horizons, a final historical step is fractionalized linearly in CLR space when the horizon is not divisible by the selected window. This is an explicit approximation because the source pages do not specify how to handle partial final steps.
Source record
The reproducibility record, URL list, retrieval date, paraphrased evidence, and hashes are in pop_baseline_provenance.json. The primary sources are the first-party 2018 simulation method, 2022 simulation, 2026 simulation, and method page, retrieved 2026-08-27.
The 2022 post reports a later 88-day step. That later setting is intentionally not silently substituted: this baseline is the documented three-window method that the project set out to improve, and all approximations are explicit.
Usage
uv run python -m scripts.pop_baseline --origin 2022-08-14 --horizon 28 --samples 5000 --seed 12345
The module returns deterministic seeded draws and machine-readable diagnostics. It requires an exact date in the processed PoP series when --origin-pop is not supplied; nearby dates are never substituted silently.
Matched benchmark
The separate scripts.pop_baseline.benchmark harness writes benchmark_report.json and benchmark_case_metrics.csv under data/processed/pop_baseline_benchmark/. It uses the same exact stored origin, target date, horizon, sample count, seed, and outcome for the baseline and frozen RC1 dynamics. The rolling component skips missing exact rows and origins with fewer than the Candidate A minimum transition pool; it never imputes a nearby date. Election 2018/2022 results are retrospective comparative evidence, not independent holdout validation. Seat metrics are explicitly unavailable because PoPBaseline does not fabricate geographic or seat draws.
The current checked-in run uses 256 draws per case and 3,634 scored cases out of 3,750; the remainder are explicit early-history/end-of-series skips. This is benchmark evidence for the harness and should not be presented as a universal winner claim. The report’s proper-score tables and implementation/input hashes are the source of any later model decision.