ElectionSimulator v1.0-rc1 certification and prospective evidence

ElectionSimulator v1.0-rc1 is the frozen Candidate A model. The statistical model, polling calibration, geographic projection, turnout assumptions, election residuals, strategic-voting treatment, and mandate rules are not changed by the comparative benchmark.

Two-commit provenance

Release certification uses two commits:

1. a clean source-code commit containing implementation, tests, fixtures, benchmark code, and documentation;
2. a separate artifact commit containing forecasts, audits, hindcasts, and prospective snapshots generated from the first commit.

Every generated artifact records source_git_commit and source_worktree_clean=true. The cleanliness flag covers source, tests, configuration, and input data; only the explicitly generated certification, hindcast, prospective-archive, and benchmark output directories are ignored for this flag. It is intentionally impossible for an artifact to contain the hash of the later commit that stores that artifact.

The historical report claiming L 3.56%, S 31.25%, and a 31.8% Tidö majority has no reproducible source artifact or command log. It is marked UNREPRODUCIBLE_INVALID and is not a canonical forecast.

Daily prospective archive

The archive is append-only by as_of date. A collision, duplicate deterministic payload, or duplicate identity fails closed; existing snapshots are never overwritten. Snapshots contain compact quantiles and histograms, not the 100,000 raw draws, and link to the canonical forecast through both its file SHA-256 and deterministic payload SHA-256.

Run one snapshot after producing the canonical forecast:

uv run python -m scripts.prospective_archive \
  --as-of 2026-08-23 \
  --election-date 2026-09-13 \
  --samples 100000 --seed 12345

The command writes data/processed/prospective_forecasts/<as_of>/snapshot.json and updates index.json. Do not rerun the same as_of date into the same archive; the command is designed to refuse that collision. Preserve the poll-data hash, source commit, model/config hashes, and timestamp from every run.

REST is an aggregate residual category for otherwise modeled-as-ineligible parties. It is never independently eligible for the 4% threshold or seats.

Botten Ada comparison

The benchmark harness treats this model as Candidate A and does not mutate it. Botten Ada is pinned to the public MansMeg/ada_code repository commit 2dfe246b86c5cab517e4a0cb87fd57e5a9c62512, with source and data URLs recorded in the benchmark output. The current repository does not contain an independently generated Botten Ada predictive-draw bundle, so the default benchmark status is NOT_RUN. No point forecast is converted into an invented distribution.

An external adapter bundle must use the standardized schema and fixed party order. It must provide identical election date, as-of cutoff, horizon, and sample count for both candidates. Retrospective 2018/2022 scores are comparative historical evidence, not holdout validation. The harness computes:

• vote CRPS by party and macro-average;
• joint vote Energy Score;
• party-specific inclusive 4% threshold Brier score;
• mean and median vote-share MAE;
• central predictive coverage and interval width;
• seat CRPS and joint seat Energy Score when independently generated seat draws and realized seats are available.

The pre-registered pivot rule is stored in the machine-readable report. A tie or Candidate A win keeps Candidate A unchanged. Only a consistent, material Candidate B advantage at late horizons and on threshold scores permits targeted layer investigation; it does not authorize automatic adoption.

Generate a Candidate A exchange bundle (raw draws are for benchmark exchange, not the compact prospective archive) with:

uv run python -m scripts.botten_ada_benchmark.candidate_a \
  --as-of 2026-08-23 --output /tmp/candidate_a.json

Then compare it with an independently produced Botten Ada bundle:

uv run python -m scripts.botten_ada_benchmark \
  --candidate-a /tmp/candidate_a.json \
  --candidate-b /path/to/botten_ada_bundle.json \
  --output data/processed/botten_ada_benchmark/benchmark_report.json