Swedish Riksdag Election Simulator v1.0-rc1

1. Overview & Architecture

ElectionSimulator v1 connects the frozen components of the Swedish Riksdag forecasting pipeline into a unified, reproducible, mathematically exact production simulator:

OpinionState v1.1 (Leakage-safe ALR estimation + House Effects + Kish weighting)
       │
       ▼
Dynamics v2 (symmetric_all_history exact transitions in CLR space, NO sqrt(h) scaling)
       │
       ▼
ElectionNoise (pp_centered_noise historical poll-to-election residual transfer)
       │
       ▼
National Vote Composition (Bounded simplex-safe 9-party vector)
       │
       ▼
GeographicProjection v1 (Deterministic IPF raking on 2022 baseline to 2026 constituencies)
       │
       ▼
Deterministic Integerization (Exact-margin bipartite residual flow rounding)
       │
       ▼
MandateAllocator v1 (Vallagen 14 kap. 1–5 §§ with iterative return convergence and keyed lottery)
       │
       ▼
Seats, Threshold Probabilities, Discrete Histograms, and Majority Analysis

Target Election: 2026-09-13.

2. Statutory Mandate Allocation & Convergence Invariants

The Riksdag mandate allocator strictly implements Vallagen (2005:837) 14 kap. 1–5 §§ and Regeringsformen 3 kap. 7 §:

349 Total Seats: 310 fixed constituency seats + 39 adjustment seats.
First Divisor: Exact Fraction(6, 5) ($1.2$).
Thresholds via Integer Cross-Products:
- National 4.0%: $25 \times \text{party_votes} \ge \text{total_valid_votes}$
- Constituency 12.0%: $25 \times \text{party_votes} \ge 3 \times \text{constituency_valid_votes}$
Keyed Deterministic Lottery: Ties resolved via SHA-256 hash over canonical legal allocation state (seed, phase, scenario_id, constituency, comparison_quotient, sorted_tied_parties, seat_state).
Iterative Return Convergence (14 kap. 4a–4c §§): Let $Q$ be the set of nationally qualified parties ($\ge 4\%$) and $L = \sum_{p \notin Q} F_p$ be fixed seats held by sub-4% parties qualifying locally. At convergence before adjustment seats: $\sum_p F_p = 310, \quad \sum_{p \in Q} E_p = 349 - L, \quad F_p \le E_p \;\; \forall p \in Q$ $U_p = E_p - F_p, \quad L + \sum_{p \in Q} (F_p + U_p) = 349, \quad \sum_{p \in Q} U_p = 39$

3. Geography Baseline Sensitivity Audit (2018 vs 2022)

Across 1,016 scenarios (central forecast, threshold sensitivity sweeps for L, KD, MP, and 1,000 Monte Carlo draws):

Deterministic Agreement Rate: 100.0% (0 / 1,016 differed) for the reported national seat vectors.
Cell-level outputs: constituency $\times$ party allocation differences are reported separately; national agreement does not certify every cell-level tie-break outcome.
Separation of Concerns: Statutory stress scenarios are tested and reported separately from empirical forecast probabilities.

4. 2026 Constituency Vote Volumes & Exact Integerization Scale

National Pseudo-Votes Total: Chosen as $6,500,000$ for exact statutory threshold representation: $0.04 \times 6,500,000 = 260,000 \text{ votes exactly.}$
Multiples of 25: Constituency target vote totals $R_c^{\text{int}}$ are apportioned as exact multiples of 25 ($12\% = 3/25$): $0.12 \times R_c^{\text{int}} = 3 \times k_c \text{ votes exactly.}$
Biproportional Controlled Rounding: Solves binary residual bipartite transportation $Y = \lfloor X \rfloor + Z$ preserving both constituency row totals $R_c^{\text{int}}$ and national party column totals $C_p^{\text{int}}$ with $|Y_{c,p} - X_{c,p}| < 1.0$.
Integerization Sensitivity: Comparing 6.5M pseudo-votes against high-precision 650M scale integer allocation across 5,000 samples yielded 0 / 5,000 seat differences (100.0% agreement).

5. REST Category Modeling & Ineligibility

REST represents the collective sum of all non-parliamentary minor parties.
Statutory Ineligibility: Labeled internally as OTHER_INELIGIBLE. REST contributes to valid-vote denominators but is legally ineligible for parliamentary seats and strictly receives 0 seats in all simulations.
Limitation Notice: The simulator assumes no individual minor party within REST independently qualifies for the Riksdag. Any minor party with realistic qualification potential must be broken out and modeled as an explicit 9th or 10th party entity.

6. Reproducibility Manifest

Every simulation emits metadata capturing all cryptographic dependencies:

model_version: 1.0.0-rc1
as_of & election_date
samples & base_seed
SHA-256 hashes of swedishpolls_individual_polls.csv, riksdag_election_results.csv, historical_certified_mandates.csv, and constituency_party_votes_2014_2022.csv
model_config_hash, source_git_commit, and source-worktree cleanliness at generation time
Deterministic payload SHA-256 excluding timestamps and runtime fields
UTC timestamp.

Given the same seed and input data, outputs are bit-for-bit identical across executions.

source_git_commit identifies the clean code/data commit used for generation. The artifact may be committed afterward and therefore must not claim that its own containing commit generated it.

The release certification, immutable prospective archive, and external-model benchmark contract are documented in election_simulator_rc1.md.

7. Performance & Benchmark Profile

Production benchmark: the freeze audit records an isolated cold-subprocess run separately from an in-process warm run. Use the generated benchmark artifact for measured values; performance is environment-dependent.
Profiler / Tracing Benchmark:
- Unchunked tracing and full Python instrumentation measures 100k in ~1,025 seconds.
- The historical discrepancy (975s vs 91s) was due to tracing memory allocation and unvectorized loop instrumentation during early development vs the optimized vectorized production pipeline.

8. CLI Commands

# Run standard 100k production simulation
make simulate

# Run performance and memory benchmarks
make simulate-benchmark

# Run baseline and integerization sensitivity audits
make simulate-audit

# Run simulator test suite
make test-simulator

9. Evidence and Evaluation Labels

SeatHindcast results are labelled “Retrospective historical evaluation (not independent holdout validation)”. Coverage and horizon patterns are descriptive, and the baseline and simulator joint Energy Scores are computed per stored election-by-horizon case before aggregation.

The Valmyndigheten Example 5 fixture archived under tests/fixtures/ is an official three-constituency municipal worked example. It is not executable by the 29-constituency, 349-seat Riksdag allocator without changing its semantics; the freeze report labels this external-fixture limitation and keeps synthetic Riksdag stress tests separately named.

edvin listo zec