docs/methodology.md

Methodology

Research question

When an election-related AI or algorithmic system is documented in relation to political representation, attention, information, participation, or decision support, what capability is observed; what control mechanism, if any, is documented; is an attempt to displace a decision-relevant capacity observed; is any agency change measured; and what remains a falsifiable hypothesis?

The dataset does not assume that AI use is manipulation, that control of a communication channel changes a person's agency, or that a documented influence attempt succeeds. It is an evidence index for testing control-shift hypotheses—not evidence that “agency transfer” occurred. Manipulation is assessed separately from agency outcomes using process evidence such as deception, unauthorised representation, opacity, coercion, exploitation, covert coordination, loss of contestability, or control over a decision environment.

Research layers and units

The release has three non-interchangeable layers.

1. Claim-coded core. A bounded record with atomic claims, sources, actors, technology, pathway assessment, and case–election links. record_type identifies whether the unit is an election-wide case, observed network, observed incident, campaign set, or preparedness file.
2. Screening register. Comparative leads, boundary cases, null samples, reserves, and records needing disaggregation or corroboration. These are not included incidents.
3. Empirical model studies. Model audits and experiments. They measure capability or experimental outcomes under a specified design, not real-world incident prevalence.

Preparedness files and empirical studies have incident_count_eligible=false. The remaining core records still require stratification by record_type; they are not automatically exchangeable observations.

Inclusion criteria

A core record must:

1. be materially connected to an election, referendum, voter participation, candidate choice, or electoral trust;
2. involve a documented AI, synthetic-media, recommender, automated-distribution, coordinated-network, or AI-enabled response mechanism;
3. have an authoritative primary record plus appropriate corroboration, or multiple credible independent sources;
4. permit occurrence, actor, mechanism, reach, effect, response, counterevidence, and uncertainty to be coded separately;
5. identify the unit and election relationship without aggregating unrelated artifacts into one implied operation;
6. retain the maximum conclusion warranted by the evidence, not the most dramatic interpretation;
7. specify what observation would count against a control-shift hypothesis and what comparator would be needed to test agency change or preservation.

Transparent or authorised AI may be retained as a comparison when it tests the boundary between representation, persuasion, manipulation, and agency preservation.

Screening and coverage

Selection is purposive. Anchor cases were chosen for legal, institutional, technical, or comparative value. Candidate leads were then added to broaden geography, mechanism, authorization status, model access, channel, and positive/null comparison.

sampling_frame.csv records:

• region and selection basis;
• search status and languages;
• source channels;
• included and candidate counts;
• whether a negative/null search was recorded;
• known gaps and last search date.

The release may be described as internally complete for its declared records and tables. It must not be described as a comprehensive election catalogue or globally representative. A regional prevalence statement requires a defined universe, systematic search, recorded intensity, positive and negative decisions, and comparable denominators.

Evidence workflow

1. Inventory official, legal, regulatory, platform, technical, peer-reviewed, civil-society, and independent reporting sources.
2. Atomize source statements into claims.
3. Distinguish the fact itself from the fact that an actor or institution made an assertion.
4. Code creator, commissioner, sponsor, distributor, infrastructure provider, beneficiary, target, and responder separately.
5. Add counterevidence and competing explanations.
6. Code actors, technology uses, content items, pathways, and quantities from supported claims.
7. Attempt to recheck URLs, publication dates, access status, and known proceeding status at the cutoff; record blocked, dynamic, or unverified items explicitly.
8. Run deterministic relational, semantic, causal-safeguard, and official-data consistency checks.

Coding and review protocol

Version 0.4.0 was produced by one primary coder with automated consistency checks and targeted record-level review. It does not report intercoder reliability because no independent second-coder sample was completed. Numeric case and claim confidence scores were removed: they were uncalibrated self-assessments that created false precision and could not substitute for agreement or source review.

Before any inferential or publication-grade v1.0 release, the protocol requires an independent coder to review a stratified sample spanning record types, regions, claim types, source grades, and evidence labels. Disagreements must be reconciled in a dated adjudication log, and agreement must be reported separately for categorical labels and quantitative extraction. Until then, users should rely on claim wording, evidence relations, source directness, counterevidence, and the explicit maximum conclusion. No intercoder-reliability claim is made.

claim_evidence.csv is the normalized claim–source relation. relation=supports means the source states the coded factual or attributed proposition; premise means it supports a factual premise used in a researcher inference but does not state or prove that inference; counterevidence qualifies or contests it. evidence_scope distinguishes an underlying fact, official action, measurement, attributed assertion, review-scope statement, or researcher inference. An official allegation, party filing, platform statement, or defence supports assertion_was_made; it does not establish the alleged conduct. directness records whether the source is an artifact, measurement, official record/allegation, or secondary report.

Claim status

evidence_label preserves the reasoning form:

• established_evidence
• established_as_campaign_report
• strong_inference
• researcher_hypothesis
• not_identified_in_declared_review
• open_question

claim_status records whether wording is supported as written, supported only as an attributed measurement, partly supported, an untested hypothesis, a review-bounded non-identification, or unresolved. evidence_basis is a non-numeric category; it is not an ordinal score.

not_identified_in_declared_review means only that no supporting source was found among the sources linked to that record by the cutoff. It is not evidence of global absence and is intentionally not linked to selected sources as though those sources proved a negative.

Falsifiable control-shift pathway

The pathway table separates capability, controller, vector, target, agency dimension, hypothesized control shift, possible power recipient, potential harm, and observed outcome. These are fields to test independently, not a causal ladder, and no transition is assumed merely because the fields share a row.

pathways.csv records:

• the agency dimension examined;
• who is hypothesized to hold or gain relevant control;
• technical/social mechanism;
• opacity, personalization, contestability, and duration;
• capability and control-mechanism status;
• attempted-transfer, observed-agency-change, and agency-preservation status;
• the maximum conclusion allowed by the evidence;
• a falsification condition and comparator;
• observed evidence separated from researcher inference;
• supporting claim identifiers and reviewer note.

The allowed maximum_conclusion values are control_mechanism_observed, attempted_transfer_observed, agency_change_observed, agency_preserved_or_extended, insufficient_evidence, and not_an_incident. They are conclusions, not severity ranks.

An observed capability is not an observed control mechanism. A control mechanism is not an attempted transfer. An attempt is not an agency change. Disclosure, authorization, contestability, or successful user verification can support an agency-preservation hypothesis, but none is treated as proof without an outcome measure. not_observed means the record contains no direct observation of that outcome; it does not mean the outcome did not occur.

In v0.4.0, none of the eight incident-count-eligible core records has agency_change_status=observed or agency_preservation_status=observed. Two support only an observed attempt, four support only an observed control mechanism, and two transparent contested uses remain boundary cases with insufficient evidence. The two preparedness files are not_an_incident.

The dataset deliberately calculates no agency-transfer or control-shift score. Such weights would be normative and falsely precise. The renamed control_shift_hypothesis claims have no direct source links: their factual premises are linked separately through source_claim_ids, and their truth must be tested against the stated comparator.

Quantitative observations

Selected decision-relevant quantities are normalized as separate rows in observations.csv. Numeric source context can remain only in claims when it is not suitable for comparison. Each observation requires:

• metric family and name;
• low, best, high, and operator where available;
• unit and denominator;
• unique-entity status;
• platform, geography, and time window;
• measurement method and source;
• causal status and limitations.

Accounts, transactions, posts, views, followers, likes, calls, questions, users, people, voters, and votes are different units. A platform event is never relabelled as a unique person without explicit deduplication. Heterogeneous aggregates are not normalized as one observation: the reported Moldova aggregate of more than 138,000 accounts and transactions is retained only as a warning claim and excluded from observations. Descriptive quantities are not causal estimates.

Empirical model studies

model_evaluations.csv keeps model tests outside the incident corpus. A row states systems/versions, interface, languages, sample unit and size, design, metrics, result summary, data availability, and causal scope.

Randomized experiments can support causal claims only for the assigned interaction and measured outcome in the experimental population. They do not establish covert deployment, durable effects, actual ballots, or aggregate election outcomes. Point-in-time model audits cannot be generalized to later versions, all prompts, consumer interfaces, or all user populations.

Official electoral-data protocol

Official election data are descriptive context and outcomes. They are linked to cases but never treated as evidence that influence changed votes.

For each official source, the manifest records the institution, a landing URL, status, format, granularity, access method, retrieval time, snapshot cutoff, and reuse terms. A distinct download URL and optional content hash, bytes, or schema version are included only when captured and verified; blanks are not evidence of a missing public record.

This release does not archive or redistribute the referenced source corpus. Claim locators are populated only when a page, paragraph, timestamp, table, cell, or section was verified against the cited object. locator_coverage=full_claim means that pinpoint covers the full coded wording; partial_claim means it covers only part; not_verified leaves the locator blank. A verified locator is not a source snapshot or preservation hash.

At the v0.4.0 cutoff, 57/124 claim–source relations have a verified full-claim locator, 47/124 have a verified partial/premise locator, and 20/124 are not_verified because the cited object was inaccessible during the locator audit. Relations for which no supporting passage was identified were removed or reclassified as explicit factual premises; the mutable YouTube count was removed because no stable snapshot was preserved. Separately, 0/70 sources have an archive URL or content hash, and 0/31 official-data sources have verified content hashes, byte counts, or source-schema versions. The relational model has substantially better pinpointing, but the source corpus is not independently preserved and partial locators do not validate uncited parts of a claim.

For each measure:

1. retain the original field code and label when available;
2. separate numerator, denominator, authority-reported rate, and computed rate;
3. distinguish legal role from descriptive role;
4. preserve legal validation or annulment chronology separately from the tally;
5. never create future result rows;
6. never force a turnout rate from incompatible denominator concepts;
7. reconcile complete result groups to the official valid-vote total.

Moldova denominator rule

Moldovan protocols distinguish:

• A: electors on basic lists;
• B: electors on supplementary lists;
• C: electors who received ballots;
• D: electors who participated;
• E: C − D;
• F: invalid ballots;
• H: valid votes;
• I/J: additional ballot-accounting fields where available.

For the 2024 presidential rounds and referendum, the Constitutional Court used C/(A+B) for the legal validity threshold. The prior wide table used D/(A+B), which is a descriptive participation rate but not the legal rate. Both are now retained. For the 2025 parliamentary election, the authority-reported rate uses D/(A+B).

Result completeness

Romania 2024 first round, Romania 2025 first round, Moldova 2024 first round, Moldova 2024 second round, the Moldovan referendum, and Moldova 2025 parliament are complete at national contest level. New Hampshire remains partial; its source remains authoritative for the full field.

Causal safeguards

The release blocks these common errors:

• institutional alarm or annulment does not prove votes changed;
• a close result is not evidence of influence;
• views, engagement, or followers do not prove exposure or persuasion;
• cross-election change is not a counterfactual;
• aggregate geography does not identify individual behavior;
• a hybrid operation does not permit its total outcome to be attributed to AI alone;
• platform or state attribution is not the same as final adjudication;
• rapid removal can explain low reach and is not proof of low technical capability;
• a non-identification is bounded to the declared review and is not evidence of global absence;
• institutional action, including annulment, is not part of the manipulation mechanism unless a causal relationship is independently established;
• production or representational control does not by itself establish a change in a voter's, candidate's, or represented person's agency.

Quality assurance

The deterministic v0.4.0 audit is designed to check all generated tables for:

1. explicit record type and manipulation assessment;
2. exclusion of preparedness and empirical studies from incident counts;
3. normalized foreign keys and evidence joins;
4. controller, target, mechanism, separate falsifiable states, maximum conclusion, falsification condition, and comparator in every pathway;
5. actor and technology provenance, including explained unknown versions;
6. typed, bounded, sourced observations and causal status;
7. no platform proxy coded as a unique person or voter;
8. model-study sample/design/scope;
9. sampling-frame counts and negative-search flags;
10. election, source, legal, turnout, and future-date status;
11. complete-result shares and totals;
12. reproduction of research and coverage views;
13. absence of numeric coder-confidence fields and legacy agency-transfer labels;
14. bounded handling of absence claims, official allegations, and heterogeneous quantities.

Audit success establishes structural and semantic consistency only. It does not validate source truth, intercoder reliability, sampling representativeness, causal effects, or a control-shift hypothesis.