apol/ai-election-manipulation-cases

AI, Elections and Agency Transfer Evidence Index

Version 0.4.0 · released 15 August 2026 · research cutoff 12 August 2026

Viewer hotfix on main, 15 August 2026: Data Studio now exposes one orientation view. The frozen v0.4.0 tag, its 26 tables, ZIP, workbook and checksums are unchanged.

This is a purposive, exploratory evidence index. It makes claims, sources, counterevidence, uncertainty, mechanisms, and selected measurements easier to inspect. It does not empirically prove the agency-transfer hypothesis, estimate how common election manipulation is, or identify votes or election outcomes caused by AI.

The repository slug is inherited. The canonical title is the one above: the index also contains preparedness files, boundary comparisons, screening leads, official electoral context, and empirical model studies. Inclusion is not a finding that every record is manipulative or harmful.

Start here

• Open the 10-row research view — fastest orientation.
• Read the Spanish quick guide — concise interpretation guide.
• Read the methodology — units, coding and limitations.
• Read the v0.4.0 red-team report — findings, fixes and residual limits.
• Inspect the claim–evidence table — trace claims to source records.
• Download the complete v0.4.0 package — CSV, Parquet, documentation, ledgers, scripts and workbook.
• Download the v0.4.0 Excel workbook — navigation and QA convenience artifact.
• Verify release checksums — hashes for the frozen ZIP and workbook.
• Browse the Agency Transfer collection — related research datasets.
• See citation metadata — cite this index and the original sources.

The coded tables and field values are in English. Spanish is limited to the quick guide and selected descriptive metadata; this is not a bilingual corpus.

What is in v0.4.0

These counts describe the generated v0.4.0 build and must be refreshed if that build changes.

Layer	Rows	What the count means
Claim-coded records	10	Mixed units: 2 election-wide files, 2 networks, 3 incidents, 1 campaign set and 2 preparedness files
Incident-count eligible	8	Excludes preparedness; it is not a prevalence denominator
Atomic claims	100	Occurrence, attribution, mechanism, reach, effects, response and inference remain separate
Claim–evidence relations	124	supports, premise or counterevidence links, not 124 independent findings
Sources linked to claims	56	Distinct source records used by claim_evidence
Source records overall	70	Also includes records used elsewhere in the corpus
Screening leads	47	Research queue; not included incidents
Empirical model studies	7	Laboratory/audit evidence; not real-world deployments
Pathway assessments	10	Researcher-coded analytic syntheses, not measured outcomes
Typed observations	22	Selected quantities with unit, denominator, window, method and causal status
Official data sources	31	Separate electoral and administrative provenance layer

The strongest pathway conclusion is deliberately limited: 4 records show an observed control mechanism, 2 show an observed attempted transfer, 2 remain insufficient, and 2 are preparedness records rather than incidents. No record codes observed agency change or observed agency preservation/extension. These categories describe evidence ceilings, not severity scores.

The 26 tables

Use	Table names
Orient and scope	research_view, case_catalog, coverage_summary, sampling_frame
Claims and evidence	cases, claims, sources, claim_evidence, analytic_record_claims
Mechanisms and provenance	pathways, observations, events, case_sources, case_actors, technology_uses, content_items
Election relations and context	case_elections, election_sources, official_elections, official_election_metrics, official_turnout, official_results, official_data_sources
Discovery	candidates, watchlist
Empirical capability evidence	model_evaluations

All 26 canonical tables are available as CSV in data/ and as native Parquet in parquet/. Relational types, keys and foreign keys are declared in data/datapackage.json.

Only research_view is declared as an interactive Data Studio configuration. This keeps the public preview small and avoids conversion-lock contention among 26 unrelated schemas. It does not remove or merge any table; load the other 25 directly from their CSV or Parquet files.

Files under research/ are transparent build inputs and historical working ledgers, not canonical v0.4.0 tables. Some retain superseded draft annotations; see research/README.md before using them.

Load the data

Hugging Face Datasets

from datasets import load_dataset

view = load_dataset(
    "apol/ai-election-manipulation-cases",
    "research_view",
    revision="v0.4.0",
    split="data",
)

To load any of the 26 tables directly, replace claim_evidence with a table name from the map above:

from datasets import load_dataset

table = "claim_evidence"
url = f"https://huggingface.co/datasets/apol/ai-election-manipulation-cases/resolve/v0.4.0/parquet/{table}.parquet"
records = load_dataset("parquet", data_files={"data": url}, split="data")

pandas

import pandas as pd

url = "https://huggingface.co/datasets/apol/ai-election-manipulation-cases/resolve/v0.4.0/parquet/research_view.parquet"
view = pd.read_parquet(url)

Polars

import polars as pl

url = "https://huggingface.co/datasets/apol/ai-election-manipulation-cases/resolve/v0.4.0/parquet/research_view.parquet"
view = pl.read_parquet(url)

Remove the pinned revision only if you deliberately want the mutable latest version.

Read the evidence correctly

Each pathway keeps these fields separate: capability, controller, influence vector, target, affected agency, hypothesized control shift, possible power recipient, potential harm, and observed outcome. Their presence in one row does not imply that one caused the next.

pathways makes that reasoning inspectable. It does not establish that every link occurred, and it does not produce a validated agency-transfer score. In particular:

• strong_inference is a researcher inference, not a directly observed effect;
• platform reach is not belief, behaviour, turnout or vote choice;
• an institutional response is not proof of behavioural or electoral harm;
• official election results are context, not causal evidence;
• transparent, authorised, defensive and preparedness uses are not automatically manipulation.

Legacy numeric confidence fields were removed because they were neither calibrated probabilities nor intercoder-reliability estimates. This release has one primary coder and reports no intercoder-reliability statistic.

Evidence provenance

Use normalized joins for factual work:

claims.case_id                 → cases.case_id
claim_evidence.claim_id        → claims.claim_id
claim_evidence.source_id       → sources.source_id
observations.claim_id          → claims.claim_id
pathways.case_id               → cases.case_id
case_elections.election_id     → official_elections.election_id
election_sources.official_source_id → official_data_sources.official_source_id

The index retains public URLs but does not redistribute source snapshots. Of 124 claim–source relations, 57 have a verified pinpoint covering the full claim, 47 have a verified pinpoint covering only a factual premise or part of the claim, and 20 remain not_verified because the cited object was inaccessible during the locator audit. locator_coverage preserves that distinction and locator_note explains every unresolved pinpoint; a nonblank pinpoint is not a frozen source object, and a partial locator must not be treated as support for the whole claim. All archive and content-hash fields remain unresolved preservation gaps. Cite and inspect the underlying source before relying on a substantive claim.

What the audit does—and does not do

The automated audit is internal structural and semantic QA. It checks declared files, identifiers, relations, enums, date formats, selected consistency rules and recorded hashes within its stated scope. It does not establish that a source is still live, that a source supports every word of a claim, that an allegation is true, that the sample is representative, or that a causal interpretation is valid.

The v0.4 auditor retains the historical filename scripts/audit-v03.mjs and compatibility output data/audit-v03.json; its internal auditor identifier is audit-v04, and data/audit.json is byte-identical. The old entry point scripts/audit-data.mjs is a read-only wrapper.

Six visible warnings record regional frames without systematic negative/null searches. They are a reason not to infer prevalence, not evidence that no cases exist in those regions.

Appropriate use

Use this index for evidence tracing, comparative case research, hypothesis generation, institutional-response analysis, source discovery, measurement reconciliation and bounded model-study comparison.

Do not use it as:

• proof of the agency-transfer thesis;
• a global or regional incident count or prevalence estimate;
• a causal estimate of votes, turnout, seats or winners changed;
• a list of proven malicious AI incidents;
• evidence that visible or authorised political AI is inherently manipulative;
• operational material for voter targeting, impersonation, deception or safeguard evasion.

Builds, releases and citation

The supported tabular pipeline is:

npm ci
npm run build:data
npm run build:audit
npm run build:parquet
npm run validate
npm run test:hardening

After the externally generated workbook is frozen in release/, maintainers run:

npm run build:package
npm run verify:package:checkout
npm run build:release-manifest
npm run validate

npm run verify:package verifies the frozen ZIP internally against its embedded manifest and checksums. Use npm run verify:package:checkout only while cutting a release to additionally require byte-for-byte parity with the current checkout; later documentation-only changes on mutable main do not rewrite an existing release.

Do not run build:package to overwrite an already tagged version from mutable main; bump the version and cut a new tag first.

npm run audit:links is a separate, time-dependent network check; it is not evidence that a blocked or rate-limited page is absent. Within the declared runtime and inputs, the deterministic scripts rebuild and check the tabular and packaged artifacts they cover. That remains narrower than end-to-end source reproducibility: external pages are not frozen, 47 pinpoints cover only a factual premise or part of their claim, 20 relations have no verified pinpoint, and automated QA does not validate source truth.

The Excel workbook is a convenience artifact, not the canonical analytic source. Its builder requires the @oai/artifact-tool runtime, which is not part of the public package.json; use CSV or Parquet unless that environment is available. The ZIP and workbook should be verified against the release manifest, not assumed equivalent merely because their filenames contain the same version.

main is mutable. Use the v0.4.0 revision for a fixed citation and record the exact commit hash used. No DOI has been assigned; none should be inferred from the repository URL. A DOI can be added only after a frozen revision is deposited and its identity is reconciled with the tag, citation metadata and release manifest.

The CC BY 4.0 license covers the original coding and synthesis. Source materials remain under their publishers' or issuing institutions' terms. Cite both this index and the original evidence sources for substantive claims.