docs/start-here.md

Start here

This repository is useful only if its three kinds of evidence stay separate: documented election records, screening leads, and model evaluations. It is an evidence index, not a list of proven AI-caused election harms.

Count hierarchy

The dataset contains 6 documented-manipulation records, not 1,087 cases. The 6 sit inside 8 incident-eligible records; the other 2 are transparent contested-use boundary cases. Adding 2 preparedness files produces the 10-record claim-coded core. Adding 47 screening leads and 7 model studies produces the 64-entry catalogue. The 1,087-row index then sums rows across 26 related tables, including claims, sources, joins, metrics and derived views. These are relational rows, not unique cases.

Choose your route

You are	Begin with	Then inspect	Safe output
Researcher	research_view	claims, claim_evidence, sources, pathways, observations	A bounded case comparison, evidence audit, measurement design, or preregistered hypothesis
Policy analyst	case_catalog	research_view, events, technology_uses, case_actors, official_data_sources	A threat-and-response brief that states evidence ceilings and gaps
Policymaker	Policy brief	Policy-analysis guide	Proportionate actions, evidence requests, and monitoring priorities
Data user	record_index	Codebook, methodology, datapackage.json	Reproducible joins and descriptive counts

What the evidence currently supports

The claim-coded core contains ten records. Eight are incident-count eligible: six are coded manipulation_confirmed or mixed_documented_manipulation, while two are transparent contested-use boundary cases. Two additional records are preparedness files. Across the ten pathways, four reach control_mechanism_observed, two reach attempted_transfer_observed, two have insufficient_evidence, and two are not_an_incident. No record reaches agency_change_observed or agency_preserved_or_extended.

This supports a narrow conclusion: documented systems have already been used to control provenance, apparent identity, salience, distribution, or participation cues in election settings. The index does not show how often this happens, how many unique voters were persuaded, whether effects lasted, or whether any result changed.

Five-minute workflow

1. Identify the unit in case_catalog. Do not pool an election-wide file, a network, an incident, and a model experiment as equivalent observations.
2. Read the case's maximum_conclusion and main_uncertainty in research_view.
3. Open its atomic claims in claims.
4. Follow every material claim through claim_evidence to sources. Treat partial_claim and not_verified locators as unresolved review work.
5. Use observations only with its unit, denominator, window, uniqueness and causal-status fields.
6. Read pathways as a falsifiable researcher assessment, not as a measured outcome.
7. Cite the original source as well as the pinned dataset revision.

Stop conditions

Do not use this dataset to claim:

• global or regional prevalence;
• a number of voters reached, persuaded, deterred or converted;
• votes, turnout, seats or winners caused by AI;
• state attribution where the record preserves uncertainty;
• that all synthetic or automated political content is manipulative;
• that an institutional response proves the alleged mechanism or its effects.

Reproduce the core counts

Run the dependency-free example from the repository root:

python examples/reproduce_core_findings.py

It reads the canonical CSV files, checks core joins and prints descriptive counts. The script deliberately computes no severity, risk, manipulation, or agency-transfer score.