Upload audited dataset v0.1.0 metadata

.gitignore

@@ -0,0 +1,5 @@
+output/
+*.xlsx
+*.png
+*.inspect.ndjson
+.DS_Store

CHANGELOG.md

@@ -0,0 +1,13 @@
+# Changelog
+
+## 0.1.0 — 2026-08-11
+
+- Initial claim-level schema.
+- Deep anchor files for Romania, Moldova, New Hampshire, Brazil 2026, and U.S. 2026.
+- Prospective watchlists for the two October/November 2026 elections.
+- Explicit separation of occurrence, reach, attribution, behavioural effect, electoral effect, and institutional outcome.
+- Claim-level anchor enrichment from the Romania and Moldova source ledgers, including separate domestic-financing, platform-network, synthetic-persona, and InfoLeaders evidence.
+- Global comparative queue of 36 source-linked candidates across the Americas, Europe, Asia, and Africa, with boundary and second-review flags.
+- Reproducible relational audit with identifier, reference, enum, date, evidence, and source-minimum checks.
+- Source/current-output integrity hashes and a release audit with zero errors and seven explicit live-update warnings.
+- Pre-release corrections to the Moldova payment denominator, candidate-date ambiguity, timeline-to-claim joins, and source-role vocabularies.

CITATION.cff

@@ -0,0 +1,15 @@
+cff-version: 1.2.0
+message: "If you use this dataset, please cite it and the original sources for substantive claims."
+title: "AI-Enabled Election Manipulation Cases"
+type: dataset
+authors:
+  - name: "Apol"
+version: 0.1.0
+date-released: 2026-08-11
+license: CC-BY-4.0
+keywords:
+  - elections
+  - artificial intelligence
+  - political manipulation
+  - deepfakes
+  - agency transfer

CONTRIBUTING.md

@@ -0,0 +1,26 @@
+# Contributing corrections and new cases
+
+This dataset accepts evidence corrections, source replacements, and proposed cases. A submission should identify the stable record ID, state one atomic correction or addition, link the strongest available source, and explain whether it changes occurrence, reach, attribution, behavioural effect, electoral effect, or institutional outcome.
+
+## Minimum evidence
+
+- Prefer an authoritative primary record plus independent corroboration.
+- Quote only the smallest passage necessary to locate the evidence; do not redistribute copyrighted articles or harmful media.
+- Separate creator, sponsor, distributor, platform, beneficiary, attributed state actor, and responding institution.
+- Report the denominator for every scale figure. Accounts, calls, completed calls, views, followers, impressions, recipients, and unique voters are not interchangeable.
+- Do not infer belief, turnout, votes, seats, or a result from reach alone.
+- Include counterevidence, denials, acquittals, reversed orders, and unresolved attribution.
+
+## Promotion from `candidates`
+
+A candidate can enter the included corpus only after it has a stable `case_id`, claim-level coding, source-role links, an event timeline, an agency-transfer path, and an explicit statement of the maximum conclusion warranted by the evidence. Boundary cases may remain in `candidates` when the main value is comparison rather than demonstrated harm.
+
+## Local validation
+
+```bash
+node scripts/build-data.mjs
+node scripts/validate-data.mjs
+node scripts/audit-data.mjs
+```
+
+A release must finish with zero audit errors. Warnings require written disposition in the changelog or source ledger.

LICENSE

@@ -0,0 +1,6 @@
+Creative Commons Attribution 4.0 International
+
+This work is licensed under the Creative Commons Attribution 4.0 International License.
+To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/
+
+The license covers the original selection, coding, and metadata in this repository. Third-party sources remain subject to their own rights and terms.

README.md

@@ -1,3 +1,188 @@
 ---
 license: cc-by-4.0
+language:
+  - en
+pretty_name: AI-Enabled Election Manipulation Cases
+size_categories:
+  - n<1K
+tags:
+  - elections
+  - political-disinformation
+  - generative-ai
+  - deepfakes
+  - agency-transfer
+  - dataset-card
+configs:
+  - config_name: cases
+    data_files:
+      - split: data
+        path: data/cases.csv
+  - config_name: claims
+    data_files:
+      - split: data
+        path: data/claims.csv
+  - config_name: sources
+    data_files:
+      - split: data
+        path: data/sources.csv
+  - config_name: events
+    data_files:
+      - split: data
+        path: data/events.csv
+  - config_name: case_sources
+    data_files:
+      - split: data
+        path: data/case_sources.csv
+  - config_name: watchlist
+    data_files:
+      - split: data
+        path: data/watchlist.csv
+  - config_name: candidates
+    data_files:
+      - split: data
+        path: data/candidates.csv
 ---
+
+# AI-Enabled Election Manipulation Cases
+
+An auditable, claim-level dataset of election-related manipulation in which AI, algorithmic systems, synthetic media, coordinated networks, or automated distribution materially shaped the incident or the response.
+
+**Research cut-off:** 11 August 2026. The rows are English-language research metadata; source material spans several languages. This is an evidence index, not a benchmark or a labelled text-classification corpus.
+
+Version `0.1.0` centers five case families:
+
+1. Romania's annulled 2024 presidential election and 2025 repeat election.
+2. Moldova's 2024 presidential election and EU referendum, with linked evidence from the 2025 parliamentary campaign.
+3. The January 2024 New Hampshire AI-cloned Biden robocall.
+4. Brazil's October 2026 general election, currently an ongoing case and monitoring file.
+5. The United States' November 2026 midterm election, currently an ongoing case and monitoring file.
+
+The dataset intentionally does **not** treat every case as a “deepfake case.” It separately records synthetic media, recommender-system amplification, coordinated inauthentic behaviour, bots, illicit finance, telecom delivery, and institutional response.
+
+## Release structure
+
+The release has two deliberately different evidentiary layers:
+
+- **Included corpus:** incidents and ongoing files that have been decomposed into atomic claims and joined to a source register and event timeline.
+- **Comparative candidates:** a global research queue. These rows are useful leads and boundary cases, but they are not counted as fully included cases until claim-level coding is complete.
+
+This distinction prevents a long list of reported “AI election incidents” from masquerading as a validated event dataset.
+
+### Coverage in `0.1.0`
+
+| Layer | Rows | What the count means |
+|---|---:|---|
+| Included cases | 9 | Five case families: Romania, Moldova, New Hampshire, Brazil 2026, and U.S. 2026 |
+| Atomic claims | 85 | Separately coded occurrence, mechanism, reach, attribution, effect, response, and counterevidence statements |
+| Sources | 57 | Primary records, independent reporting, technical investigations, and explicit counterevidence |
+| Timeline events | 31 | Observed, contested, and prospective events kept distinct |
+| Case–source links | 63 | Source role and supported claim types |
+| Defensive watch indicators | 12 | Brazil and U.S. 2026 monitoring rules; not incident predictions |
+| Comparative candidates | 36 | Twelve Americas, twelve Europe, and twelve Asia/Africa leads, including boundary and reserve cases |
+
+The release audit reports **0 errors** and **7 expected live-review warnings**. All seven warnings identify time-sensitive Brazil or U.S. 2026 claims that require updating as proceedings and campaigns evolve; they are not silently treated as final facts.
+
+## What is distinctive
+
+- The unit of evidence is a **claim**, not a news article.
+- Occurrence, reach, attribution, behavioural effect, electoral effect, and institutional outcome are separate claim types.
+- Every non-open claim has an evidence label, confidence score, and explicit source links; unsourced open questions are explicitly labelled.
+- Prospective and ongoing elections are never encoded as completed harms.
+- The `agency_transfer_path` records who controls the system, what capacity is transferred away from citizens or institutions, and where power concentrates.
+- Counterevidence and unresolved questions are first-class fields.
+
+## Files
+
+| Config | Unit | Purpose |
+|---|---|---|
+| `cases` | Case or case family | Core comparative metadata and assessed mechanism |
+| `claims` | Atomic claim | Evidence status, confidence, support, and counterevidence |
+| `sources` | Source | Provenance and quality metadata |
+| `events` | Dated event | Auditable case timeline |
+| `case_sources` | Case–source link | Many-to-many source roles |
+| `watchlist` | Observable indicator | Prospective monitoring for Brazil and U.S. 2026 |
+| `candidates` | Comparative lead | Global comparison queue kept separate until full claim-level coding |
+
+`data/stats.json` reports current row counts and `data/audit.json` records the automated quality-assurance result. `data/schema.json` contains the generated column order for every config.
+
+## Quick start (after Hub publication)
+
+```python
+from datasets import load_dataset
+
+cases = load_dataset(
+    "apol/ai-election-manipulation-cases",
+    "cases",
+    split="data",
+)
+
+claims = load_dataset(
+    "apol/ai-election-manipulation-cases",
+    "claims",
+    split="data",
+)
+```
+
+The configs are relational rather than independent training sets. Join `claims`, `events`, `case_sources`, and `watchlist` to `cases` with `case_id`; join `case_sources` and the pipe-delimited `claims.source_ids` to `sources` with `source_id`. Stable IDs are never reused for a different incident.
+
+## Evidence labels
+
+- `established_evidence`: supported by a primary record or multiple high-quality independent sources.
+- `strong_inference`: the inference is well supported but not directly observed or adjudicated.
+- `plausible_hypothesis`: coherent and evidence-informed, but important links remain unverified.
+- `speculative_scenario`: a forward-looking possibility; never counted as an observed event.
+- `open_question`: the dataset records the uncertainty rather than resolving it.
+
+Confidence is reported on a `0–1` scale and refers to the claim **as worded**, not to the case as a whole.
+
+## Reading a claim correctly
+
+The dataset distinguishes six questions that are often collapsed in public discussion:
+
+1. **Occurrence:** did the content, network, payment, call, rule, or institutional act exist?
+2. **Reach:** what denominator was measured—attempts, completed calls, views, followers, accounts, or recipients?
+3. **Attribution:** who created, financed, directed, distributed, or merely benefited from it, and at what evidentiary level?
+4. **Behavioural effect:** did a person believe, prefer, abstain, turn out, or vote differently?
+5. **Electoral effect:** is there evidence about votes, seats, margins, or the result?
+6. **Institutional outcome:** what did a court, regulator, election authority, platform, or law-enforcement body do?
+
+Evidence for one question is not silently promoted into another. In particular, views and virality are not treated as proof of persuasion, and an institutional decision is not automatically a causal estimate of changed votes.
+
+## Central limitation
+
+Evidence of electoral effect is scarce. Reach, virality, exposure, or institutional alarm must not be translated into votes changed. Most cases therefore have strong evidence about occurrence and response, but `unknown` or low-confidence evidence for behavioural and electoral effects.
+
+The source register also mixes different kinds of authority. A court order is authoritative evidence that an order was issued; a regulator's allegation is authoritative evidence of that allegation; a platform report is evidence of the platform's measurement or attribution. None is treated as omniscient evidence of every underlying fact.
+
+## Intended and unsuitable uses
+
+Suitable uses include comparative case research, source discovery, institutional-response analysis, hypothesis generation, and testing how different evidence thresholds change a case narrative.
+
+Do not use the release as:
+
+- a prevalence estimate for all elections or all online political content;
+- a causal estimate of votes changed;
+- a list of proven foreign operations without consulting `actor_attribution_status` and the atomic claims;
+- training data for reproducing deceptive media, targeting voters, or evading detection;
+- a static forecast of the unfinished 2026 elections.
+
+## Ongoing-election protocol
+
+Brazil and the United States in 2026 are versioned as-of files. Observed events, disputed allegations, legal rules, future deadlines, and defensive monitoring indicators remain separate. A scheduled election or plausible risk never becomes an incident row by implication. Later releases may revise ongoing records while preserving the changelog.
+
+## Methodology and ethics
+
+See [`docs/methodology.md`](docs/methodology.md), [`docs/codebook.md`](docs/codebook.md), and [`docs/ethics-and-release.md`](docs/ethics-and-release.md).
+Corrections and promotion criteria are documented in [`CONTRIBUTING.md`](CONTRIBUTING.md).
+
+To reproduce the generated files locally:
+
+```bash
+node scripts/build-data.mjs
+node scripts/validate-data.mjs
+node scripts/audit-data.mjs
+```
+
+## Citation
+
+See [`CITATION.cff`](CITATION.cff). Cite original sources for substantive factual claims and this dataset for the coding and synthesis.