Reporting-stats — measuring asymmetry in Israel–Palestine conflict reporting

This project measures whether news coverage of the Israel–Palestine conflict is asymmetric relative to a ground-truth record of events, and if so, in which direction and on which dimension. It is deliberately built to be falsifiable: it can return "bias toward Israel", "bias toward Palestine", or "no measurable asymmetry" depending on what the data shows.

⚠️ Framing note. A study designed to prove bias is not credible. A study designed to measure asymmetry — with a pre-registered metric that could come out either way — is. Everything here is structured around the second goal. Read docs/METHODOLOGY.md before drawing conclusions.

Why a baseline alone is not enough

A list of rocket attacks on Israel, by itself, cannot demonstrate reporting bias. Bias is a relative measurement. If you only have events on one side you cannot separate "coverage is skewed" from "that side had more newsworthy events."

So the baseline (rockets → Israel) is built as one arm of a two-sided event dataset. The other arm (Israeli strikes/operations → Gaza & West Bank) is pulled the same way, from the same source, so the two are comparable.

The three things we measure

Dimension	Question	Data needed
Coverage volume	Articles / words per event and per fatality on each side	events + per-outlet article counts
Framing & language	Active vs. passive voice; who is named as the actor; word choice (killed vs. died)	article headlines/bodies
Selection / omission	Which real events get covered at all vs. ignored	events + coverage match

Ground truth: ACLED

Events come from ACLED (Armed Conflict Location & Event Data). ACLED codes each event from a deliberate mix of pro-Israeli and pro-Palestinian sources and documents uncertainty — the most defensible "truth" layer available, and it covers both directions of fire.

A sourced seed dataset (data/seed/rocket_episodes_seed.csv) is included so you can explore the rocket baseline immediately, before wiring up ACLED. It is coarse (episode-level, approximate) and is superseded by the ACLED pull.

Outlets analyzed

Configured in config.yaml:

• Western legacy: BBC, New York Times, Reuters, AP
• Regional anchors: Al Jazeera, Times of Israel, Haaretz
• Broad aggregate: all GDELT-indexed outlets (for volume/tone at scale)

⚠️ Network policy constraint

ACLED and GDELT are not reachable from the default Trusted Claude-Code-on-the-web network level (Host not in allowlist / HTTP 403). Set the environment's Network access to Custom and allowlist the data hosts — full instructions and the exact domain list are in docs/NETWORK.md. The framing judge needs no network change: the Anthropic SDK host is already allowlisted, and the in-session subagent path needs neither a key nor network. Running locally sidesteps the allowlist entirely.

Quick start

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env          # add your ACLED key + email

# 1. Pull both-sides events (needs network + ACLED creds)
python -m src.ingest.acled --start 2023-10-01 --end 2024-06-01

# 2. Build the rocket-attack baseline (success/failure outcomes)
python -m src.baseline.rocket_baseline

# 3. Pull per-outlet coverage volume (needs network; GDELT, no key)
python -m src.ingest.gdelt --start 2023-10-01 --end 2024-06-01

# 4. Harvest real headlines for the framing analysis (needs network; GDELT)
python -m src.ingest.headlines --start 2023-10-01 --end 2024-06-01

# 5. Run the three analyses
python -m src.analysis.coverage_volume
python -m src.analysis.framing
python -m src.analysis.selection

Layout

config.yaml                        outlets, query terms, date windows
data/seed/rocket_episodes_seed.csv sourced, usable-now rocket baseline
data/seed/SOURCES.md               provenance for every seed figure
src/ingest/acled.py                pull both-sides events from ACLED
src/ingest/gdelt.py                per-outlet coverage volume + tone
src/ingest/headlines.py            harvest per-outlet headlines (GDELT artlist)
src/baseline/rocket_baseline.py    derive rocket→Israel events + outcomes
src/analysis/coverage_volume.py    articles per event / per fatality by side
src/analysis/framing.py            agent attribution & passive-voice metrics
src/analysis/selection.py          which events get zero coverage, by side
docs/METHODOLOGY.md                assumptions, threats to validity, caveats