ImpactFeed

Discoverable, EA-aware tracking of EU public consultations (extensible to other jurisdictions), with cause-area filtering and impact scoring so individuals and experts can act where it counts most.

Built for the EAGx Stockholm 2026 mini hackathon — 1 hour, 3 people.

---

1. The Problem

Lawmakers in the EU and most democracies are constitutionally obliged to consult citizens, experts, and civil society before adopting new laws or regulations. In the EU this is codified in Article 11 of the Treaty on European Union and operationalised through the Have Your Say portal, which has hosted every Commission consultation since November 2018.

In theory, this is one of the highest-leverage points an EA-aligned individual or expert can act: a well-argued submission can directly shape the wording of a law that affects hundreds of millions of people. In practice, the consultation pipeline is broken at almost every stage:

Problem	What the evidence says	Source
Citizens don't know consultations exist	The European Commission's own 2019 Stocktaking identifies "lack of knowledge on behalf of the EU citizens with regards to the participatory opportunities" as the #1 problem of the consultation regime.	European Court of Auditors — Have Your Say special report (2019)
Dense legalese blocks ordinary participation	Studies of EU consultations show citizen participation drops sharply when consultation documents use complex/technical language.	'Have Your Say' in practice (Tandfonline, 2025)
Lobbyists and industry dominate	"Very limited participation of citizen groups… reduces the diversity of viewpoints and expertise provided to policymakers, potentially resulting in policy decisions less attentive to the public interest."	Conceptualizing consultation approaches (Springer, 2020)
Black-box afterward	Citizens have no visibility into how their input was weighed, leaving little incentive to engage again.	Public Consultations Unpacked (European Law Blog)
Analysis is enormously expensive	The UK government estimates manual analysis of its consultations consumes ~75,000 working days and £20m / year.	AI tool 'Consult' — UK government case study
AI-generated comment floods are an emerging threat	Generative models can flood consultations with synthetic-but-plausible comments, drowning out genuine input.	How AI/LLMs may impact transparency (WFD)

The EA community is well-positioned to do something about this: many of the most consequential consultations (AI governance, biosecurity, animal welfare, global health) are exactly the cause areas where EA-aligned experts have outsized counterfactual value — if they know the consultation is happening.

Today, they usually don't. The EA Forum surfaces individual consultations occasionally and ad-hoc (e.g. UK aid review, hen cages), but there is no systematic tracker.

---

2. The Existing Landscape

There are useful tools in this space, but each is built for a different audience. None aggregates across jurisdictions with an EA cause-area lens and an opportunity-for-impact score.

Tool	Discovers consultations	Cross-jurisdiction	EA cause-aware	Impact-scored	Citizen-side	Free / OSS
EU "Have Your Say" portal	✅ EU only	❌	❌	❌	✅	✅
gov.uk consultations	✅ UK only	❌	❌	❌	✅	✅
EC directorate newsletters (Digital, R&I etc.)	✅ siloed by topic	❌	❌	❌	✅	✅
Consult (UK Incubator for AI)	❌ (analyses responses)	❌	❌	❌	❌ (gov-side)	❌
Talk to the City (AI Objectives Inst.)	❌ (analyses responses)	❌	❌	❌	❌ (gov-side)	✅
PolicyPulse (academic)	❌ (analyses discussions)	❌	❌	❌	❌ (researcher-side)	✅
Quorum / Plural Policy (commercial)	✅ broad	✅	❌	❌	❌ (lobbyist-side)	❌
EA Forum — Policy topic	Ad-hoc, occasional	Partial	✅	❌	✅	✅
ImpactFeed (this project)	✅ EU at MVP, multi-jurisdiction post-v0	✅ (extensible)	✅	✅	✅	✅

The wedge: discovery + EA cause classification + opportunity score + opt-in alerts, citizen-and-expert-facing, open source.

---

3. Our Idea

Build a tool that:

1. Crawls the websites and feeds of the EU Commission ("Have Your Say") for active consultations. Designed to extend to other jurisdictions (national EU member-state portals, UK, US federal rulemaking, etc.) post-v0.
2. Aggregates the results into a structured store.
3. Classifies each consultation against a borrowed 80,000 Hours / EA Forum cause-area taxonomy (multi-label).
4. Scores "opportunity for impact" using an LLM judgment against an explicit rubric — and shows the reasoning so users can override it.
5. Publishes through a clean website + opt-in email digests, with subscribers selecting cause areas and minimum impact thresholds.

Long-term (explicitly out of MVP): an opt-in expert directory that surfaces relevant consultations to subscribed experts. No automated cold outreach — that path is borderline spam, hurts trust, and runs into EU AI Act transparency obligations.

---

4. The 1-Hour Hackathon Build

4.1 What we ship in 60 minutes

A working demo, not a deployed product. The story: "Here's a real list of open EU consultations, classified by EA cause area and scored for impact, with the reasoning you can read."

4.2 Audience

EA-aligned individuals and individual experts (e.g. a biosecurity researcher who would respond to a relevant consultation if prompted). Not EA orgs with their own policy teams (v2 audience). Not the general public.

4.3 Initial cause areas (narrow on purpose)

Two only, multi-label:

• AI governance & safety — active EU AI Act delegated and implementing acts provide steady consultation flow.
• Animal welfare / factory farming — the EU regularly consults on cage bans, welfare labelling, transport rules.

Anything else → Other bucket, not surfaced in alerts.

4.4 Data source (v0)

EU Have Your Say only. No documented public API, so:

• Plan A: quick BeautifulSoup scrape of the open-consultations listing.
• Plan B (fallback): if the portal is JS-heavy / fights us inside 15 minutes, hand-curate ~10 real open consultations into consultations.json. The pipeline + UI is the demo, not the scraper.

Other jurisdictions (national EU member-state portals, UK, US federal rulemaking, etc.) are post-v0.

4.5 Impact scoring

LLM-judged (Claude), 1–5, with a written rationale visible to the user. Single number, multidimensional rubric:

Dimension	What it captures
Scope	How many people / animals / future generations are affected?
Reversibility	Is this a one-shot decision or revisable later?
Tractability	Is the consultation genuinely open to influence, or rubber-stamping?
Urgency	Deadline proximity — and is there time for a quality response?
Counterfactual	Will EA-aligned voices likely otherwise be heard? (low → higher score)

One Claude call per consultation: input is title + summary + cause tags + rubric; output is {score: 1-5, rationale: string, dimension_notes: {...}}. Scores are not authoritative — they are visible-reasoning suggestions users can disagree with.

4.6 Tech stack

Layer	v0 (shipped)	Post-hackathon
Pipeline	Python + Anthropic SDK (Claude Opus 4.7), Pydantic structured output, prompt caching	Same
Consultation store	JSON (data/scored.json, committed)	Postgres
Scraper schedule	Run locally on demand	GitHub Actions cron, daily
Frontend	Single static index.html + vanilla JS, Inter font	Astro (post-v0)
Subscriber store	Neon Postgres (via Vercel Storage marketplace; subscribers table, see api/_schema.sql)	Same
Subscribe API	Vercel serverless function api/subscribe.js (Node, @neondatabase/serverless)	Same + send-digest cron
Hosting	Vercel (static + functions)	Same + custom domain

---

4.6.1 Local development

# 1. Python deps + .env (one-time)
pip install -r requirements.txt
cp .env.example .env  # then fill in ANTHROPIC_API_KEY

# 2. Score consultations (re-run whenever data/consultations.json changes)
python -m pipeline.classify_and_score   # writes data/scored.json

# 3a. Static-only preview (no /api/subscribe — form submissions will fail)
python -m http.server 8000              # open http://localhost:8000/

# 3b. Full local stack (static + /api/subscribe). Requires the Vercel CLI.
cd api && npm install && cd ..          # install function deps once
npx vercel link                         # link to your Vercel project
npx vercel env pull api/.env.local      # pull DATABASE_URL into the function env
npx vercel dev                          # open http://localhost:3000/

4.6.2 Deploying to Vercel

One-time setup, ~5 minutes:

1. Push to GitHub (already done if you're reading this on github.com).
2. Go to vercel.com/new, Import this repo. Framework preset: Other. Build settings: leave default. Click Deploy. The first deploy will make the static frontend live; /api/subscribe will return 500 until the database is connected.
3. In the project dashboard → Storage → Create database → Postgres (this provisions a managed Neon Postgres via Vercel's marketplace) → pick a region close to you → Connect. Vercel auto-injects DATABASE_URL (and friends) as env vars on all functions.
4. Open the new database → Query tab → paste the contents of api/_schema.sql and run it. (One CREATE TABLE, one CREATE INDEX.)
5. Trigger a fresh deploy (Deployments → ⋯ → Redeploy) so /api/subscribe picks up the new env vars.

After that, every push to main redeploys automatically.

To inspect subscribers: Vercel dashboard → your database → Query → SELECT * FROM subscribers ORDER BY created_at DESC;

4.7 60-minute task split (3 people)

Time	Person A — Data	Person B — LLM	Person C — UI
0–10	Repo setup, scaffold dirs	Repo setup, draft scoring prompt	Repo setup, scaffold static page
10–40	Scrape Have Your Say → consultations.json (or hand-curate 10 if scraper fights you)	classify_and_score.py: reads JSON, calls Claude, writes scored.json	Page reads scored.json; filter by cause + min score; click row → show rationale
40–50	Help wherever stuck	Help wherever stuck	Add Tally subscribe embed
50–60	Integration test, demo dry-run	Integration test, demo dry-run	Integration test, demo dry-run

4.8 Out of scope for v0 (deferred deliberately)

• Expert outreach bot and any automated cold contact
• Other jurisdictions: national EU member-state portals, UK, US federal rulemaking, etc. (post-v0)
• SQLite / database (JSON suffices for ~tens of consultations)
• Real email sending — subscriptions are persisted to Postgres, but no digest is mailed yet (v1 = Resend, weekly Monday digest)
• Cron / hosting for the scraper (v1 = GitHub Action)
• Multilingual (EU consultations are normally available in English)
• "Did my comment matter?" — scrape final adopted text, LLM-map which submitted themes survived (v3)
• AI-generated comment / astroturf detection on the response side
• User accounts (preferences encoded in the subscribe form)

4.9 Sustainability / handoff

TBD. Identify a partner before / during the hackathon so the project doesn't die at demo day. Candidates: EA Forum moderators, an EA policy org (e.g. Center for Long-Term Resilience), Apart Research, or a 80,000 Hours team member.

---

5. License

MIT.

---

6. Status

🚧 Pre-build. Specs locked, ready to start the 60 minutes.