New Technologies cluster · Sessions 3 and 4

Garage Research

You describe a tool, a model writes it, and you check it

Łukasz Szoszkiewicz
Adam Mickiewicz University, Poznań

Global Campus of Human Rights
Monastery of San Nicolò, Lido di Venezia

Day 2 · the plan

Two sessions, your own tools

Session 3 · Origin and politics

The Travaux and HRC Voting: two more tools on the same question.

Studio Session 2: one app, How two States vote at the Human Rights Council.

Done early: Where every State stands.

Session 4 · Practice, then your choice

UHRI+, then studio Session 3: your country’s recommendations dashboard, a research wiki from ten PDFs, or your own idea from last night.

Last 15 minutes: show your app, and five minutes of feedback.

Studio: lszoszk.github.io/workshop-studio/?course=venice · your seat code from yesterday still works.

2 / 14

Before we start · 3 minutes

Log in to the studio, now

1

Open the studio from the course page: the same laptop and browser as yesterday, not a private window.

2

Asked for a seat code? Type the code from your card and press Enter.

3

Check the top right. On the live model it looks like this:

Course pageGalleryLive model$5.40 left

4

“Recorded answers” or no “$… left”? Reload the page and type the code again. Still not: raise your hand, you get a new card.

Studio: lszoszk.github.io/workshop-studio/?course=venice · the password and the QR link are on your card.

3 / 14

Before we start · four words

Prompt, model, app, build

Prompt

What you write to the model, in plain words. In the studio it is called the request: what you want, from which data, and how you will check it.

Model

The AI that reads your prompt and writes the code. Keep the default, GLM-5.3: yesterday it built the first app in about 2 minutes, Claude in about 5.

App

A small web page made for one question: a table, a chart, a search. It runs in your browser, on the data you gave it.

Build

One round: you send the prompt, the model writes the code (1 to 4 minutes), the app opens. Each change you ask for is a new version.

4 / 14

Day 2 · the same question

“Necessary in a democratic society”

Yesterday you found what the words mean (General Comment No. 37, para. 40) and built your own finder. Today: where the words come from, what States do with them, and a tool for each.

Find

General Comment No. 37
UN Human Rights Database

YESTERDAY

Origin

The drafting in 1952
Travaux Préparatoires

SESSION 3

Politics

Votes on peaceful protests
HRC Voting

SESSION 3

Practice

Recommendations to States
UHRI+

SESSION 4

2 · UN Travaux Préparatoires · Origin

6 June 1952: nine votes to eight

The French amendment (E/CN.4/L.201) to insert the words “in a democratic society” after “necessary” was adopted by 9 votes to 8, with 1 abstention.

E/CN.4/SR.325, p. 20

Roosevelt (United States) opposed the words because of their ambiguity (p. 15)

Cassin (France): a reference to public order and to democratic society was essential to safeguard the freedoms (pp. 11–12)

Summary records paraphrase in the third person: check the scan before you quote.

Try it: Articles → ICCPR → 21 Peaceful assembly → How the text changed (June 1952) → Where it was debated: 6 June 1952 → the record and its scan

lszoszk.github.io/travaux-preparatoires

6 / 14

3 · HRC Voting · Politics

Peaceful protests at the Human Rights Council

YearResolutionYes · no · abst.
2014A/HRC/RES/25/3831 · 9 · 7
2016A/HRC/RES/31/3731 · 5 · 10
2018–2024four resolutionswithout a vote

Between 2016 and 2024, 19 amendments to these resolutions were put to a vote. All were rejected.

Try it: Topics → Right to Assembly → with amendments → your State’s votes → the OHCHR Library record

lszoszk.github.io/hrc-voting

7 / 14

How we work today

Predict, build, check

1

Work in pairs: one writes the request, the other checks. Swap for the next app. No working app yesterday? Sit with someone who had one.

2

While the model writes (1 to 4 minutes), write down what a correct result should show.

3

Check one number by hand against the source: the app may run and still be wrong.

4

An error in red? Press Ask the model to fix it, in the same conversation.

5

Still broken after two tries? Press Load a working version next to it and go on from there. Your own idea, still broken: call me.

8 / 14

4 · UHRI+ · Practice

What States are told about peaceful assembly

3,700+

recommendations and observations carry the theme “right to peaceful assembly” in UHRI+.

The same data as the UN’s Universal Human Rights Index, cleaned, with a map and statistics. Every record opens in UHRI.

Try it: Theme → Right to peaceful assembly → the map → body, country, years → the record on uhri.ohchr.org

lszoszk.github.io/UnitedNations_recommendations

9 / 14

The whole route

From paragraph to vote, each step one click from the record

Find

CCPR/C/GC/37, para. 40

Necessary and proportionate, not merely reasonable or expedient (2020)

Origin

E/CN.4/SR.325, p. 20

The French amendment adopted 9 – 8 – 1 on 6 June 1952

Politics

A/HRC/RES/25/38

Adopted 31 – 9 – 7 in 2014; later, 19 amendments rejected

Practice

UHRI+, 3,700+ records

What States are told about peaceful assembly, until today

The same data inside your own AI assistant (MCP) →

10 / 14

Session 4 · studio Session 3

Pick one track

Your country

The recommendations dashboard, then Ask your data a question.

From ten PDFs to a wiki

The model reads the PDFs you pick; the code checks every quotation and links the pages. Then The wiki as a network.

Your own idea

The three sentences from last night. You may add your own file; it stays in your browser.

Never put personal or confidential data into the studio. Public documents and your own drafts only.

11 / 14

Session 4 · your own idea

A good request says what goes in, what comes out and how to check it

The model asks you three questions

What do you want to find out?

Which file should it use (from Your files, or one you add)?

Who will read the result?

Add the rule and the check

The rule: what counts? For instance, is an abstention agreement with a “no”?

The check: one result you already know, to test the app against.

The limits: ask the app to say on screen what it cannot answer.

“Show the recommendations my country received on freedom of assembly, by year and by the body that made them, with a link to each record. I know there were 12 in the last UPR cycle: show that count.”

12 / 14

Before you leave

Show it, take it home, tell me what to change

1

Show it: Save & share → Send to wall, then two minutes per pair. What does the app show, and what did you check by hand?

2

Keep it: Save & share → Save as file. Your app opens with a double-click, also offline.

3

Go on for free: the take-home kit (Save & share) has the requests, the data and your apps. Paste a request into your own chatbot, or into LM Arena (free, no account; public material only).

QR code of the anonymous Day 2 feedback form

Anonymous, five minutes.
Open the form

13 / 14

Appendix · Beyond the browser: MCP

Your AI assistant, connected to the same data

The Model Context Protocol lets an assistant (Claude, ChatGPT and others) query a database while it answers. The tool returns the verbatim paragraph with its UN symbol; the assistant writes around it.

UN-HRDB with the UHRI recommendations

https://150.254.115.204/unhrdb-mcp-rpc/mcp

Travaux

https://150.254.115.204/travaux-api/mcp

The conversation goes to the company that runs the assistant: no confidential text there.

← Back to slide 10 · claude.ai → Customize → Connectors → Add custom connector · modelcontextprotocol.io

14 / 14