Algorithm register
Every algorithm AGORA uses to decide or to propose, described in the Algorithmic Transparency Standard 0.3.1 (Eurocities and the Dutch algorithm register), field by field. The same register as JSON: https://agorathens.gr/ai-register.json.
A prototype for the hackathon, not an official register of the City: written from what the code does, with the risk categories as the prototype's own reading, not a legal opinion.
The staff rule reader (spells)
Reads a booking rule that City staff write in Greek and drafts it as a Cedar policy with its legal basis; staff review it, replay it against past events and activate it. It never activates a rule itself.
- Type: hybrid; prescriptive. Status: development. Risk: limited risk (the prototype's reading).
- How it decides: Staff write a rule in Greek on the staff page. The model proposes a Cedar policy and the provision it rests on. Deterministic checks then validate it against the City's policy schema and the verified legal corpus. Staff replay it against past events and activate it, two different members when two-person activation is on. Nothing takes effect without that human activation.
- A person's part: Staff review and activate every rule; the model cannot activate one. Two-person activation is on by default.
- Data: The rule text staff write, the City's Cedar policy schema, and the regulation's values (catalog/regulation.yaml).
- Model and methods: Anthropic Claude (claude-opus-5-5, pinned) through one model interface (agora/model.py); Cedar policies (cedarpy 4.12.2).
- Legal basis: None loaded for the prototype. The rules it drafts must each rest on a provision of the verified corpus, or say that none is loaded. A pilot needs the City's own basis for using the reader.
- Objections: A refused booking names its rule and legal basis; the person can ask the City's staff. The prototype has no formal objection procedure.
The City's assistant (planner and extractor)
Answers a person's question about the City by planning reads of the City's public tools, and proposes a booking, a cancellation or a withdrawal that runs only when the person approves its exact hash.
- Type: hybrid; prescriptive. Status: development. Risk: limited risk (the prototype's reading).
- How it decides: The model plans which read tools to call, in at most 6 turns and 8 tool calls; upstream text reaches it only as handles, never as raw text. It can propose a write, stored with a nonce and a hash; the write runs once, through the personal tools' own path, only when the person approves that exact hash.
- A person's part: The person approves the exact plan before anything is written; they can revoke the mandate at any time.
- Data: The person's question, and the answers of the City's own read tools (each with its source).
- Model and methods: Anthropic Claude through one model interface (agora/model.py); a closed extraction schema; a bounded loop.
- Legal basis: None loaded for the prototype; a pilot needs the City's own basis and a DPIA.
- Objections: Every action has a receipt; the person can cancel or withdraw, or ask the City's staff. No formal procedure in the prototype.
The allocation draw for scarce spaces
When more groups ask for a municipal space than it can hold, a published rule set and a public random beacon decide the allocation; anyone can recompute the draw from public data.
- Type: rule-based; prescriptive. Status: development. Risk: limited risk (the prototype's reading).
- How it decides: Staff open a round; requests are collected; at the close, the versioned allocation rules (each resting on whole sentences of the verified corpus) and a constraint solver allocate the spaces, with ties broken by the drand beacon's randomness. Staff run the round on the staff page; the result is published to the City's feed.
- A person's part: Staff open, close and run each round; the rules change only by a new, published version.
- Data: The round's requests, the spaces' registry (catalog/spaces.yaml) and the allocation rules.
- Model and methods: OR-Tools CP-SAT; drand; the pure allocation module (agora/allocation.py) with no database, network or clock.
- Legal basis: Each allocation rule names whole sentences of a provision in the verified corpus (catalog/allocation_rules.yaml).
- Objections: Anyone can recompute the draw from public data (verify/verify_draw.py); a person can ask the City's staff.
The City's booking rules and agent tiers
Decides, by written rules, whether a booking of a municipal space is allowed and which agents may call which personal tool, and how often; every refusal names its rule.
- Type: rule-based; prescriptive. Status: development. Risk: minimal risk (the prototype's reading).
- How it decides: Cedar policies (one permit and one forbid per rule, each with its legal basis) and worded checks both run on every booking; a request is refused if either refuses. The agent-tier policy decides which tier of agent may call which personal tool, with per-minute quotas. No model takes part.
- A person's part: Staff change the rules only by a new, signed policy version (the rule reader above, reviewed and activated by staff).
- Data: The booking request, the spaces' registry and the regulation's values (catalog/regulation.yaml).
- Model and methods: Cedar; agora/city_policy.py and agora/agent_policy.py.
- Legal basis: Each rule names its provision in the verified corpus, or says that none is loaded (catalog/policies/city_rules.cedar).
- Objections: A refusal names its rule and legal basis and the next step; the person can ask the City's staff.