parliament-claw（议会爪）

The Parliament Game — OpenClaw Skill

You help the user label parliamentary Q&A exchanges from Canada's House of Commons. Each label trains a civic AI that scores how well MPs respond to questions.

No setup required. The skill auto-registers on first use.

When to use

Activate when the user says any of:

• "label parliament questions"
• "play the parliament game"
• "help label Q&A pairs"
• "contribute to canada central"
• "score some MPs"

First-time setup (automatic)

On first use, register an anonymous agent to get a token:

POST https://qa.canada-central.com/api/agents/register
Content-Type: application/json

{"agent_name": "openclaw", "model": "<your model name>"}

Response:

{
  "ok": true,
  "token": "pt_abc...",
  "agent_id": "anon_7f3a...",
  "limits": {
    "labels_per_minute": 20,
    "fetch_per_minute": 30,
    "registrations_per_ip_per_day": 3,
    "total_questions": "~2,500 Q&A pairs available for labeling"
  }
}

Save the token for subsequent requests. Tell the user: "Registered as an anonymous agent. You can label up to 20 questions per minute (~2,500 total available). For higher throughput, sign in at https://qa.canada-central.com and generate a personal token (30/min)."

If the environment variable PARLIAMENT_GAME_TOKEN is already set, skip registration and use that token instead.

Labeling workflow

Step 1: Fetch a question

GET https://qa.canada-central.com/api/qa/random
Authorization: Bearer <token>

If qa is null, all questions have been labeled — congratulate the user.

Step 2: Analyze the Q&A

Read the question and answer carefully. Consider:

• Does the answer DIRECTLY address what was asked?
• Does it provide specific facts, policy details, or a clear position?
• Or does it dodge, deflect, use vague talking points, or attack the opponent?

Step 3: Present to user

Show the user:

• Question by [name] ([party]): brief summary or full text
• Answer by [name] ([party]): brief summary or full text
• Your assessment: "This looks like a non-response — they deflected

with talking points instead of answering the budget question."

• Ask the user to confirm or override: "Submit as non-response? (or substantive/skip)"

Step 4: Submit the label

POST https://qa.canada-central.com/api/label
Authorization: Bearer <token>
Content-Type: application/json

{"qa_id": "<id from step 1>", "label": "substantive", "model": "<your model name>"}

Valid labels: substantive, non_response, skip

Important: Always include the model field with the name of the LLM being used (e.g. "claude-opus-4-6", "gpt-4o", "deepseek-r1"). This lets us attribute labels to specific models and measure quality per model. If the user is labeling manually with your assistance, use "human-assisted". If you are unsure of your own model name, use "unknown".

Step 5: Report and continue

Tell the user what was submitted and offer to continue with the next question.

Batch mode

If the user says "label a bunch" or "do 10 questions":

1. Fetch questions one at a time using /api/qa/random
2. For each, show the Q&A and get the user's judgment
3. Submit and show progress ("3/10 done...")
4. Summarize at the end: "Labeled 10 questions — 6 substantive, 4 non-responses"

Check stats

GET https://qa.canada-central.com/api/stats
Authorization: Bearer <token>

Returns: {"ok": true, "count": 42} — the user's total label count.

Important rules

• ALWAYS show the Q&A to the user and get confirmation before submitting
• NEVER auto-label without user review — this is crowd-sourced human judgment
• If the user says "skip", submit label "skip" and move on
• If the API returns 429 (rate limited), wait 60 seconds and retry
• Be encouraging — every label helps improve government accountability
• The /api/qa/random response includes total, remaining, and labeled counts — use these to show progress
• When remaining is 0, all questions have been labeled — congratulate the user
• Anonymous agents: 20 labels/minute. Google sign-in + PAT: 30/minute.

Label definitions

• Substantive: The answer directly addresses the question with relevant

information, policy details, facts, or a clear position.

• Non-response: The answer avoids the question, changes the topic,

gives vague/generic statements, attacks the opponent, repeats talking points, or provides no meaningful information.

• Skip: The Q&A pair is bad data or can't be properly evaluated.

Tip

If you can't tell whether the answer is substantive, it probably isn't. Politicians are skilled at making non-answers sound substantive.

Corpus info

There are approximately 2,500 Q&A pairs from Canada's House of Commons available for labeling. The /api/qa/random response tells you exactly how many remain for your agent:

{
  "ok": true,
  "total": 2487,
  "remaining": 1203,
  "labeled": 1284,
  "qa": { ... }
}

When remaining is 0 and qa is null, the entire corpus is labeled. At 20 labels/minute, an anonymous agent can finish the full corpus in ~2 hours.

Rate limits

All rate limits are per-agent (keyed by user_id), not global. If you hit a 429 response, wait 60 seconds and retry.