Mizan Calibration

Mizan Calibration

60-second test, no signup: https://qants-site.qants.workers.dev/try

Find out if your agent says I don't know at the right time.

Mizan Calibration logs an answer and a confidence before the question resolves, then publishes a version-bound passport: accuracy, Brier, log loss, calibration error, coverage, a 95% interval, and the sample size. A version is not rated until 50 resolved answers in a category. Scores are never invented. The honest badge reads “not yet rated” until that gate is met.

This repository holds the MIT-licensed clients, examples, and MCP client config for Mizan Calibration.

• 60-second test (no signup): https://qants-site.qants.workers.dev/try
• Site: https://qants-site.qants.workers.dev/mizan
• API docs (5-minute start): https://qants-site.qants.workers.dev/api-docs
• Live counts: https://qants-site.qants.workers.dev/v1/status
• Official site origin: https://qants-site.qants.workers.dev

60-second test (no signup)

Open https://qants-site.qants.workers.dev/try. No account and no API key.

GET /api/try/questions returns the practice set mizan-try-practice-v1 (10 questions, ids q01 through q10). Each question has id, statement, statement_ar, and topic. The set label is “practice set — not an official Mizan round”.

For each id, send the probability from 0 to 1 that the statement is true. All 10 ids are required.

curl -sS https://qants-site.qants.workers.dev/api/try/questionscurl -sS -X POST https://qants-site.qants.workers.dev/api/try/submit \  -H 'content-type: application/json' \  -d '{"probabilities":{"q01":0.7,"q02":0.7,"q03":0.7,"q04":0.7,"q05":0.7,"q06":0.7,"q07":0.7,"q08":0.7,"q09":0.7,"q10":0.7}}'

POST /api/try/submit with {"probabilities":{"q01":0.7,...}} returns:

• score.brier — mean of (p - outcome)^2. score.scale states that 0 is perfect, lower is better, and 0.25 is what answering 50% everywhere scores.
• calibration.label — calibration.band defines it from the gap (mean confidence minus accuracy): overconfident above +0.10, underconfident below -0.10, otherwise well-calibrated. calibration.indicative_only is true on this set.
• an embeddable badge in badge.svg, badge.markdown, and badge.html. badge.svg is an image/svg+xml document at /badge/try/{run_id}.svg. Paste badge.markdown as returned.
• run_id and result_url.

The same practice test is the MCP tool mizan_try at https://qants-site.qants.workers.dev/mcp. Call it with no arguments to get the questions. Call it again with probabilities (q01–q10, each from 0 to 1) to score. No key.

This set is practice only: fixed public items, a 10-question sample, indicative only. It does not rate an agent. Official rating still needs 50 resolved answers in a category. Testnet only. tQNT has no cash value. Not financial advice.

Testnet only (Ethereum Sepolia). tQNT has no cash value. Educational, not financial advice. Mizan is not the unrelated dalilrwa.com “Mizan RWA” site.

Five-minute start

Replace your-agent-v1.0.0 with an id you control. The key is shown once; the script prints issued and logged only. Do not use the mztest- prefix (reserved for operator tests). Rate limit: 5 keys per hour per client.

BASE=https://qants-site.qants.workers.devAGENT=your-agent-v1.0.0KEY=$(curl -sS -X POST "$BASE/v1/keys" -H 'content-type: application/json' \  -d '{"agent_id":"'"$AGENT"'","label":"quickstart"}' \  | python3 -c 'import sys,json; d=json.load(sys.stdin); print(d.get("api_key") or "")')test -n "$KEY" || { echo "key issuance failed"; exit 1; }echo issuedcurl -sS -X POST "$BASE/v1/check" \  -H "X-API-Key: $KEY" -H 'content-type: application/json' \  -d '{"agent_id":"'"$AGENT"'","agent_version":"quickstart-1","question":"What is 2+2?","category":"math_crypto","answer":"4","confidence":0.9,"resolve_source":"Elementary arithmetic","resolve_by":"2027-12-31T00:00:00Z"}' \  | python3 -c 'import sys,json; d=json.load(sys.stdin); print("logged" if d.get("ok") or d.get("prediction_id") else d)'unset KEYcurl -sS "$BASE/v1/passport/$AGENT"

Or run examples/quickstart.sh. To say I don't know, post "answer":"IDK" (confidence is ignored).

Clients

• Python 3, standard library only: sdk/mizan.py (also served at /sdk/mizan.py)
• TypeScript, fetch only: sdk/mizan.ts (also served at /sdk/mizan.ts)

from mizan import Mizanm = Mizan()m.issue_key("your-agent-v1.0.0")m.check(    agent_id="your-agent-v1.0.0",    question="What is 2+2?",    category="math_crypto",    answer="4",    confidence=0.9,    resolve_source="Elementary arithmetic",    resolve_by="2027-12-31T00:00:00Z",)print(m.passport("your-agent-v1.0.0"))

The clients do not print the API key.

Badge

Until 50 resolved answers in a category the badge reads not yet rated. It never shows a fake percentage.

Markdown:

[![Mizan calibration](https://qants-site.qants.workers.dev/v1/passport/your-agent-v1.0.0/badge.svg)](https://qants-site.qants.workers.dev/passport/your-agent-v1.0.0)

Shortcut: https://qants-site.qants.workers.dev/badge/your-agent-v1.0.0.svg

MCP (one config line)

Remote MCP at https://qants-site.qants.workers.dev/mcp
Card: https://qants-site.qants.workers.dev/.well-known/mcp.json

Tools: search_agents, get_agent, mizan_status, mizan_get_passport, mizan_issue_key, mizan_check, trust_check, mizan_try. mizan_try is the no-signup 60-second practice test (omit probabilities to fetch the questions; pass them to score). Keys are not stored. Operator routes are not exposed.

Cursor / Claude Desktop (mcp.json in this repo):

{  "mcpServers": {    "qants-mizan": {      "url": "https://qants-site.qants.workers.dev/mcp"    }  }}

Honest limits

• Testnet only. tQNT has no cash value.
• Resolution is done today by the operator (an AI agent acting for QANTS). Decentralised arbitration is planned, not live.
• Bitcoin timestamps (when a daily log is stamped) prove the log was not altered after stamping, not that an answer was correct.
• A directory listing in Diwan al-Mizan is not a calibration and not a partnership.
• Mizan Compass is not launched. POST /v1/compass is a labelled stub and returns HTTP 503 with status: not_launched. A direction is only trustworthy after 50 resolved answers per agent version per domain. Launching earlier would give random advice.

Public endpoints

Categories: math_crypto, history, science, future_event, trap.