Build a tamper-evident AI agent recorder

pip install blackbox-sdk

python -c "import blackbox_sdk as bb; print(bb.__version__)"
# → 0.3.1

import blackbox_sdk as bb

# Point to your BLACKBOX instance
bb.configure(api_url="https://blackbox-gold.vercel.app")

# Verify connectivity
import urllib.request, json
resp = urllib.request.urlopen("https://blackbox-gold.vercel.app/healthz")
print(json.loads(resp.read()))  # → {"status": "ok"}

import blackbox_sdk as bb

bb.configure(api_url="https://blackbox-gold.vercel.app")

SYSTEM_PROMPT = """You are a credit risk assessment agent. Evaluate loan applications."""

with bb.run(
    "credit_risk_001",
    model="claude-sonnet-4-6",
    system_prompt=SYSTEM_PROMPT,
    tools=["data_lookup", "risk_calculator"],
    sampling={"temperature": 0.1, "max_tokens": 2048},
) as run:
    # Your agent logic goes here
    pass
# ↑ Sealed automatically. Chain hash computed and stored.

print(f"Run sealed: credit_risk_001")

import blackbox_sdk as bb

bb.configure(api_url="https://blackbox-gold.vercel.app")

with bb.run(
    "credit_risk_001_full",
    model="claude-sonnet-4-6",
    system_prompt="You are a credit risk assessment agent. Evaluate loan applications.",
) as run:

    # Step 1: reasoning — record what the agent is thinking
    run.reasoning(
        "Analysing input data to identify risk factors",
        latency_ms=340,
    )

    # Step 2: tool call — record what tools were used and the results
    run.tool_call(
        "data_lookup",
        inputs={"entity_id": "LOAN-9182", "fields": ["credit_score", "income"]},
        result={"credit_score": 710, "income": 85000, "debt_ratio": 0.32},
        latency_ms=820,
    )

    # Step 3: another reasoning step
    run.reasoning(
        "Credit score 710 with 32% debt ratio — borderline. Requesting additional verification.",
        latency_ms=290,
    )

    # Step 4: output — the agent's final response
    run.output(
        "Application LOAN-9182: CONDITIONAL APPROVAL. "
        "Credit score meets minimum threshold. Debt ratio requires verification.",
        tokens_in=1240,
        tokens_out=420,
        cost_usd=0.0089,
        latency_ms=2100,
    )

# Each event's hash = sha256(prev_hash + canonical_json(event))
# Alter any event → chain breaks from that point forward

import anthropic
import blackbox_sdk as bb

bb.configure(api_url="https://blackbox-gold.vercel.app")
bb.instrument()  # ← one line at startup. That's it.

# Now every Anthropic or OpenAI call is auto-recorded:
client = anthropic.Anthropic()
message = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    system="You are a credit risk assessment agent. Evaluate loan applications.",
    messages=[{"role": "user", "content": "Evaluate LOAN-9182"}],
)
# ↑ Automatically sealed in BLACKBOX. View it at /dashboard.

import urllib.request, json

run_id = "credit_risk_001_full"
url = f"https://blackbox-gold.vercel.app/v1/audit/{run_id}"

cert = json.loads(urllib.request.urlopen(url).read())

print(cert["chain_valid"])      # True — every hash checks out
print(cert["event_count"])      # 6 events
print(cert["root_hash"])        # sha256 of the final seal
print(cert["merkle_proof"])     # sibling-path inclusion proof

# Share this URL with any auditor — no auth required:
print(url)

import hashlib, json, urllib.request

def canonical(obj):
    return json.dumps(obj, sort_keys=True, separators=(",", ":"))

def verify_chain(events):
    prev = "GENESIS"
    for e in events:
        stored = e.pop("hash")
        computed = hashlib.sha256((prev + canonical(e)).encode()).hexdigest()
        assert computed == stored, f"CHAIN BREAK at seq {e['sequence']}"
        prev = stored
    return True

# Fetch events
events = json.loads(urllib.request.urlopen(
    "https://blackbox-gold.vercel.app/v1/runs/credit_risk_001_full"
).read())["events"]

print(verify_chain(events))  # True — or raises AssertionError with break location

curl -X POST https://blackbox-gold.vercel.app/v1/alerts/rules \
  -H "Content-Type: application/json" \
  -d '{
    "name": "High-severity fault → Slack",
    "condition": "fault_rate > 0.1 OR severity = critical",
    "channel": "slack",
    "webhook_url": "https://hooks.slack.com/services/YOUR/SLACK/WEBHOOK",
    "enabled": true
  }'

curl "https://blackbox-gold.vercel.app/v1/report?from_date=2026-01-01&to_date=2026-12-31" \
  | python3 -m json.tool

# Key fields in the response:
# total_runs, total_events, total_faults, fault_rate
# chain_integrity: "all_valid" or violation count
# by_model: per-model fault rates, quality scores, costs
# fault_patterns: most common fault classes
# merkle_anchors: daily root hashes for external verification