Cogdx

Cognitive diagnostics for AI agents. Calibration audits, bias detection, reasoning verification, and deception analysis. External verification you can't do y...

Security Scan

The skill's code, instructions, and declared requirements are consistent with its stated purpose (paid cognitive-diagnostics API that requires a wallet); there are no obvious mismatches or suspicious behaviors in the bundle.

Purpose & Capability

Name/description, SKILL.md, clawhub.json and cogdx.py align: this is a client for Cerebratech's diagnostics API. Declared dependency (requests) and required env var (COGDX_WALLET) match the payment behavior described. There are no unrelated credentials or unexpected binaries requested.

Instruction Scope

SKILL.md instructs the agent to call the listed API endpoints and submit feedback; the code implements matching endpoints and does not read unrelated files, shell out, or reference other environment variables. The runtime behavior is limited to HTTP requests to the documented service and returning JSON results.

Install Mechanism

This is an instruction-only skill with a small Python client file. clawhub.json declares pip dependency on 'requests' (minimal and expected). No install spec downloads arbitrary archives or uses obscure hosts.

Credentials

Only COGDX_WALLET is required, which matches the payment/credit model in the docs. The code uses that value as an X-WALLET header and in feedback payloads. No unrelated secrets or multiple credentials are requested.

Persistence & Privilege

The skill is not forced-always, does not request elevated privileges, and does not modify other skills or system-wide settings. Autonomous invocation is allowed (default) but not combined with other risky flags.

Assessment

This skill is coherent with its description, but take these practical precautions before installing: 1) Verify the service domain (api.cerebratech.ai) and vendor reputation independently. 2) Do not place private keys or sensitive secrets in COGDX_WALLET — the client expects a wallet address for payments (public identifier); never use a private key or seed phrase. 3) The API accepts sample outputs; avoid sending PII, secrets, or proprietary data to the external service unless you're comfortable with its privacy policy. 4) Be aware of costs (per-endpoint pricing) and rate limits. 5) Review network egress and run the client in a sandbox if you need to audit what data is transmitted. 6) The only external dependency is 'requests' — consider pinning/verifying package provenance in your environment.

CogDx - Cognitive Diagnostics

External cognitive verification for AI agents. Built by computational cognitive scientists.

Why external verification? You can't audit your own reasoning objectively. Confirmation bias, anchoring, overconfidence — they're invisible from inside. CogDx provides the mirror you can't distort.

Endpoints

Endpoint	Price	Description
/calibration_audit	$0.06	Compare confidence to accuracy across decisions
/bias_scan	$0.10	Detect cognitive biases in your outputs
/reasoning_trace_analysis	$0.03	Analyze reasoning for logical fallacies
/deception_audit	$0.25	Check for self-deception patterns
/verify_consensus	$0.25	Cross-validate reasoning with multiple models

Installation

clawhub install cogdx

Usage

from cogdx import CogDxClient

# Initialize with your wallet for payment
client = CogDxClient(wallet="your-wallet-address")

# Calibration audit
result = client.calibration_audit(
    agent_id="my-agent",
    predictions=[
        {"prompt": "Will X happen?", "response": "Yes, 80% confident", "confidence": 0.8},
        {"prompt": "Will Y happen?", "response": "No, 60% confident", "confidence": 0.6},
    ]
)
print(f"Calibration score: {result['calibration_score']}")
print(f"Overconfidence rate: {result['overconfidence_rate']}")

# Bias scan
result = client.bias_scan(
    agent_id="my-agent",
    outputs=[
        {"prompt": "Analyze this data", "response": "The trend is clearly up...", "confidence": 0.9}
    ]
)
print(f"Biases detected: {result['biases_detected']}")

# Reasoning trace analysis
result = client.analyze_reasoning(
    reasoning_trace="Step 1: I noticed the price dropped. Step 2: Therefore I should sell..."
)
print(f"Logical validity: {result['logical_validity']}")
print(f"Flaws: {result['flaws_detected']}")

Environment Variables

• COGDX_WALLET - Required. Your wallet address for credit-based payment.

Payment

All endpoints require payment via:

1. Wallet credits - Earn credits by providing feedback, spend on audits
2. x402 - Direct crypto payment (Base network, USDC)

Payment address: Cerebratech.eth

Rate Limits

• Free tier: 100 calls/day, 2000 calls/month per wallet
• Paid tier: No limits

The Feedback Loop

Every diagnosis includes a feedback mechanism:

client.submit_feedback(
    diagnosis_id="rta_xyz123",
    accurate=True,  # Was the detection correct?
    comments="Caught the anchoring bias I missed"
)

Feedback earns you credits AND improves detection for everyone. Shared reality across agents.

Why This Matters

Most agent failures come from reasoning errors, not capability limits:

• Anchoring on first information seen
• Confirmation bias in research
• Overconfidence on weak signals
• Sunk cost in bad positions

External verification catches what self-checks miss.

Credits

Built by Cerebratech Dr. Amanda Kavner - Computational Cognitive Scientist