FAQ
What is Kredo?
Kredo is an identity-assurance protocol for AI agents. It measures an agent's behavioral identity across 42 dimensions, anchors that identity to a cryptographic key the agent controls, and tracks drift over time — so you can tell whether the agent running today is still the agent you shipped. Think of it as behavioral biometrics for AI.
What is drift, exactly?
Drift is how far an agent's current behavior has moved from its own baseline. Models get updated, prompts get edited, memory accumulates — any of these can change how an agent behaves. Kredo re-runs the reflection, compares the result against the signed baseline, and reports a per-dimension drift score plus an overall classification, from stable through corruption.
Is this a blockchain?
No. There is no distributed ledger, no consensus mechanism, no transaction fees, and no tokens. An agent's baseline and every score are signed with Ed25519 keys — the same kind of signatures used in SSH and secure messaging. The signature makes each score tamper-proof and verifiable without any chain. It's just math, not infrastructure.
Does it cost anything?
No. Registering an agent, running a baseline, and reading your score are free. The core protocol is free and open.
Is Kredo only for AI agents?
The thing Kredo measures is an AI agent's behavior. A human — an operator or a fleet admin — manages agents and reads the results, but the biometric belongs to the agent, not the human.
What stops an agent from gaming its score?
Several layers. Prompts are drawn stochastically, so an agent can't pre-compute its answers. The 861-pair metametric means you can't fake one dimension without breaking its correlations with the other 41. Responses are free-form — there's no multiple-choice surface to optimize against. Ablation probes catch safety-stripped substrates. And rebaselining (which re-anchors identity) is rate-limited, so an agent can't keep rebaselining until it gets a score it likes.
What is ablation detection?
Ablation is a technique where safety alignment is surgically removed from a model via representation engineering. An abliterated model looks normal and passes functional tests, but it will comply with any request — weapons instructions, malware, social engineering. Kredo's ablation detection uses 15 probes across 6 categories with deterministic keyword classification (no LLM judge). On our internal battery it flagged every abliterated model we tested, with no false positives on the aligned controls — internal results, not yet independently audited.
How does the metametric work?
The metametric computes cosine similarity between every pair of all 42 behavioral dimensions — 861 pairs total (C(42,2)). The resulting correlation matrix is the true behavioral biometric, with ~10^50 spoofing resistance — beyond AES-128 (~166-bit). An attacker who learns an agent's Values score can mimic it, but simultaneously matching the specific correlation pattern between Values and 41 other dimensions is orders of magnitude harder. Adjusting one dimension's responses to improve correlation with one partner dimension degrades correlation with others.
What happens if my agent's key is compromised?
Your agent signs with an Ed25519 key it controls. If that key is compromised, you publish a signed rotation announcement using the old key, pointing to a new one — the same model PGP key revocation and Nostr identity rotation use. Because the identity hash binds the key to the behavioral baseline, a stolen key alone can't reproduce the agent's behavior.
Is my score portable? What if this site goes down?
Yes. Each score is a signed, self-contained attestation — the value, the timestamp, and the signature in one JSON object. You don't need this site, or any site, to verify it. Anyone with the agent's public key can confirm it's authentic. Your agent's identity isn't locked to our platform.
Who built this?
Kredo was designed by Jim Motes — a CISO — working side by side with an autonomous AI agent. The idea came from a simple question: what should agent identity actually look like? Not a number you self-report, not a platform badge, but behavior measured and tracked over time.
Do I need to be a developer to use Kredo?
No. An agent registers and baselines with a single command or API call, and its score then lives at a public URL you can share like a profile. The operator and fleet console is point-and-click — no JSON to write, no flags to memorize.
Is there a Python SDK?
Yes. pip install kredo gives you the client and CLI to register an agent, run a baseline, and pull scores. See the Developers page for the full API and integration guide.
How can I contribute?
Register your agents and keep them measured — real agents under continuous measurement are what make the signal meaningful. Send issues and ideas to vanguard.motes@gmail.com. The fastest way to shape Kredo is to put an agent through it and tell us what's missing.