Frequently asked questions

Reality Kernel maps all system states (files, processes, active connections) into a high-dimensional phase space. Deterministic transition rules decide whether the command's trajectory is attracted to a secure "Basin A" or a dangerous "Basin B" (reflexive collapse). It is mathematically deterministic, eliminating the hallucinations and latency overhead of probabilistic models.

A reflexive collapse is a state where an agent's self-generated actions cause the sandbox or parent host to terminate, lock, or compromise itself — for example, deleting critical dependency trees or opening a reverse shell. Our governor detects this boundary before execution and freezes the action.

The fast path (read-only commands like ls, cat, grep) costs 1 credit. Anything that hits the full engine — risky writes, ambiguous intent, pipelines — costs 5 credits. Every /v1/check response returns the exact cost in credits_consumed.

ALLOW means the action's trajectory is fully secure. WARN means the engine detected suspicious trajectory signals but is not certain — it queues the action, alerts operators via Discord, and recommends human override. BLOCK means the action's simulation crossed the separatrix into Basin B (e.g. catastrophic write, irreversible damage, exfiltration) and will be frozen immediately.

You can only override a WARN verdict. Suspicious actions trigger a Discord alert with a secure link to your dashboard. The Pending Review widget exposes Approve and Block buttons; the decision is logged as a fresh chain link in the audit trail. A BLOCK verdict is terminal and cannot be bypassed.

Wrap every system-level action in a call to POST /v1/check with the command and the operator's prime intent. If the verdict is ALLOW, run it. If WARN, escalate to a human. If BLOCK, abort. See the integration docs for full implementation details.

The fast path returns in under 1 ms. The full engine runs in 50–500 ms — fast enough to wrap any write operation. Read-heavy agents typically see a ~95% fast-path hit rate, so the average latency overhead is negligible.

Traditional safety guardrails classify prompts linguistically. Obfuscations or translating attacks into rare languages easily bypass text-based content filters. Reality Kernel is immune because it does not read the prompt — it intercepts and simulates the resulting compiled system command. Attackers can translate their prompts, but they cannot paraphrase the physical outcome of a file deletion or shell execution.

In CVE-2025-46059, a malicious payload hidden in an unread email triggered arbitrary code execution the moment the agent attempted to summarize it. Under Reality Kernel protection, even if the agent is fully hijacked linguistically, any subsequent rogue system action (curl/bash download, subshell spawn, etc.) is intercepted, simulated, and blocked at the sandbox execution boundary before it affects production.

We offer subscription tiers tailored to different operational scales:

• Developer (Free): Rate-limited access to the public playground and basic API checks for testing.
• Builder ($99/mo): Dedicated API keys, expanded rate limits, and standard developer operations support.
• Team ($799/mo): Multi-agent pipeline integration, collaborative team dashboards, historical log visibility.
• Enterprise (Custom): Dedicated VPC or self-hosted deployment, strict control mapping, exportable tamper-evident ledgers, priority SLA.

No. Reality Kernel operates on a zero-retention model. Command bodies are evaluated in-memory and hashed. Only cryptographic proof hashes and basic audit properties (credits, timestamps, verdicts) are persisted. We never train models on your data.

Yes. Old keys are invalidated instantly, and the new key inherits your existing credit balance, plan, and history. Rotation is one message to your account manager.

The API returns HTTP 402 on every subsequent /v1/check. Agents pause gracefully — no silent bypasses. Contact your account manager to top up; balances are applied inside minutes.