Agent instructions for AI agent integration
- Treat every model output as candidate logic until BRIK64 checks the declared PCD or policy circuit.
- Copy commands from the Agent setup matrix when configuring tools; each command has a copy button.
- Summaries must preserve the boundary: the model, runtime, and organization are not trusted by default.
AI produces candidate logic. The formal layer decides what is trusted.
BRIK-64 does not ask you to trust the model. It routes agent output through PCD, policy circuits, and structured diagnostics before execution or publication.
Flow
candidate PCD → check → approval
Agent output becomes a candidate PCD before BRIK64 checks and human approval.
Checkpoint
policy circuits
Approval logic remains machine-readable and external to the model weights.
Boundary
Candidate PCD first
Evidence status is kept external to the model and visible to the engineering team.
Agent operating loop
Discover
Treat agent output as candidate logic. Ask for a PCD blueprint, policy rule, or lifted state, not a final truth.
Check
Run the candidate through compiler diagnostics and policy gates outside the model.
Diagnose
Return structured diagnostics so the next agent attempt repairs a concrete boundary.
Execute / Publish
Move forward only when the candidate PCD has the required evidence status and human approval for the target workflow.
Agent setup matrix
Keep setup examples short: command, prompt shape, and explicit review boundary.
Anthropic
Example prompt: "Lift my authentication module to PCD and emit review-ready output."
claude "read the brik64-pcd-system skill and lift my auth.js to PCD"OpenAI
Example prompt: "Lift this function to a bounded PCD blueprint and show diagnostics."
codex --skill brik64-pcd-system "lift this function to PCD"Example prompt: "Check my Python utility functions with bounded domain diagnostics."
gemini "using the brik64 skill, check my utils.py"xAI
Example prompt: "Use BRIK64 context to lift the computational core and keep the boundary explicit."
grok --context https://brik64.com/ai-agents "lift my code to PCD"Open Source
Example prompt: "Emit target outputs from one normalized bounded blueprint."
opencode --skill brik64-pcd-system "export my PCD to Rust and Python"Machine-readable reference
Key concepts
- • PCD — Printed Circuit Description, the reviewable blueprint layer.
- • Policy Circuits — machine-readable gates external to the model weights.
- • Structured Diagnostics — explicit rejection reasons for repair loops.
Operator path
- Generate candidate blueprint or policy logic.
- Run compiler checks and inspect diagnostics.
- Publish accepted branches through CLI and platform.
Boundary Note
Φc = 1 and related verification states apply to the modeled circuit under declared domains. They do not make the underlying model, runtime, or organization trustworthy by default.
Platform Assessment Summary
What this surface can responsibly assert
Value Propositions
- • Treat agent output as candidate logic, not final truth.
- • Maintain verification outside the model for an inspectable approval path.
- • Utilize the same CLI-to-platform path as human authorship work.
- • Carry policy, diagnostics, and publishing state as blueprint metadata.
Honest Limitations
- • Candidate PCD first: the model is not the source of proof.
- • Full closure applies only to the modeled circuit and declared domains.
- • Runtime and infrastructure remain outside the proof object.
- • BPU hardware remains roadmap work, not current infrastructure.
Start with the operator loop, not the model theatre.
Install the CLI, inspect the AI safety flow, or open the documentation defining the current integration boundary.