Home

Percolation Inversion Compiler Wiki

Percolation Inversion Compiler, or PIC, is an AI agent output checker and workflow verification toolkit.

In plain terms, PIC helps an agent or engineer turn a piece of text, a repository signal, a message, a feed item, or a proposed capability into a structured report:

• what is being claimed;
• what evidence is attached;
• what is still missing;
• what can be safely reused;
• what should be routed to a verifier;
• what must remain as residual work.

PIC is useful when agent output should be audited before it is reused, routed, promoted, or treated as completed work.

Why Agent Output Is Treated As Candidate Work

An AI agent can write a confident answer before the supporting evidence has been checked. It can also mix correct details with missing context, stale facts, unsafe next steps, or claims that belong to an external domain.

PIC treats agent output as candidate work so the system can ask simple operational questions before reuse:

• What is the claim?
• What evidence supports it?
• What evidence is missing?
• Which verifier route should check it?
• What residual work must stay visible?
• Is the result safe to route, or should it remain diagnostic?

This does not make the agent slower for its own sake. It prevents unverified text from being silently promoted into reusable work.

One Sentence Description

PIC is a workflow verification and AI agent output checking toolkit for preserving evidence, proof obligations, residual ledgers, verifier routes, and safe reuse decisions.

What PIC Is

PIC is a local-first finite checking system. It emits deterministic JSON reports, residual ledgers, proof obligations, salience queues, packet-promotion decisions, identity checks, and collective-phase certificates.

The shortest mental model is:

candidate output
-> structured packet
-> evidence and obligation checks
-> residual ledger
-> next safe routing decision

PIC does not prove that an AI system is generally intelligent. It does not prove real-world truth. It does not execute unsafe actions. It makes incomplete work visible instead of hiding it.

First Pages To Read

• Getting Started: pip install, source checkout, CLI, and Python SDK.
• Quickstart: shortest command path after setup.
• Use Cases: common uses such as LLM output checking, GitHub Actions, OpenClaw, and agent messages.
• Core Concepts: candidate, evidence, proof obligation, residual ledger, verifier route, and reusable abstraction capital.
• Concepts in Plain Language: common terms without heavy theory.
• Agent Guide: how an AI agent should read PIC output.
• What PIC Does Not Do: limits and non-claims.
• Safety and Limits: what PIC does not authorize or prove.
• Theory Map: how ECPT, BIT, TRC, SQOT, and ALT fit together.
• CLI Recipes: practical command patterns.
• Glossary: short definitions.

Common Use Cases

• Check AI-generated text before reusing it.
• Verify AI agent workflows before treating candidate work as completed work.
• Run a read-only GitHub Actions checker that uploads residual-preserving JSON.
• Use the separate OpenClaw / ClawHub skill as a pre-action checklist for external-effect actions.
• Check agent-to-agent messages before using them as packet candidates.
• Preserve unresolved evidence and proof obligations.
• Route missing work to a verifier.
• Keep repository, message, feed, or web intake as candidate packets.
• Separate declared metadata from checked evidence.
• Build residual-preserving reports for CI or agent loops.
• Check signed agent populations and Sybil-resistance ledgers.
• Decide whether an abstraction can become reusable capital.

Search Terms

AI agent output checker, workflow verification, LLM output validation, residual ledger, proof obligations, verifier routing, evidence routing, safe reuse checks, packet promotion, salience queue, abstraction liquidity, agent identity checks, Sybil resistance, PIC, ECPT, BIT, TRC, SQOT, ALT.

What To Trust

Trust the structured report more than the agent text that produced it.

Read these fields first:

• accepted: the finite envelope was accepted by the checker.
• operationally_usable: the result can be used for routing under the selected profile.
• settled: all scoped finite obligations are discharged. This is often false.
• missing_obligations: work that remains unresolved.
• residual_ledger: debt or limits that must not be erased.
• agent_tasks: recommended next finite tasks.
• route_execution_requests: verifier route requests, not permission to execute arbitrary commands.

settled=false is not a command failure. It means the remaining obligations are explicit.

Main Repository

• Repository: https://github.com/kadubon/percolation-inversion-compiler
• README: https://github.com/kadubon/percolation-inversion-compiler#readme
• Agent guide: https://github.com/kadubon/percolation-inversion-compiler/blob/main/docs/for-agents.md
• CLI reference: https://github.com/kadubon/percolation-inversion-compiler/blob/main/docs/cli-reference.md
• Safety boundary: https://github.com/kadubon/percolation-inversion-compiler/blob/main/docs/11-safety-boundary.md