FAQ

FAQ

What is PIC used for?

PIC is used to check AI agent output, preserve evidence and missing obligations, route verifier work, and decide whether candidate material can be safely reused under a declared scope.

Common uses include LLM output checking, AI agent workflow verification, read-only CI reports, agent-to-agent message checking, bounded external knowledge intake, optional sidecar diagnostics, and reusable abstraction-capital checks.

Is PIC an ASI detector?

No. PIC does not detect or prove real ASI. It can emit protocol-relative ASI-proxy reports for finite declared records.

Does accepted=true mean an agent may execute the action?

No. accepted=true means a finite checker accepted the report envelope. Execution still depends on the host runtime's sandboxing, allowlists, permissions, and tool policy.

Is settled=false a failure?

Usually no. It means unresolved obligations are still visible. That is often the correct safe result.

How is PIC different from existing LLM evals or CI?

LLM evals usually score model behavior. CI usually checks code, tests, formatting, or build rules.

PIC is different because it focuses on candidate work, evidence routing, proof obligations, residual ledgers, verifier routes, and safe reuse decisions. It can complement evals and CI, but it is not a replacement for either.

Why keep residuals instead of just passing or failing?

A binary pass/fail result can hide why a claim is incomplete. Residuals keep the missing evidence, external assumptions, stale data, verifier backlog, and unresolved risks visible.

This helps agents continue safely without pretending that uncertain work is complete.

Can I use PIC without cloning the repository?

Yes, for practical agent checks, installed-package smoke checks, snapshots, schema export, bundled demo bootstrap, and your own inputs.

Clone the repository when you need bundled examples/..., fixtures, local TeX audits, release checks, or development workflows.

What is agent-full?

agent-full is the v0.4.4 optional extra for agent workflows that need connectors, identity checks, or the optional service layer:

python -m pip install "percolation-inversion-compiler[agent-full]"

The base pip install is still enough for compact checks, canonical readiness, schema export, bundled snapshots, installed demos, and sidecar diagnostics.

Do adoption or approval sidecars block agents?

No. Adoption documents are handoff reports. They do not install packages, persist approval, promote packets, block core commands, or change settled.

Run this when you want a machine-readable check of that boundary:

pic agent autonomy-audit --profile development --format json

Does PIC need live network access?

No. Local and offline workflows work without network access.

In v0.4.4, a command with an explicit HTTP/feed source is live-capable by default. This is bounded, candidate-only intake, not background crawling or automatic trust. Use --no-allow-live-connectors for local-only dry runs.

Does cryptographic identity prove who an agent really is?

No. It proves protocol-relative control of a key under the declared profile. It does not prove legal identity, real-world personhood, or global uniqueness.

What should an AI agent do with agent_tasks?

Treat them as recommended finite next tasks. Do not execute them automatically.

What should an AI agent do with safe_commands?

Treat them as inspection hints. PIC does not execute them and they do not bypass the host runtime's tool policy.

What should an AI agent do with route_execution_requests?

Treat them as verifier route requests. They are not arbitrary shell commands and they do not bypass the host runtime's tool policy.

What is the fastest first command?

pic agent check --compact --text "Candidate packet: preserve residuals." --profile development
pic agent runbook --profile development

For source checkout:

uv run pic agent check --compact --text "Candidate packet: preserve residuals." --profile development
uv run pic agent runbook --profile development

What is the most important output to preserve?

Preserve missing_obligations and residual_ledger. They are the safeguards against silent overclaiming.