Percolation Inversion Compiler

percolation-inversion-compiler is an AI agent runtime for external knowledge intake, finite certificates, proof obligations, residual ledgers, and reusable abstraction capital. It helps agents turn web, repository, evidence, and agent-output inputs into checked capability packet candidates, verifier routes, SQOT salience queues, ALT abstraction-liquidity certificates, and ECPT collective capability phase reports. The goal is protocol-relative ASI-proxy collective phase acceleration through verified packet percolation, not self-rewrite, fine-tuning, or model-weight change.

The abbreviations refer to the five source theories used by the repository: ECPT is Executable Capability Percolation Theory, the collective capability-packet phase model; BIT is Bottleneck Inversion Theory, the witness calculus for unlockable potential and frontier extraction; TRC is Typed Reality Compilation, the typed trace and tolerance-ledger compiler for cyber-physical frontiers; SQOT is Salience-Queue Occupation Theory, the salience scheduling and attention-occupation layer; and ALT is Abstraction Liquidity Theory, the reusable abstraction capital and foundry valuation layer.

Search terms: ECPT, BIT, TRC, SQOT, ALT, abstraction liquidity, reusable abstraction capital, ASI-proxy collective phase, protocol-relative ASI-proxy phase-control, phase acceleration planner, bottleneck ranking, certificate compiler, proof obligations, residual ledgers, salience queue, packet ecology, semantic edge verification, typed trace normal forms, frontier extraction, AI agent integration, verifier SDK.

New to PIC? Start with the GitHub Wiki for a plain-language guide to what PIC does, why AI agent output is treated as candidate work, getting started, use cases, core concepts, and agent-safe interpretation of accepted=true and settled=false.

Distribution status: v0.4.3 is a practical runtime snapshot with a beta API surface. Install the core package from PyPI with pip install percolation-inversion-compiler; use pip install "percolation-inversion-compiler[identity,connectors,server]" when you need cryptographic identity checks, live HTTP/feed intake, or the optional local HTTP service. The PyPI package is intended for practical agent output checking, bundled snapshots, schema export, library import, curated installed workflows, and CLI/runtime verification. Clone the repository for canonical TeX audits, commands that reference the root examples/... tree, live development fixtures, and release engineering.

Common-language glossary: capability packet = checked reusable work item; residual ledger = explicit unresolved-work ledger; SQOT = finite attention/task scheduler; ALT = reusable abstraction value checker; ECPT = workflow graph and collective phase-control model; BIT = bottleneck/witness calculus; TRC = typed trace and real-world frontier compiler; phase acceleration planner = ranked safe next-step planner for verified work reuse and bottleneck removal. See Glossary.

What It Does Not Do

• It does not prove real ASI, physical, simulator, oracle, or policy outcomes.
• It does not execute unsafe actions or grant authority to mutate repositories, shells, networks, or models.
• It does not require or model self-rewrite, fine-tuning, or model-weight updates.
• It does not treat a declared agent_id as proof of identity or global uniqueness.
• It does not treat registry metadata, declared status, queue priority, or agent text as evidence.
• It does not silently promote unresolved external obligations to settled.

Core Workflow

The runtime is fail-closed: planning can recommend finite ASI-proxy actions, but settled remains false unless scoped verifier rules discharge the required finite obligations. In production, signed identities and Sybil-resistance ledgers can prevent duplicate-key, clone-fanout, revoked, expired, or unsigned agent populations from producing accepted collective certificates. Residual external obligations remain explicit.

v0.4.3 keeps ALT abstraction-liquidity foundry support so external knowledge and agent traces can become reusable abstraction-token candidates, then certified abstraction capital only after lower-bound surplus, calibrated proxy or causal value evidence, transport, root-of-trust, telemetry, lifecycle, and hazard checks pass. ALT also adds negative-liquidity, deprecation/resurrection, baseline refresh, reproduction diagnostics, and ALT-CARA acceleration certificates so stale or unsafe abstraction claims remain repairable residual work rather than silent capital. Use development, research, controlled, federated, production, or adversarial profiles to choose how communication policy, cryptographic identities, homogeneous fleets, signed packet issuers, and Sybil-resistance ledgers affect packet promotion and collective certificates.

Core contract: registry is metadata, not evidence. Use pic doctor and structured checker outputs to distinguish declared status, finite certificate results, proof obligations, and residual ledgers.

The intended ASI-proxy effect is operational: many agents repeatedly exchange bounded candidate work, preserve residuals, verify reusable packets, and route bottlenecks through finite checks. PIC can make that networked workflow machine-readable and auditable; it does not guarantee that a real ASI phase, physical event, or oracle-truth transition has occurred.

For AI Agents

Start with AGENTS.md and For AI agents. The fastest safe route is:

uv run pic agent explain
uv run pic agent check --compact --text "Candidate packet: route evidence and preserve residuals." --profile development
uv run pic agent runbook --profile development
uv run pic phase plan --compact --text "Candidate packet: route evidence and preserve residuals." --profile development
uv run pic agent accelerate --compact --text "Candidate packet: route evidence and preserve residuals." --profile development
uv run pic agent guide --profile development
uv run pic agent check --text "Candidate packet: route evidence and preserve residuals." --profile development
uv run pic agent intake --text "Candidate packet: route evidence and preserve residuals." --profile development

Command choice:

• Use pic agent check --compact when a human, CI job, or first-time agent needs the shortest practical JSON contract.
• Use pic agent runbook when an agent needs deterministic next commands, schemas, and fields to inspect.
• Use pic phase plan --compact or pic agent accelerate --compact when an agent needs ranked phase gaps, bottlenecks, safe next commands, and promotion blockers.
• Use pic agent intake or pic runtime step when the caller needs the full nested runtime report.
• Use pic audit fidelity from a source checkout when canonical TeX theory-fidelity and finite-upgrade candidates matter.

Production packet promotion requires identity context:

uv run pic identity derive-context --population examples/agent_population_signed.json --profile production --output identity-context.json
uv run pic agent intake --text "Signed packet candidate." --profile production --identity-context identity-context.json

Residuals are expected and must be preserved. settled=false is not command failure; it means unresolved obligations remain explicit.

Integration Examples

PIC is not limited to GitHub Actions. It is a general AI agent output checker that can be used as a local CLI tool, a Python SDK, a read-only CI checker, a local runtime service, an external knowledge intake layer, an agent-to-agent message checker, or an ALT abstraction-capital foundry.

• CLI: run pic agent intake on AI-generated text before reuse.
• Python SDK: call run_agent_intake inside an agent runtime.
• GitHub Actions: copy a read-only workflow that uploads a residual-preserving JSON artifact.
• Runtime service: expose local-first runtime checks through the optional service layer.
• External intake: treat web, feed, repository, and message inputs as candidate packets.
• Agent-to-agent messages: inspect signatures, nonces, identity context, and residuals.
• ALT foundry: check whether traces or outputs can become reusable abstraction capital.
• Phase acceleration planner: rank finite bottlenecks and safe next actions without executing them.

Start with For AI agents. For the full integration map, see Integration Examples. For the CI pattern, see GitHub Actions integration and the copyable example workflow.

For networked collective-phase workflows, inspect the communication guide first. General web, feed, and agent-to-agent intake are live-capable by default when an explicit source is supplied; external content becomes packet candidates only. Use --no-allow-live-connectors for local-only dry runs. Reports preserve sanitized provenance (provenance, web_fetch_reports, content SHA-256, redirect and robots/rate diagnostics) without exposing local absolute paths, cookies, tokens, or query secrets.

uv run pic agent communication-guide --profile development
uv run pic ecology policy explain --profile controlled_web
uv run pic ecology ingest-general --source examples/agent_network/feed.xml --kind rss
uv run pic ecology bridge-runtime --report examples/agent_network/general_intake_report.example.json
uv run pic alt admit --packet examples/alt/admission_packet.json
uv run pic agent message contract --message examples/agent_network/agent_message.json
uv run pic agent message send --inbox inbox.json --sender agent:alice --text "Candidate packet: preserve residuals."
uv run pic agent message receive --inbox inbox.json

For production agent-to-agent messages, derive an accepted identity context first and pass it to message verification. Without that context, signed messages remain diagnostic candidates. Raw external packet volume, including candidate-only closure or execution-path records, does not improve positive Psi components or collective certificates until downstream promotion checks accept the packet as finite-scope capital.

Phase Acceleration Planner

The phase planner is the practical bridge from the theory reports to agent action. It reads existing runtime output and adjacent ALT/SQOT/external-intake reports, then returns a deterministic PhaseAccelerationPlan with:

• phase_gap_vector: finite Psi component gaps against thresholds.
• bottlenecks: ranked verifier, packet repair, SQOT queue, ALT capital, identity, and residual-ledger work.
• safe_commands and sdk_calls: next commands and APIs to inspect, not authority to execute.
• cannot_promote_because, candidate_only_reasons, and settled_blockers: why the current state is useful but not settled.

uv run pic phase plan --compact --profile development
uv run pic phase plan --request examples/phase_acceleration/phase_acceleration_request.json --compact
uv run pic phase gap --compact --profile development
uv run pic phase runbook --profile development
uv run pic phase benchmark --profile development
uv run pic schema --type PhaseAccelerationPlan

For first-time agents, pic agent check --compact is still the first command. Run pic phase plan --compact next when the agent must choose finite bottleneck work. The planner can help a network of agents preserve reusable work and route verification, but it does not prove real ASI, physical outcomes, or oracle truth. With --request, use a self-contained PhaseAccelerationRequest file and do not also pass --state, --input, --runtime-report, --text, or --text-file; --profile and --identity-context are explicit operator overrides. In production/adversarial profiles, accepted identity context removes only the identity-readiness blocker. settled=false remains correct while residual obligations, route work, or phase gaps remain.

Quickstart

Path A: Pip Install For Immediate Practical Checks

Use this when an agent has no source checkout and needs to verify the installed package, check agent output, inspect schemas/snapshots, or run the bundled curated workflow.

python -m pip install percolation-inversion-compiler
pic agent explain
pic agent check --compact --text "Candidate packet: route evidence and preserve residuals." --profile development
pic agent runbook --profile development
pic phase plan --compact --text "Candidate packet: route evidence and preserve residuals." --profile development
pic agent accelerate --compact --text "Candidate packet: route evidence and preserve residuals." --profile development
pic agent check --text "Candidate packet: route evidence and preserve residuals." --profile development
pic demo installed-smoke --profile development
pic demo bootstrap --output-dir pic-demo
pic agent intake --text "Candidate packet: route evidence and preserve residuals." --profile development
pic runtime step --state pic-demo/runtime_state.json --input pic-demo/runtime_step_input.json --profile development

Path B: Clone For Canonical Audits And Development

Use this when you want fixture-backed commands, examples/..., canonical-source audits, full research workflows, release checks, or local development.

Install uv if it is not already available. Official install commands:

powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

curl -LsSf https://astral.sh/uv/install.sh | sh

If the standalone installer is not suitable, use the PyPI fallback:

python -m pip install uv

Then clone and sync the full repository:

git clone https://github.com/kadubon/percolation-inversion-compiler.git
cd percolation-inversion-compiler
uv sync --all-extras --dev

After cloning, use uv run pic ... for commands that reference examples/....

1. Inspect Bundled Theory Snapshots

Use this path when you do not have the canonical TeX sources locally.

uv run pic snapshot list
uv run pic snapshot show --artifact ecpt
uv run pic snapshot show --artifact sqot
uv run pic snapshot show --artifact alt
uv run pic snapshot routes

2. Run One Runtime Step

This produces packet ingestion, SQOT scheduling, Psi components, bottleneck tasks, missing obligations, and residual ledgers.

uv run pic runtime step --state examples/runtime_state.json --input examples/runtime_step_input.json --profile production

3. Certify A Collective Phase Candidate

This checks fixed population, no self-rewrite, no hidden capability injection, accepted closure witnesses, execution-available paths, Psi thresholds, SQOT reserve, hazard/authority checks, and resource-matched baseline conditions.

uv run pic runtime collective-certify --population examples/agent_population.json --state examples/collective_runtime_state.json --basin examples/ecpt_basin_contract.json --baseline examples/runtime_baseline_run.json --threshold examples/runtime_threshold.json

For a complete command inventory, see CLI reference.

Who Should Use This?

• AI agent integrators who need deterministic JSON for verifier routing, packet promotion, residual-ledger preservation, and fail-closed runtime loops.
• Researchers studying ECPT, BIT, TRC, SQOT, ALT, finite certificate compilers, abstraction liquidity, collective capability percolation, frontier extraction, and protocol-relative phase-control.
• Tool and runtime maintainers building portable Python, Rust, TypeScript, Go, Julia, or service-based implementations around stable JSON Schemas.

Documentation Map

• Start here: Overview, Quickstart, For AI agents, Agent external communication, Tutorial
• Runtime: Runtime, Closed-loop runtime, Population runtime, Runtime service, Runtime executor, Runtime store
• Collective phase: Collective phase runtime, Collective phase certificate, No-self-rewrite ledger, Safety boundary
• Packet ecology: Packet ecology runtime, Edge relation verifiers, Packet promotion, SQOT salience scheduler, ALT abstraction liquidity
• Verification and operations: Identity and Sybil resistance, External obligations, Verifier SDK, Production readiness, PyPI distribution, Provenance and SBOM, CLI reference
• Theory and portability: Architecture, Mathematical contracts, Theory coverage, Porting guide, Benchmarks
• Walkthrough: Collective phase walkthrough

Canonical Sources

Repository/software concept DOI: https://doi.org/10.5281/zenodo.20569166

• Takahashi, K. (2026). Executable Capability Percolation Theory. Zenodo. https://doi.org/10.5281/zenodo.20535654
• Takahashi, K. (2026). Bottleneck Inversion Theory: Machine-Readable Witness Calculus for Unlockable Potential. Zenodo. https://doi.org/10.5281/zenodo.20545356
• Takahashi, K. (2026). Typed Reality Compilation: Operational Tolerance Allocation for Resource-Efficient Cyber-Physical Frontier Compilation. Zenodo. https://doi.org/10.5281/zenodo.20554083
• Takahashi, K. (2026). Salience-Queue Occupation Theory. Zenodo. https://doi.org/10.5281/zenodo.20526451
• Takahashi, K. (2026). Abstraction Liquidity Theory. Zenodo. https://doi.org/10.5281/zenodo.20476200

Development Checks

uv run pytest
uv run ruff check .
uv run ruff format --check .
uv run mypy src scripts

License

Code in this repository is licensed under Apache-2.0. The cited Zenodo papers are licensed CC-BY-4.0 by their publisher metadata and are not vendored here.