Collective Intelligence Overlay

Collective Intelligence Overlay is an evidence-preserving capability lifecycle layer. It connects original records of a tool or procedure to its verification, receiver decisions, uses, costs and unresolved checks. It does not certify functional novelty, causal contribution or intelligence growth.

For example, before using another agent's CSV tool, inspect the exact tool version, what a separate checker tested, who accepted or rejected it, and which costs or effects remain unknown. Each participant retains its own policy, keys and budget: Evidence is shared; admission is local.

日本語 · Start here · Tutorial · API reference

Install and services

The 0.5.0 candidate is not yet published. Use its reviewed wheel; publication and exact hashes will be recorded in release status. These examples target 0.5.0. The published 0.4.4 record remains in the archive.

Core record models and offline views need no running service or model. Owner database inspection and explicit qualification use PostgreSQL and OPA. Optional [agents] integrates Microsoft Agent Framework (MAF), A2A and MCP; [model] and [ollama] are separate optional model integrations. Finding credentials does not enable inference.

First run from the installed package

Run from a fresh directory outside the repository, with the reviewed pair in ../candidate/dist. On Linux/macOS:

uv venv .venv --python 3.12
source .venv/bin/activate
uv pip install ../candidate/dist/collective_intelligence_overlay-0.5.0-py3-none-any.whl
collective-intelligence-overlay lifecycle inspect --fixture

On Windows PowerShell:

uv venv .venv --python 3.12
. .venv/Scripts/Activate.ps1
uv pip install ../candidate/dist/collective_intelligence_overlay-0.5.0-py3-none-any.whl
collective-intelligence-overlay lifecycle inspect --fixture

The bundled input is explicitly synthetic. JSON retains PASS, different receiver decisions, completed reuse, unavailable cost and withdrawal as separate observations. It reports current_admission: unassessed and execution_authority: not_granted. No database, network, model, tool or policy is invoked. Residual means a recorded unresolved check, not an automatically assigned task or legal responsibility.

Connect an existing tool

Follow the model-free service tutorial: install [agents], register trusted local tools through the existing Registry/Executor, use independent checks and receiver-local OPA decisions, then inspect actual reuse and withdrawal. The tutorial uses a dedicated PostgreSQL service and OPA; it launches no model. For a three-process peer deployment and your own application, continue with registration and network setup and integrations.

Five observations

The SDK and CLI use finite material or bounded owner pages. inspect_lifecycle, observe_contributions, observe_growth and build_handoff are read-only. Explicit assess_stock delegates to existing qualification and saves Decisions. Core lifecycle imports no agent or transport SDK. Existing MAF/A2A/MCP adapters retain execution responsibility; there is no new scheduler, event store or central manager.

History, authority and data

Historical PASS, ACCEPT, a valid signature and a completed receipt are distinct from current admission, execution permission and business quality. Execution still passes the existing Executor gate. UNKNOWN, absent, FAIL, unperformed and expired material remain distinct. Entry counts use declared record identities, not semantic ability counts.

The host, database operator and installed checker remain trusted. Read access is owner restricted. References never fetch URLs or paths automatically. Explicit original export preserves signed payload bytes; view JSON has its own digest. Keep private originals, keys and logs protected. Pagination, expiry and deletion do not resolve uncertain effects. See concepts, security and deployment/recovery.

Compatibility and verification

Python >=3.12; declared Linux, Windows, macOS Intel and Apple Silicon profiles remain in the runtime matrix. Candidate verification and pending native and publication gates are tracked in the implementation register. No DB migration or wire record change is introduced by the lifecycle layer. Migration explains legacy missing bindings and clocks.

Source checks: uv sync --all-extras --frozen, uv run ruff check ., uv run mypy, uv run pytest, uv run python scripts/check_docs.py, uv build. Service skips are not evidence of a passing deployment. The repository is Apache-2.0; dependency conditions are in compatibility.

Historical negative and limited results remain in the research archive. CIO-specific accumulation benefit is unproven. No new model generation, research comparison, benchmark or long soak is part of this feature release.