Full documentation
A system you trust with your work should be one you can inspect, down to the last line.
The code and its README are on GitHub. Below is a map of it: every file, class and function in G6 and how they connect, 2,588 parts in 10 layers, with a 15-step tour. Pan, search, and click any part to read what it does.
Machine-generated, not reviewed. An AI model read the code at
commit 14de30a6e8 (19 October 2026, 445 files) and wrote every description here and in the
viewer. The code is the authority; the map is regenerated before each release.
The viewer is the dashboard of Understand-Anything, built unchanged, under the MIT License.
The 10 layers
| Layer | Files | What it holds |
|---|---|---|
| Kernel | 44 | The g6.core kernel that every component depends on (the three reliability labels, the reliability envelope, block contracts, the component registry, approvals, the hash-chained ledger, storage, the emergency stop and the sandbox), plus the package root and the standalone file-manifest integrity check and Python error triage helpers that import no component. |
| Evidence, Claims and Reasoning Components | 52 | Components that ground claims in sources, keep belief revision honest, issue evidence certificates and check result envelopes, together with the reasoning, debate, formal proof, counterexample-guided synthesis and abstraction components that test what a claim really shows. |
| Action and Runtime Safety Components | 50 | Components that decide whether an agent action may run: action gating and its action logic, the safety engine, advisory prompt and source-code guards, the goal engine with its plan tree and controls, and safety bounds on agent models (reachability bounds, finite-trace temporal logic and rollback feasibility). |
| Coding Agent and Workforce Components | 56 | The coding-agent gate and the workforce component that runs a local coding mission under a frozen standard, with its role templates (doer, prober, puller, verifier, gatekeeper) and test-run parameters. |
| Task, Context and Governed Learning Components | 64 | Components that turn intent into plans and solved tasks (intent engine, reliability pipeline, solver, navigator, context) and the governed learning components (learning, self-training, distillation) that change behaviour only through checked promotion. |
| Composition, Surfaces and Apps | 26 | g6.harness, the one place components are composed into an end-to-end run, and the ways people reach it: the g6-mcp server, the g6 command line, the eleven single-problem bundles and the operator loop with its approval queue and budget stops. |
| Metaharness | 25 | The layer on top that tunes G6 to a recurring task by composing many components, proposing changes that a person approves or rolls back; only the MCP surface imports it. |
| Off-by-Default Research Code | 36 | Research code that stays off unless switched on, including evidence-constrained and governed harness adaptation behind G6_ECHA_RESEARCH=1 and the mission bridge behind G6_MISSION_BRIDGE=1 with its replay fixtures. |
| Documentation Website | 53 | The Django documentation site in web/ that is built to static files: settings and URL routing, page views and the static build command, HTML templates, Markdown content pages and the Tailwind stylesheet build. |
| Validation Studies, Examples and Project Support | 36 | Scripts and protocols of the pre-registered validation studies, runnable demos, maintainer scripts for export, catalogue generation and bundle packaging, and the root README, agent guide, notice and pyproject.toml. |