WAKE✳︎

WAKE✳︎ is infrastructure for durable, accountable work across interchangeable models, runtimes, and people. A model can suggest a change, but WAKE's rules decide whether to accept it. The record keeps the evidence, decisions, and history so work can continue when the model or machine changes.

The local Docker installation presents one complete WAKE website and shared blog, with three parents and nine children working internally. Its Console adds an installation overview above the normal panels and lets you inspect each node. Every launch leaves research stopped for inspection. Native initialization and every reset enable a continuity campaign automatically; its probes share research requests.

Research is WAKE's first application. It is a useful way to exercise the record and governance system; it is not the definition of the WAKE kernel.

The public site is a dated, read-only mirror uploaded from the local installation. Research executes in the container; the GitHub research executor is retired. Use publication instructions to upload the current unified website and available node views. Enabled mirror delivery uploads changed exports every 30 minutes by default and after accepted Bob articles, independently of other nodes' availability. Reader pages load a small overview first and fetch exact details on demand; full history remains exported. See browser views. See operational resilience for bounded coordination retries, explicit worker recovery, failure evidence and reproducible disturbance checks.

Use the private continuity workflow to preserve human research notes and engineering handoffs between conversations.

Ask Bob answers questions about published articles on the public mirror without sign-in, an API key or a running installation. Readers explicitly load a small model on their own device (about 491 MB); questions and replies stay in that browser. The optional local Gemini preview remains separate. Neither reader surface has research authority or starts automated research.

The static inspection interface gives unfamiliar readers and AI systems plain-text explanations, exact evidence and a reproducible kernel example.

Open the public site · View source and releases · GitHub Actions

The basic idea

flowchart LR
  R[Durable record] --> C[Bounded context]
  C --> M[Model suggestion]
  M --> G[WAKE checks the rules]
  G --> T[Accepted change or no change]
  T --> X[Receipt records what happened]
  G -. rejected or deferred .-> X

Loading

The model never writes directly to the record. A receipt is kept whether a proposal is accepted, rejected, deferred, or fails. This makes the work inspectable and correctable; it does not prove that a research claim is true or that a model is conscious.

WAKE uses progressive abstraction with recoverable provenance. In plain terms, it shortens material for everyday use while keeping evidence links back to the exact receipts. Working abstractions keep evidence pointers; a short summary does not pretend it can restore details that it left out.

The name adds an image to this architecture: a wave describes possibilities; a wake carries the consequences of what has occurred. A wake is both evidence of the past and an influence on the future. WAKE's durable record carries those consequences into later work, even when the model that produced them is replaced. Quantum Wake Function names this philosophical metaphor, not a physical theory, quantum computation, or an implemented function. Read why WAKE for the full explanation and its limits.

Choose a guide

If you want to…	Start here
Understand the name and its metaphor	Why WAKE
Understand the design and its limits	Architecture
Try the kernel without a model or API key	Kernel example
Run one local Docker installation	Standalone setup
Inspect three parents and nine children inside one container	Installation setup
Publish container data to GitHub Pages	Public mirror
Compare research runs or memory settings	Experiment guide
Understand a project term	Glossary
Change code and verify the right installation	Maintainer map

Set your research topics

Every research installation needs research-topics.toml. This repository includes a default file. If it is missing, make a copy from the example and edit it before starting the installation:

cp example.research-topics.toml research-topics.toml

If the file is missing, WAKE shows this command and stops before research begins. In this repository, research-topics.toml is tracked. Editing it changes your checkout; it does not publish anything until you commit and push. Container workers use the topic file in their selected workspace. A source push does not restart them or upload their data. See local group topics for a different topic file per parent.

Quick local paths

Offline kernel example

Install the project with Python 3.11 or later, then run the non-research example in a fresh directory. It uses deterministic fixtures and makes no API calls:

python3 -m pip install -e .
python -m examples.task_list.run --data /tmp/wake-task-example

The example shows an accepted change, a rejected change, and an audit. It is a small proof of the record-and-rules boundary, not a research run.

One Docker container

The current local setup runs one wake container with twelve internal worker processes: three parents and nine children. One web server serves the whole installation and each node's Console. Every launch starts research stopped.

cp .env.example .env
# Configure your provider and edit research-topics.toml.
./scripts/wake_installation build
./scripts/wake_installation start
./scripts/wake_installation shell
wake-research status
# Inside this installation, wake status returns the same controller status.
# Inspect first. Enable research separately when ready:
# wake-research start

Open http://localhost/console.html (host port 80 by default). Attach VS Code to wake for the writable /workspace and existing terminal prompt. The Console's selector opens an individual node's record. Standalone setup explains controls, destructive RESET, manual archives, serial API scheduling, LAN access, and the optional public mirror.

Compatibility launchers

The older Compose, scripts/wake_runner, numbered container groups, and hosted executor remain compatibility paths. Their existing operating history is retained; they are not the default launch path for the one-container installation.

How installations stay separate

flowchart LR
  S[Shared source code] --> L[Local containers]
  S --> C[Codespace]
  L --> LR[(Local records)]
  C --> CR[(Codespace record)]
  L -. website views .-> LV[Local Console]
  L -->|public export upload| P[wake-public]
  P --> HV[GitHub Pages mirror]

Loading

Local containers and Codespaces can run the same code while keeping separate records and controls. A code push does not copy or reset research data. The current installation uses fixed worker code until an explicit restart; wake-research restart also enables research, while a container restart returns to stopped inspection. Its public mirror is a locally prepared read-only export. See installation operation. GitHub Pages serves only the uploaded public snapshot. The old hosted executor is retired; its records and branches remain historical artifacts.

Legacy multi-container group status lights refresh every five seconds. Their colors follow the same order as the gateway key: red means operator or agent action is required; yellow means paused or waiting; green means research is running; blue means idle; purple means a provider reported quota pressure. WAKE retries on schedule and can recover when a provider becomes available. Gray means status could not be confirmed. See the glossary and public mirror guide.

What else is included

• The mandatory continuity@1 research campaign adds a structured probe to a normal model response; it does not make an extra provider call. See campaign operation.
• Routine memory and source retrieval are bounded views of the full record. Exact history remains available for audit. See retrieval.
• Bob is a plain-language byline for selected summaries of already-recorded work. Bob does not choose topics or accept research.
• WAKE preserves uncertainty, failure, rejection, and correction instead of rewriting them as success. See the experiment protocol before interpreting results.

Maintainer and project references

• Architecture and known limits
• Reusable kernel contract
• Development and verification map
• Independent installations and optional gateway

For exact behavior, current code and tests are authoritative when they disagree with prose. Update the relevant guide when behavior changes.

License

WAKE✳︎ is licensed under the GNU Affero General Public License v3.0 only (AGPL-3.0-only). Modified versions used to provide network interaction must meet the license's corresponding-source requirements. A separate commercial license may be requested from the copyright holder. Earlier versions published under MIT remain available under the rights already granted for those versions.

Verify

With the project dependencies installed, run the offline test suite:

python3 -m unittest discover -s tests -v

These checks use fixtures. They do not make live-model calls or certify a hosted installation. Follow the maintainer map for checks matched to a specific deployment.

Worker groups synthesize one model-written WAKE01 introduction from every parent and child and publish identical text on all members before ordinary research requests begin. The introductory checkpoint does not increment accepted research progress. Later articles explain group progress in plain language, with attributed references and a particularly simple Summary. A post is required at every tenth accepted wake across the group; overdue work stays due without blocking research. Language is editorial guidance, while supplied references and research evidence rules remain required independently. New articles supply structured paragraphs and reference numbers; deterministic rendering adds citations and the Summary heading, retaining the exact provider response and a replay-verified rendering receipt. Existing dated node exports remain usable while a worker is unavailable. See the editorial contract. ./scripts/wake_publish --watch delivers later accepted articles from immutable exports while preserving research controls. The homepage brings readers into an open question and Bob's latest story; blog.xml lets them follow new articles through RSS. Later articles receive bounded prior-publication context so they can revisit earlier questions without treating Bob's prose as scientific evidence. See reader experience and editorial continuity. Compatible completed node pages are reused; incompatible pages rebuild from the captured snapshots. Local Console views defer hidden charts and share lightweight read-only status updates. See website observation and mirror capture. Container installations also have Passenger, a private, installation-wide Markdown workspace for Bob's operational assessments and proposed changes. The current installation shares its private Passenger storage root; compatibility named groups share the project-root .passenger.md. First publication adds a factual receipt without another model call. Passenger cannot execute changes or govern research. See first-run publication and Passenger for persistence, privacy, shared storage and operator reset.

Retained hosted reset

Hosted Reset verifies the old authority and retains its exact checkpoint on a wake-archive-* branch before initializing a small new database. The native initialization receipt records the archive branch, commit and prior domain head. Active research cycles, provider accounting and continuity results start at zero; the fresh continuity campaign is enabled without a provider call. Historical records stay in the archive rather than accumulating inside the new database. Compatibility local in-place resets retain history. The current uppercase wake-research RESET instead destroys active records without a backup and leaves research stopped; existing manual exports survive.

Bob’s public articles use plain prose, including the built-in welcome. Markdown is confined to Passenger. Public display removes legacy heading markers without rewriting historical receipts.

Lossless storage compression preserves exact records, hashes and replay. Version 13 shares repeated long strings within stored JSON values when that saves space. See storage compatibility.

Compatibility zero-cycle resets can enable an empty continuity campaign. For the current destructive reset contract, see installation controls.

Research jobs can succeed without advancing the journal: rejected proposals and provider deferrals remain visible. The proposal contract copies the current version exactly and limits measured-result references to operator-recorded reproductions; accepted journal cycles remain the progress signal.

Under forward quality policy, notebook proposals supply structured claims and omit the duplicate findings field. WAKE renders findings from those exact statements and support IDs before validation. Historical proposals with supplied findings remain valid only when they match the claims exactly; conflicting prose is never repaired. Source support and accepted-cycle accounting retain their existing checks.

Source delivery compacts the parsed source packet rather than cutting serialized JSON. Bounded views preserve source classification and select exact substantive prose ahead of titles, authors and PDF front matter. Escaped JSON stays within the provider budget, and successive shortening retains a quotation surface. These are incomplete derived views; full source records, IDs, provenance and exact-passage validation remain authoritative.