OBSERVE → PROPOSE → TEST → REVIEW

Self-correction,
with a visible record.

WAKE can prepare a real change to its own presentation code, test it in isolation, and open a draft for review. Follow what ran, what passed, and what is still waiting.

A passing candidate is a proposed source change. A merge, website publication and research-runtime upgrade are separate steps.

SOURCE CORRECTION

Latest runs

Checking the correction workflow…

    Proposed changes

    Checking review branches…

      Open all correction runs → · Review proposed changes →

      MAKE A CORRECTION

      Describe one repair. WAKE prepares the candidate.

      1. Open the correction workflow. Sign in to GitHub with permission to run this repository’s workflows. Choose Run workflow on master.
      2. Describe the problem and expected behavior. In request, explain one concrete repair. In paths, list one to four existing files separated by commas. This first capability covers wake/assets/*.js, *.css, *.html and docs/*.md.
      3. Run it. WAKE sends only that request and those selected source files to Gemini. It validates the returned edits, runs checks in a separate container with no credentials or network, then opens a draft pull request if checks pass.
      4. Review the result. Open the draft’s diff and linked run. Check the behavior and limitations. Merge an acceptable change through GitHub; publish the website through its normal release process. Failed runs keep their logs and any proposal artifact.

      Start a source correction ↗

      Example request and file paths

      request

      Fix the ant-farm zoom controls so a two-finger pinch magnifies the habitat around the midpoint of the gesture. Preserve saved colonies, one-finger panning, and keyboard zoom.

      paths

      wake/assets/ant-farm.js, wake/assets/ant-farm.css

      The repository needs its existing GEMINI_API_KEY Actions secret and GitHub’s setting allowing Actions to create pull requests. Requests, proposal artifacts and review branches are public repository material; do not include secrets. There is one bounded model request per run. This page never receives a credential or submits a privileged command.

      INSTALLED BEHAVIOR ADAPTATION

      What the research installation has recorded

      There is also a native capability that can choose a deadline-priority context policy, evaluate it and adopt or roll it back automatically. It requires an installation grant. The table reads the latest published node snapshots; it does not enable that grant.

      Loading the installation snapshot…

      NodeRecorded grantPolicyTrials / adoptions / rollbacksEvidence

      How native automatic adoption works →

      THE CHANGE BOUNDARY

      What can change, and what cannot

      The source workflow can write a bounded review branch. It cannot edit the kernel, governance, authority records, provider credentials, dependencies, tests, deployment bindings or this workflow. Its evaluator is installed code, outside the model’s edit allowance.

      The native policy lane can change only its installed context-selector setting after its own evidence checks. Neither lane can erase earlier receipts, count its own changes as research progress, change spending limits or grant itself broader permission.

      Review does not require you to copy code by hand: the proposed files are already on the draft branch. Broader runtime self-modification needs a separately versioned capability and compatible runtime release.

      Earlier operational recovery experiment

      The earlier rate-limit/retry example remains available as exact fixture results. It used synthetic services and made zero production changes. Those results are historical test evidence, separate from the correction runs and installation observations above.

      Reproduce that isolated experiment →