Skip to content

xxyoudeadpunkxx/canon-boundary-guard

v0.3.2CC-BY-SA-4.0

Canon Boundary Guard provenance posture for ChatGPT, active from the beginning of the conversation. Keeps evidence, discussion, authorized changes, assistant instructions, and model assumptions distinct before reusable output.

GitHub

Canon Boundary Guard — ChatGPT plugin

A personal documentary plugin by XxYouDeaDPunKxX. Version: 0.3.2.

Canon Boundary Guard applies a provenance posture throughout a ChatGPT conversation. It separates inspected evidence, discussion, explicitly authorized changes, instructions to the assistant, and unverified assumptions before they enter reusable answers or artifacts.

Personal import

  1. Download the plugin ZIP — version 0.3.2.
  2. In ChatGPT web, open the plugin creation dialog from the Plugins page and upload the downloaded ZIP.
  3. Complete the add-plugin flow. Activate Canon Boundary Guard through your ChatGPT custom instructions or explicitly select/tag the plugin in the first message of a new conversation.

For custom instructions, use:

Use Canon Boundary Guard from the start of every conversation. Read its complete skill before the first substantive response and apply it throughout the chat.

Once loaded, the complete posture applies before the first substantive response and remains active throughout the chat. No initialization declaration or working file is required.

The skill metadata targets CHAT and permits implicit invocation. The description is not an executable session-start hook: installation alone does not establish that the skill was loaded. Activation through custom instructions or an explicit tag is the supported setup here.

Operating scope

The complete rules are in SKILL.md. Ordinary conversation needs no status announcement. Reusable content receives the applicable provenance, authorization, proof, and label checks.

Explicit approval remains valid within its scope across turns of the current conversation. The plugin does not carry decisions between conversations. Previously supplied material is assessed through its actual contents and available authorization; there is no automatic continuity or recovery process.

Missing evidence blocks the output that depends on it. It does not disable unrelated work. A generated draft or an unchecked saved artifact does not become evidence for its own claims merely because it is read again.

Included files

canon-boundary-guard/
  plugin.json
  LICENSE
  README.md
  skills/canon-boundary-guard/
    SKILL.md
    agents/openai.yaml
    scripts/
      extract_proof.py
      artifact_fingerprint.py

The optional helpers extract textual proof and compute file fingerprints. They use Python's standard library. Python and accessible files are needed to execute them; the posture also supports other mechanical inspection tools. Their implementation need not be read during ordinary conversation. For a file with no text lines, the proof helper reports line_range: null, zero words and empty word lists. Its text output says none (empty file).

Verification

The repository tests cover packaging, resource references, unchanged helper copies, archive contents, and execution of extracted helpers from another working directory. Development commands are in the repository's maintainer notes. These checks do not establish ChatGPT's loading or instruction-following behavior.

To verify the installed plugin, activate it in a new ChatGPT web conversation through custom instructions or an explicit tag. Check that the complete skill is read before substantive output, then exercise source-based reusable output, an authorized change, an unapproved addition, and unavailable evidence. A claim that the skill is active is not evidence that the required reads or checks ran.

Attribution

Adapted from Canon Boundary Guard by XxYouDeaDPunKxX, under CC BY-SA 4.0. The license is included.

Changes dated September 23, 2026: conversation-scoped operation, consolidated instructions, removal of continuity bookkeeping and its implementation, retained proof/hash helpers, and mechanical output checks replacing linguistic decontamination heuristics. The original source distribution remains separate in the repository.

Version 0.3.2 corrects the proof helper's empty-file range. Posture instructions and authorization rules are unchanged.

Format references: plugin packaging, skill loading, and package validation.