Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/XxYouDeaDPunKxX/canon-boundary-guard-for-gpt-project/llms.txt

Use this file to discover all available pages before exploring further.

Canon Boundary Guard for GPT Projects is a source-bundle frame that helps ChatGPT keep project files, chat context, project instructions, working assumptions, and generated drafts separated during long or complex sessions. It is not a native ChatGPT skill — it works by combining a zipped source bundle added to Project Sources with a set of Project Instructions that tell ChatGPT how to classify, inspect, and gate material before anything becomes reusable project content.

The Problem It Solves

In a long ChatGPT Project session, many things can start looking equally authoritative: files already attached to the Project, things said in the current chat, older project memory or moved chats, instructions that tell ChatGPT how to behave, assumptions made by the model, and drafts generated during the session. Without a clear frame, any of these can silently drift into canon. Canon Boundary Guard prevents that by assigning every piece of information to a source class — L0 through L3 — and requiring a classification step before anything crosses the persistence boundary. A chat message, draft, assumption, or generated file should not silently become canon.

High-Level Architecture

The frame has three interlocking parts:
  1. Source bundle — the folder canon-boundary-guard-gpt/ is distributed as a zip and uploaded to Project Sources. It contains the skill definition (SKILL.md), reference documents, JSON schemas, and optional Python validators. At runtime, ChatGPT must locate and inspect the bundle before substantive work begins.
  2. Project Instructions anchor — the contents of PROJECT_CUSTOM_INSTRUCTIONS.txt are pasted into the Project’s instructions field. This tells ChatGPT to run a Status Check bootstrap at the start of every new session, inspecting the bundle before producing any persistent output.
  3. Simulated persistence gate — because GPT Projects do not expose a user-defined pre-write hook, the adapter defines a simulated gate at the semantic persistence boundary. Before any durable output is written, ChatGPT must classify provenance, determine the operating mode (A, B, or C), and — for Mode C operations — produce a full dossier and stop unless the operator explicitly authorizes the promotion.

Key Concepts

Source Classes

Understand the L0–L3 provenance layers that drive every classification decision, from inspected project files down to unverified model assumptions.

Operating Modes

Learn when ChatGPT should proceed silently (Mode A), require a compact dossier (Mode B), or stop and request explicit authorization (Mode C).

Simulated Gate

See how the persistence boundary works in GPT Projects, what triggers the gate, and how save labels are applied deterministically.

Installation

Step-by-step instructions for building the zip bundle, uploading it to Project Sources, and pasting the Project Instructions anchor.

Limits

Canon Boundary Guard reduces silent promotion — it is not a hard guarantee. ChatGPT can still answer incorrectly, skip instructions, lose context, or fail to inspect available files. The frame is a working discipline for GPT Projects. It cannot block every answer, save, file action, or UI action; it only defines what should be recognized as canon. Material saved through the ChatGPT UI without passing the gate may be inspected as evidence of its own contents, but remains L1 or recovery material unless later admitted through the gate.

The Source Bundle Model

The entire frame travels as a single zip — the folder canon-boundary-guard-gpt/ compressed and uploaded to Project Sources. At runtime ChatGPT locates the bundle, extracts it if necessary as a source-staged extraction, and inspects the entrypoint canon-boundary-guard-gpt/SKILL.md before any substantive work. The zip, Project Source record, and any /mnt/data file are not canon by default; they become L0 only for the specific surface actively inspected in the current task. A Project can also be used without the bundle by uploading it manually to an ordinary chat, but continuity is weaker because the instruction anchor lives only inside the conversation.

Build docs developers (and LLMs) love