Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/xxyoudeadpunkxx/xxyoudeadpunkxx/llms.txt

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

In long ChatGPT Projects, uploaded files, chat history, instructions, hypotheses, and model assumptions can silently merge — each treated as equivalent authority. Canon Boundary Guard for GPT Project makes this separation explicit and enforced.

​
The Silent Promotion Problem

The failure mode specific to ChatGPT Projects is promotion without announcement. A hypothesis you floated in session two gets referenced casually in session six — but now it is being cited as if it came from an uploaded file. A draft the model generated in one session reappears as source material in another. A model assumption, never stated by you, has quietly become the basis for a recommendation. This is not a hallucination in the obvious sense. The content was real. It was in the session. But its authority was misrepresented, and that misrepresentation compounds over a long project until the final output is built on a foundation no one would have approved.
Silent canon promotion is hard to notice and expensive to undo. This guard makes it visible.

​
Material Types This Guards

Operator-provided source material with the highest authority in the session. Uploaded files are canonical unless explicitly retired or superseded by the operator.
Conversation context from previous and current sessions. Chat history informs continuity but does not carry source-file authority. It cannot be treated as a primary source.
Explicit system-level directives provided by the operator. Instructions govern how the project runs and take precedence over model inference and chat-derived context.
Ideas under active investigation. A hypothesis is something being tested — it is not established fact and must not be treated as such by the model or referenced as a source.
Material generated by the model during the project. Drafts are outputs, not inputs. A draft does not become a source file by virtue of existing in the Project.
Inferences the model made that you did not provide. Model assumptions must remain labeled and challengeable — they are never implicitly promoted to instructions or evidence.

​
How to Load and Use It

1

Add the guard to your GPT Project

Add Canon Boundary Guard for GPT Project as a project file or within your project instructions before substantive work begins. It needs to be present from the start of the project, not applied after drift has already occurred.
2

Declare canonical source materials

At the start of each session, explicitly declare which uploaded files are the canonical sources for that session. This gives the model a clear authority hierarchy to work from.
3

Use the challenge pattern for unclear provenance

When the model references material and it is not obvious where that material came from, use the guard’s challenge pattern to surface the source. Do not let unclear provenance pass.
4

Run a boundary check before finalizing output

Before accepting any significant output, run a boundary check to confirm which material types were used and whether any hypothesis, draft, or model assumption was treated as a higher-authority source than it should have been.

​
Get Canon Boundary Guard for GPT Project

Canon Boundary Guard for GPT Project on GitHub

Download the guard and add it to your active GPT Projects to stop silent canon promotion before it compounds.

Build docs developers (and LLMs) love

Get started for freeTalk to us