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.

Before any persistent write, Canon Boundary Guard classifies the operation into one of three modes. The mode determines whether the assistant proceeds silently, produces a compact dossier as a lightweight provenance record, or stops entirely and waits for explicit operator authorization. Selecting the correct mode is not discretionary — it is a mechanical consequence of what sources the output draws on.

​
Mode A — Mechanical L0 operation

Definition: The output draws entirely on inspected L0 evidence. No non-L0 material enters the persistent artifact. The operation is mechanical: extracting, copying, reformatting, or reorganizing content that is already fully inspected, without interpretation or synthesis. Behavior: Proceed silently. No dossier is required. Use cases:
  • Copying a field value from an inspected schema into a generated config file
  • Renaming a variable using a naming convention already defined in an L0 source
  • Extracting a list of endpoints from an inspected OpenAPI file
  • Reformatting inspected JSON to YAML without changing content
Because Mode A output is mechanically derived from verified sources, it carries the lowest gate overhead. The tradeoff is strict: if any non-L0 material enters the artifact — even a single inferred value — the mode must escalate.

​
Mode B — Semantic reorganization of L0

Definition: The operation semantically reorganizes existing L0 evidence — drawing conclusions, rewriting structure, merging sources, or summarizing content — without introducing any L1, L2, or L3 material. The assistant is interpreting and arranging what is already inspected, not importing new assumptions. Behavior: Produce a compact dossier if persistence is involved. The compact dossier documents the target artifact, the selected mode, and the evidence list so provenance can be traced later. Use cases:
  • Summarizing an inspected requirements document into a structured spec
  • Merging two inspected schemas into a unified data model
  • Rewriting an inspected architecture diagram description into a narrative section
Compact dossier format: 
Target: <artifact name or path>
Mode: B
Evidence: <list of inspected L0 sources used>
The compact dossier is intentionally lightweight. It is not a gate stop — it is a provenance receipt that stays attached to the output.

​
Mode C — Promotion of non-L0 material

Definition: The persistent output contains L1, L1A, L2, or L3 material. The assistant is promoting something from conversation, operator approval, agent-control context, or model memory into a durable artifact. Behavior: Produce a full dossier AND stop before writing unless the operator has explicitly authorized the delta in the current turn.
Mode C does not forbid persistence. It requires explicit operator authorization before the assistant proceeds. Once the operator authorizes the specific delta, the assistant may write the artifact and apply the [SAFE TO SAVE] label.
Use cases:
  • Promoting a draft design decision agreed on in chat into a spec document
  • Incorporating a chat-agreed naming convention into a protocol file
  • Persisting a new workflow rule the operator proposed during the session
  • Writing a policy document that includes model-prior best-practice guidance the operator approved
Full dossier format: 
Target: <artifact name or path>
Mode: C
Evidence: <list of inspected L0 sources used>
Authorized delta: <L1A material explicitly approved by operator, or "none">
Rejected shaping: <L1/L2 material excluded from the artifact, or "none">
Rejected model prior: <L3 material excluded from the artifact, or "none">
Conflicts: <any evidence conflicts identified, or "none">
Decision needed: <what the operator must authorize before the assistant proceeds>
Write none when a field is empty. Do not invent rejected items — if nothing was excluded, say none.

​
Mode selection logic

Use the following decision process before every persistent write:
1

Does the output contain any L1, L1A, L2, or L3 material?

If yes → Mode C. Produce the full dossier and stop pending operator authorization.
2

Does the output semantically reorganize L0 evidence?

If yes (and no non-L0 material is present) → Mode B. Produce the compact dossier and proceed.
3

Is the operation purely mechanical L0 extraction or reformatting?

If yes → Mode A. Proceed silently with no dossier.
If there is any ambiguity about whether non-L0 material is present, treat the output as Mode C and surface the question to the operator before writing.

​
Scratch operations

Writes inside /mnt/data/scratch/** are Mode A by default and do not require a dossier. Scratch is a disposable, non-canon zone. Content there is not considered persistent output and does not trigger the gate. However, promotion from /mnt/data/scratch/** to any canon or final destination always triggers the gate, regardless of whether the scratch file was created as Mode A. At promotion time, the assistant must classify provenance fresh and select the appropriate mode for the promoted artifact.

Build docs developers (and LLMs) love

Get started for freeTalk to us