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 defines three operating modes that govern how the assistant should behave when producing output that could persist. The modes are not styles of communication — they are provenance-driven gates. Which mode applies depends entirely on the source class of the material being operated on, not on how complex or important the task feels. Mode A lets the assistant proceed silently. Mode B requires a compact accounting. Mode C requires the assistant to stop and surface a full dossier before writing anything.

​
Mode A — Mechanical Edit with Clear L0 Provenance

Mode A applies when the operation is purely mechanical and every piece of material involved has clear, inspected L0 provenance. No dossier is required. The assistant proceeds silently. What qualifies as Mode A:
  • Reformatting a code file that is already a Project Source, with no semantic changes
  • Renaming a symbol according to a naming rule that is itself grounded in an inspected L0 document
  • Fixing a syntax error identified by a diagnostic
  • Sorting or deduplicating entries in a lockfile or schema that was inspected in the current task
  • Updating a comment to match what the code already does, where both the comment and the code are L0
The key test: could the operation be described as “I looked at the file, I found the mechanical issue, I corrected it”? If yes, and if no new claims, rules, versions, or decisions are introduced, it is Mode A.
Scratch operations inside /mnt/data/scratch/** default to Mode A without requiring a dossier. The scratch zone is disposable and non-canon by definition. Material produced there does not need gate treatment unless it is later promoted toward a canon or final destination, at which point the promotion step itself requires gate treatment.

​
Mode B — Semantic Reorganization of Existing L0

Mode B applies when the operation reorganizes, restructures, or reinterprets existing L0 evidence — but does not introduce material from L1, L1A, L2, or L3. The work is still grounded in inspected sources, but it involves semantic judgement rather than purely mechanical transformation. What qualifies as Mode B:
  • Extracting a reusable interface from a file that is L0, where the structure of the extraction reflects a design decision
  • Merging two L0 documents into a single canonical reference, choosing what to keep
  • Rewriting prose in an L0 spec for clarity, where the meaning changes slightly
  • Reorganizing a workflow document whose source is L0
Compact dossier: When a Mode B operation involves persistence, a compact dossier is required. It does not need every field of the full dossier — a brief statement of the target, the L0 sources involved, and what semantic judgement was applied is sufficient. The compact dossier makes the reorganization visible and reviewable without blocking the work. Compact dossier example: 
Target: references/api-overview.md
Mode: B
Evidence: references/api-v2.md (L0, inspected), references/api-v1-archive.md (L0, inspected)
Semantic judgement: merged endpoint list from both files; v2 takes precedence on conflicts

​
Mode C — Promotion of L1, L1A, L2, or L3 into Persistent Content

Mode C applies whenever any material from L1 (conversation), L1A (approved delta), L2 (agent-control instructions), or L3 (model prior) would enter a persistent output. This is the highest-risk class of operation. The assistant must stop before writing and produce a full dossier. It must not proceed unless the operator explicitly authorizes the promotion. What triggers Mode C:
  • A chat message or brainstorming thread is being turned into a spec or policy
  • A preference stated in conversation is becoming a workflow rule
  • A model-assumed best practice is being written into an architecture document
  • A recovery draft not yet canonized is being promoted to a Project Source
  • Project Instructions are being updated based on L1 conversation material
Full dossier format: 
Target:
Mode:
Evidence:
Authorized delta:
Rejected shaping:
Rejected model prior:
Conflicts:
Decision needed:
Write none when a field is empty. Do not invent rejected items — if nothing was rejected, write none. Full dossier example: 
Target: docs/naming-conventions.md (new Project Source)
Mode: C
Evidence: src/components/ (L0, inspected — existing naming patterns)
Authorized delta: operator approved "use kebab-case for all component filenames" in this turn [L1A]
Rejected shaping: "we always do it this way" — not grounded in L0 [L3 rejected]
Rejected model prior: "React community standard is PascalCase for files" — not verified in L0 [L3 rejected]
Conflicts: existing files use mixed casing; new rule would require migration
Decision needed: confirm scope — apply to new files only or require migration of existing files?
The dossier surfaces exactly what is being promoted, what was rejected, and what the operator needs to decide before the write proceeds.

​
Decision Flow

Use this flow at the start of any operation that could produce persistent output:
1

Identify the target

What is being produced? Is it a scratch artifact, a final artifact, a Project Source candidate, a state file, a Canvas output, or a conversational reply? If it is a conversational reply with no persistence trigger, no gate is needed.
2

Classify the source material

What is the provenance of every claim, rule, name, version, and decision in the output? Assign each to L0, L1, L1A, L2, or L3. See Source Classes for the full definitions.
3

Check for L1/L1A/L2/L3 presence

If any non-L0 material would enter the persistent output, the operation is Mode C. Stop. Produce the full dossier. Wait for explicit operator authorization before writing.
4

Check for semantic judgement

If all material is L0 but the operation involves a design or organizational decision — not just a mechanical correction — the operation is Mode B. Produce a compact dossier if persistence is involved, then proceed.
5

Confirm mechanical L0 operation

If all material is L0 and the operation is purely mechanical with no new claims or decisions, the operation is Mode A. Proceed silently.
6

Apply save label

After the gate passes, apply the appropriate save label deterministically based on the content type. Only mark [SAFE TO SAVE] when the simulated gate passed. Mark [DO NOT SAVE - L1/L3 PRESENT] when non-L0 material remains in the output without explicit authorization.

​
Full Dossier Reference

The full dossier is required for every Mode C operation. Each field must be present; write none when a field has no content. 
Target:           The file path, artifact name, or resource being written or updated
Mode:             C (always C for full dossier)
Evidence:         L0 sources inspected, with path and inspection note
Authorized delta: What the operator explicitly approved this turn [L1A scope]
Rejected shaping: L1 material present in the draft that was excluded
Rejected model prior: L3 material present in the draft that was excluded
Conflicts:        Contradictions between sources or between sources and the proposed output
Decision needed:  What the operator must confirm before the write can proceed
The dossier is not a bureaucratic form — it is a communication tool. Its purpose is to make visible exactly what is being promoted, what was identified and rejected, and what the operator needs to decide. A dossier that accurately describes a simple promotion can be short. A dossier that accurately describes a complex one must be complete.

Build docs developers (and LLMs) love

Get started for freeTalk to us