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.

For Mode B or Mode C persistence, Canon Boundary Guard requires mechanical proof-of-read — concrete evidence that the source was actively inspected during the current task, not recalled from memory or summarized from prior context. A paraphrase alone is never accepted as proof-of-read.

What proof-of-read is

Proof-of-read is a mechanical anchor that demonstrates the assistant actually read a specific section of a specific source file, not that it remembers or can reproduce the general content of that file. It consists of exact verbatim extracts — section headings, opening words, closing words — that could only be produced by direct inspection. This matters because L3 model memory can produce convincing paraphrases of well-known documents without any active retrieval. Without a mechanical anchor, it is impossible to distinguish genuine L0 inspection from model prior masquerading as source evidence. Proof-of-read closes that gap: if the assistant cannot produce the exact first and last five words of the inspected section, the source is treated as unavailable for that write. Proof-of-read applies per source, per section. Inspecting one section of a file does not establish proof-of-read for the entire file.

Textual source requirements

For each textual source section used as L0 evidence in a Mode B or Mode C dossier, provide all of the following:
FieldRequirement
Source identityFile path, Project Source ID, or URL — enough to uniquely identify the source.
Section headingThe exact heading as written in the source, including any leading # characters.
First 5 wordsThe exact first five words of the inspected section body (not the heading).
Last 5 wordsThe exact last five words of the inspected section body.
Location anchorLine numbers, byte ranges, page numbers, chunk identifiers, or file paths — whichever is available.
If the section has fewer than 10 words, quote the entire section rather than providing separate first/last excerpts. Example:
Source: canon-boundary-guard-gpt/references/protocol.md
Section: ## L0 Evidence
First 5 words: Persistent or verified evidence inspected
Last 5 words: and later inspected
Line range: 7-25

Structured or non-text source requirements

For structured sources (JSON schemas, YAML configs, spreadsheets, database exports, binary artifacts), paraphrasing key names is not sufficient. Provide all of the following that are available:
  • Source identity — file path, Project Source ID, URL, or artifact identifier.
  • Exact top-level keys, sheet names, filenames, object identifiers, or schema paths — the structural landmarks that confirm active inspection.
  • Hash — SHA-256 or equivalent if available; record the hash algorithm used.
  • Retrieval failure — if any of the above are unavailable, declare the missing anchor explicitly rather than omitting the field.

Proof failure

If proof-of-read is unavailable for a source — because retrieval failed, the file cannot be located, or the assistant cannot produce the required verbatim anchors — treat the source as unavailable. Block persistence for any output that depended on that source, and enter read-only mode for the target artifact if necessary. Do not substitute a paraphrase, a summary, or model memory to fill the gap.

Using extract_proof.py

The scripts/extract_proof.py utility can mechanically generate proof-of-read records for Markdown files. Given a file path and section heading, it extracts the section heading as written, the first five words of the section body, the last five words, and the line range — exactly the fields required by the textual source proof format. Use extract_proof.py when the source is available as a file in /mnt/data or as an extracted Project Source bundle. Run it before constructing the dossier rather than after, so the proof anchors are available when the dossier fields are filled in. For full usage documentation, see extract_proof.py reference.

Build docs developers (and LLMs) love