Guide
What is a canonical AI version of a document?
A canonical AI version of a document is the single, authoritative, machine-readable copy that every AI agent is meant to read, published at a permanent address and checked against the original. Instead of each assistant extracting its own interpretation from a PDF or deck, all of them read the same fidelity-checked text and data.
Why "canonical" matters
When ten people upload the same PDF to ten assistants, you get ten extractions. One reads a chart, another misses it; one attaches a forecast label, another drops it. None of them is wrong in an obvious way, which is exactly the problem.
"Canonical" borrows from the web: one URL is declared the official copy and others point to it. For AI, the canonical version is the copy you, the owner, stand behind, and that agents should prefer over their own guesswork.
What a canonical AI version contains
| Part | Purpose |
|---|---|
| Permanent URL | One address that survives updates, so links never go stale |
| Human-readable page | People and AI see the same content |
| Markdown | Clean text with headings and tables that language models read well |
| Structured JSON | Metrics with period, currency and classification; claims; qualifiers |
| Provenance | Each value points to the page, slide or cell it came from |
| Version and status | Which version this is and whether it passed fidelity check |
Canonical is not the same as converted
Converting a PDF to Markdown is a first step, not the result. A converter can move text faithfully and still lose the thing that matters: that a number was a forecast, that a column was in thousands, that a footnote said "unaudited".
A canonical version has to be checked against the source: every number, currency, date and period matches the source; nothing appears that the source does not contain; qualifiers stay attached; ambiguities the source cannot settle are resolved by the owner, not guessed.
How to tell whether you have one
- Can you send one link that will still be correct after the next update?
- Does every figure carry its period, currency and qualifier in text, not only in layout?
- Can a reader trace a number back to its source page, slide or cell?
- Is there a visible status saying the version was checked?
- Do AI tools read text and data, rather than re-parsing a visual file?
How The Papers creates it
You upload a PDF, PowerPoint, Excel, Word or Markdown file. The Papers extracts content (including chart series, speaker notes, formulas and named ranges), then checks it against the source before publishing. Where the source is ambiguous, for example whether a figure is Revenue or GMV, you confirm, and the answer is recorded as owner-confirmed.
The result is an AI Version at a permanent link, with Markdown at /{room}/{doc}.md and JSON at /{room}/{doc}.json, an AI Readiness score and a status. "Fidelity checked" requires zero unresolved critical issues and cannot be forced. When you update the document, the link stays the same and the version history keeps earlier versions.
FAQ
- Does a canonical AI version replace the original PDF?
- No. People can still read the PDF; the canonical version is what you point AI tools to, so they read checked content.
- What happens when I update the document?
- In The Papers the permanent link stays the same, the new version is checked against the source, and earlier versions remain in the history.
- Why both Markdown and JSON?
- Markdown suits language models reading prose and tables; JSON suits agents and tools that need exact metrics with their period, currency and classification.
- Does fidelity check mean the content is true?
- No. It means the AI version is faithful to your source. Whether the source is correct remains your responsibility.