AI already reads documents. That is not the question. The question is how long it takes, and which models can still answer after the ingest. PDF is where we ship first. The convention is the card inside the file, not the format.
Every chat, every agent, every pipeline that touches a contract pays the full parse again. Prefill is a wall-clock tax proportional to document size. Token prices can fall 80% in a year. The seconds to chew eight pages do not. Tail latency is the agent’s user experience. None of that improves when tokens get cheaper.
The existing fixes park the understanding somewhere the file leaves behind. Adobe keeps it in the app. RAG keeps it in your database. A sidecar next to the file is gone the first time someone emails the document. Replacement formats ask the world to stop using what they already have. Files do not circulate that way. They are forwarded, archived, named badly, and opened three years later by software that was not in the room when they were written.
open-matter is a small YAML card that travels with the file. Version 0.1 is the PDF profile: a normal attachment named open-matter.yaml. Acrobat has shown attachments since 2001. The pages do not change. DOCX, EPUB, HTML — same envelope, different pocket. A frontier model writes the card once. Every later reader — a frontier API, an 8B open-weight model, a 3B model on a phone — reads a page of hints instead of the whole file.
That last part is the claim that is hardest to copy and easiest to feel if you run models locally. Small models degrade on long context. Needle-in-haystack scores on them range from 22% to 92%. The gap versus closed models widens past 4K tokens. An 8B model drops about twenty points from 8K to 32K. Pre-digested artifacts reverse the gap: a 3B model with structured retrieval can beat a GPT-4o baseline that is still hunting through raw pages. On a phone the constraint is energy. Local inference can drain 6–25% of the battery in under fifteen minutes, and most of that energy is data movement. Every skipped re-parse saves seconds and joules.
Writing the card costs one full read. At one read per document it saves nothing. From the second read on, the arithmetic is simple. Tools that do not look for the card still get a normal file. That is the whole honesty contract: we do not claim every model on earth reads this today. We claim that five lines in an ingestion library is a smaller ask than a new file format, and that the models which need it most are the ones you already run.
A machine-readable convention without consumers is a file nobody opens. So the consumers ship with the spec: a TypeScript reader, a Python reader, an MCP server that exposes read_manifest and write_manifest, and one free app that writes a card onto your file and sits an evaluation written from the pages. The card is exhaust: written on a read you were already doing. The spec stays free.
If you write local or open-weight software, look for the card before you parse. If you issue files, card them yourself. If the card is wrong, ignore it.