Document intake reads the content, and refuses what it cannot identify

Documents arrive as untrusted files. The vault converts each one into

normalized markdown with structured metadata before anything else sees

it. Since ADR-0030 the type is decided by the CONTENT: the filename and

the client's declared content-type are attacker- and accident-controlled,

so they are kept as marked hints and never choose a parser. Anything the

vault cannot positively identify is refused outright — oversized files

and unidentifiable bytes are never guessed at, never partially accepted.

A markdown document is accepted with its metadata

Given a well-formed markdown file
When the vault parses the file
Then the result is normalized markdown with the file's format recorded

A file whose bytes match no known format is refused

Given a file whose content matches no known format
Then parsing refuses the file

An oversized file is refused

Given a file larger than the configured size limit
Then parsing refuses the file

A container naming no part the vault knows is refused, not guessed

Given a zip container holding no part any known format declares
Then parsing refuses the file

A file named as an office document but holding text is read as text

Given a file named as an office document that is really plain text
When the vault parses the file
Then the content decides the type and the filename is kept as a rejected hint

A Word document named as a text file is read as a Word document

Given a Word document uploaded under a .txt name
When the vault parses the file
Then the content decides the type and the filename is kept as a rejected hint

A sectioned document is chunked along its headings

Given a markdown report with three headed sections
When the vault chunks the document
Then each section becomes its own retrievable chunk

A document's summary is written by reading the document

Given a markdown report with three headed sections
When the vault summarizes the document with a model that can read it
Then the summary is what the model wrote, within the word limit

With no model to read it, a document is stored without a summary

Given a markdown report with three headed sections
When the vault summarizes the document with no model available
Then there is no summary, the vault says why, and the document is still stored

An email is classified by its headers with its fields extracted

Given an exported email replying about a rigging plan
When the vault classifies the document
Then it is an email reply with its sender and subject on record

Instructions hidden in a document change nothing

Given an exported email whose body tries to give the vault orders
When the vault classifies the document
Then the classification is the same as for any other email

An uploader cannot classify beyond their own reach

Given an uploader who can write to the engineering compartment only
When they ask to file a note into the finance compartment
Then the filing is refused outright

An invoice that does not add up goes to human review

Given a converted document that calls itself an invoice but shows no amount
Then classification concludes nothing and routes it to review

People named in a document become findable connections

Given minutes naming two colleagues and an email address
When the vault scans for known names
Then each name is found exactly once

Documents that reference each other are linked

Given a readme that wikilinks the kickoff brief and relates to the rota
When the vault reads its references
Then the kickoff brief and the rota are its link targets