How ShardStitch tells a fact from a guess.

A normal handoff blends the repository, old chat history, and a model's assumptions into one blob. ShardStitch separates those sources so the next AI knows which lines came from disk and which lines need checking.

DISKauth/session.ts changed in the current diffVERIFIED
CLAIMPrevious AI said the test suite passedINFERRED
CHECKThe recorded test command is older than the current diffSTALE
NEXTRun the focused auth test before expanding the refactorVERIFY FIRST

Repository facts and session context do not get blended.

Verified material is read directly from current project state such as Git status, changed files, recent commits, dependencies, and recorded checks. Inferred material is recovered intent, recap, decisions, failed paths, and next moves. It can be useful, but it is explicitly best-effort. Stale material conflicts with current evidence or needs a re-check.

The handoff carries the distinction forward.

The continuation packet can be delivered through a tool pickup file, a clipboard handoff, or ShardStitch MCP tools. A receiving tool can use the verified tier as grounded context and re-check inferred or stale lines against the current repository before acting.

The labels inform the next tool; they are not a claim that every external AI client is technically prevented from ignoring the packet.