Verified versus inferred: read an AI handoff safely
A verified line describes evidence available in the project. An inferred line describes reconstructed intent or a suggested next step. ShardStitch keeps those separate so the next AI can check a claim instead of inheriting it as fact.
What each label means
Verified: a statement grounded in current files, Git state or recorded command output. Read the evidence and its timing: a saved result describes that run, not every later edit.
Inferred: likely intent, reconstructed reasoning or a next action recovered from available session material or notes. Useful context is still a hypothesis when the repository does not establish it.
Stale: a recorded claim no longer agrees with current evidence, or a result needs to be checked again after changes.
Two worked examples
These examples illustrate the distinction; they are not claims about your repository.
VERIFIED: auth/session.ts appears in the current Git diff.
Evidence: inspect git diff -- auth/session.ts.
INFERRED: the next task may be to finish the session-expiry change.
Evidence limit: recovered intent is not proof the change is complete.The first line can be checked against the current checkout. The second may guide the next question, but needs confirmation against the task and diff.
A completion message is not test evidence
If a prior assistant said the tests passed, keep that statement separate from a recorded command and its output. Check which files and revision the result applies to. A test run from before the latest edit cannot establish that the latest edit passes.
Memory is context, not current task state
Project notes and pickup files can preserve decisions, constraints and useful commands. Files may have changed since those notes were written. Inspect the current diff, changed paths, recent commits and recorded checks before trusting an old recap.
A remembered failed attempt is worth carrying forward so the next session does not repeat it blindly. A proposed next move remains a suggestion, not evidence of an action already performed.
What the receiving tool should check
- Compare the listed changed files and diff with the current checkout.
- Read the recorded command and result, including whether it predates later edits.
- Keep inferred intent and uncertain causes labelled until checked.
- Choose a bounded next action and record the result before calling it complete.
How the distinction survives a handoff
The packet can be delivered through a tool pickup file, clipboard handoff or ShardStitch MCP tools. Available evidence limits what it can establish: unsaved work and uncaptured conversation details remain unknown.
The labels inform the receiving tool. They do not technically prevent an external AI from ignoring them, restore a provider's remote session or prove every generated suggestion correct.
Integrity Engine | Confidence labels | Session recovery | Tool integration paths