# Correct and resume: limits and conditions

> Which changes can return to a verified result, which cannot, and what continuity depends on.

Corrections and resuming: limits and conditions.

## Which corrections qualify

- **Defined correction rules.** On the core two-group path, a guarantee is restored by correcting the exact values of planned samples, or by restoring every row from originals. Other workflows recheck corrections under their own rules.
- **Restoration must be registered in advance.** Restoring from originals requires the source materials and the full-restoration rule to be fixed with the design before the first data arrive.
- **Numbers still have to pass.** A qualifying correction still has to pass every numerical check before a new guarantee is issued.

## Changes that do not restore a guarantee

- **Outside the rules.** Arbitrary value changes, partial restoration, originals added or swapped after the fact, and design or method changes after data arrive are paused or refused.
- **Approval is not enough.** A written reason, a repeated run, or a human approval alone cannot restore a guarantee.

## How rechecking and history work

- **Every dependent step.** SciGround rechecks every dependent step rather than optimizing for the minimum, so a recheck can take longer than a single recalculation.
- **History is kept.** Old versions and Records are preserved, not erased. A correction cannot make an earlier result disappear.
- **Withdrawal.** Withdrawal applies to a whole study, not to individual claims.

## What continuity depends on

- **Your AI calling SciGround.** SciGround points each new session to the current version. Whether a given AI follows that guidance is outside the guarantee; see Use with your AI.
- **Role assignment.** Who holds the proposer, approver, and owner roles is your decision. SciGround enforces the roles as assigned.

Back to [Correct and resume](https://www.licklider.ai/sciground/revisions/).