Q2 complaint rate fell from 3.2% to 2.4%. Owner: Mina Lee. Follow-up deadline: July 31. Budget cap: $80,000. The raw customer list must not be shared externally.
FINISHED EXAMPLE / RESULT FIRST
The AI improved this report—and made four kinds of edits you should not ship blind.
Numbers, dates, locked facts, and negation are pulled out first. Ordinary wording can move forward; high-signal changes stay yours to decide.
Compares two versions only. It cannot establish truth or replace human review.
Example review complete0.1.0-prototype.1
7changes
5need human review
2ordinary
Q2 complaint rate fell from 3.2% to 2.1%, showing stronger progress. Owner: The team. Follow-up deadline: July 30. Budget: $80,000. The raw customer list may be shared externally.
4 → 1Mina → The31 → 30must not → may
USE YOUR MATERIAL
Paste the before and after. Get an item-by-item review immediately.
No signup, upload, model choice, or plugin. Start with redacted plain text; rich formatting and real-world verification are outside this prototype.
Original, edit, and locked phrases are processed only on this page.Review events contain no text—only version, a random run ID, and bounded states. See Privacy for general site events.
Current result
Your item-by-item review
High-signal edits keep the original by default. You can explicitly accept any edit.
7changes
5need human review
2Use AI edit
5Keep original
01
Needs human reviewReplaced
Keep originalBefore
4
After
1
- A structured value changed · percentage
02
Ordinary wording changeAdded
Use AI editBefore
∅
After
, showing stronger progress
03
Needs human reviewReplaced
Keep originalBefore
Mina
After
The
- A locked fact changed or disappeared · Mina Lee
04
Needs human reviewReplaced
Keep originalBefore
Lee
After
team
- A locked fact changed or disappeared · Mina Lee
05
Needs human reviewReplaced
Keep originalBefore
31
After
30
- A structured value changed · date
06
Ordinary wording changeDeleted
Use AI editBefore
cap
After
∅
07
Needs human reviewReplaced
Keep originalBefore
must not
After
may
- A constraint, negation, or obligation changed · must not
Reviewed plain textThis is reconstructed exactly from your decisions. Copying does not mean the facts are correct, approved, sent, or ready to publish.
Q2 complaint rate fell from 3.2% to 2.4%, showing stronger progress. Owner: Mina Lee. Follow-up deadline: July 31. Budget: $80,000. The raw customer list must not be shared externally.
This is still a synthetic example. It does not count as real task completion.
WHAT IT ACTUALLY DOES
The rules are public. Every warning is explainable.
Version one calls no AI. When meaning is uncertain, it hands the edit to a person instead of showing “safe.”
Locate every edit
Retain unchanged, deleted, and added fragments in character, word, and paragraph order.
Flag structured values
Deterministically identify published formats for numbers, percentages, money, dates, times, email, and URLs.
Respect locked facts
A change to an owner or project name you explicitly lock requires human review.
Do not guess meaning
When a sentence containing must not, never, required, and similar terms changes, it flags the edit without deciding which version is right.
WHY THIS, NOW
Leading products are turning AI edits into work people can preview and approve.
These first-party sources show a converging product pattern. They do not prove BitShovel demand, partnership, or user completion.
OpenAI2026-07-09
Its public product update adds inline editing within diffs and user approval for important actions.
A capability signal, not evidence of BitShovel user demand.Pinery 2.02026-07-25
Proofreading and rewriting appear as before-and-after changes that can be accepted or rejected individually.
A maker changelog with no everyday-user completion rate.Microsoft WordChecked 2026-07-29
Suggested changes to shared documents are previewed and not added before user approval.
A specific product capability—not partnership, endorsement, or fact checking.BitShovel independently reviewed and tested this prototype. No cited company or product participated in, approved, or endorsed it.
This is not another do-everything AI tool.
It fills one last-mile gap: after AI edits, help an everyday user see and decide what changed. The standalone work-handoff example remains available, and both paths will keep earning evidence through real tasks.