AI Override Suggestions

What Override Suggestions Are

When you review changes in the Safety Gate, you sometimes need to correct a value before it reaches the target system. Maybe a cost center code changed format in the source, or an account mapping needs a manual adjustment for a specific edge case. If you have made the same type of correction before, VersionForge recognizes the pattern and suggests the override for you.

Override suggestions appear as a blue prompt on rows where a known correction pattern applies. You accept the suggestion with one click, or dismiss it and enter your own value.

How Suggestions Are Generated

VersionForge tracks every override you make in the Safety Gate. Over time, it builds a pattern library:

• Field + old value + new value -- "When the cost center field contains CC-410, the reviewer always changes it to CC-4100."
• Conditional patterns -- "When the account is 6010 and the subsidiary is US-West, the reviewer always maps the department to SALES-WEST."
• Format corrections -- "When a date arrives as MM/DD/YYYY, the reviewer always reformats it to YYYY-MM-DD."

When a new sync produces rows that match an existing pattern, VersionForge attaches a suggestion to those rows. The suggestion includes the proposed corrected value and the number of times you have applied this same correction in the past.

Accepting a Suggestion

When you see a blue suggestion prompt on a row:

1. Review the proposed value -- The suggestion shows the current source value and the proposed override side by side.
2. Click Accept -- The override is applied to the row. The field updates to the suggested value and the row is marked as overridden.
3. Continue reviewing -- Accepted suggestions count as overrides in the run log, with a note that the override was AI-suggested.

You can also accept all suggestions for a batch by clicking Apply All Suggestions at the top of the review queue. This is useful when a format change affects dozens or hundreds of rows and you have already verified the pattern is correct.

Dismissing a Suggestion

If a suggestion is wrong or no longer applicable:

1. Click Dismiss on the suggestion prompt.
2. Optionally, enter the correct override value manually.
3. The dismissal is recorded and fed back into the pattern model.

Training the Model

Every accept and every dismiss improves future suggestions:

• Accepts reinforce the pattern. The model's confidence in that correction increases, and it applies the suggestion more proactively in future syncs.
• Dismissals weaken the pattern. After multiple dismissals, the model lowers the confidence score and may stop suggesting that correction entirely.
• New overrides that do not match an existing pattern create a new candidate pattern. After you apply the same correction three times, it becomes an active suggestion pattern.

You can view and manage all active patterns in Settings > AI Features > Override Patterns. Each pattern shows:

| Field | Description | |-------|-------------| | Pattern | The field, condition, and correction | | Times applied | How many times this correction has been accepted | | Confidence | Current model confidence (0-100%) | | Last used | Date of the most recent application |

From this screen, you can delete patterns that are no longer relevant or adjust their scope.

Limitations

• Override suggestions work on field-level corrections only. They do not suggest whether to approve or reject an entire row.
• The model requires at least three identical corrections before it creates a suggestion pattern. One-off overrides do not generate suggestions.
• Suggestions are based on exact or near-exact value matching. The model does not currently generalize across unrelated fields or make inferences about business logic.