Skip to main content

​
Overview

The Fix Inbox is where every AI-suggested correction lives until you act on it. Whenever ORCA detects a quality issue it can fix — a misclassification, a format violation, a missing value with a confident imputation strategy — it creates a proposed correction step grouped under the file it belongs to. You review them in one place, approve the ones you trust, and apply them in bulk. This is different from one-shot auto-remediation, which generates a fix plan for a single file at the moment you click “Remediate”. The Fix Inbox is a continuous queue across every analyzed file in your workspace.

​
How fixes get into the inbox

Fixes are created automatically by several engines:
EngineTriggersExample
Rule suggesterAfter every job completes”Add a not_null rule on customer_email (3% nulls detected)“
Correction suggesterWhen a quality issue has a confident strategy”Standardize phone format to +CC NNNNNNNNNN”
Anomaly explainerWhen anomaly detection flags a metric drift”Investigate revenue — current value departs sharply from its historical baseline”
Each suggestion has:
  • Confidence score (0–1) — how sure the engine is the fix is correct
  • Source — auto (engine-generated) or user (manually created)
  • Status — proposed, active (approved), rejected, or applied
  • Preview — before/after sample so you can verify the change before applying

​
The trust system

ORCA learns which fixes you typically approve and uses that to set an auto-apply confidence threshold per workspace. High-confidence fixes are one-click approvable, and the highest-confidence fixes are pre-checked when you open the inbox so you can clear the queue in a single click. You can see and adjust the threshold from your Org profile trust history. As you accept or reject suggestions, the trust history updates and the threshold adapts.

​
Working with the inbox

Open Fix Inbox from the sidebar.
1

Review pipelines

Suggestions are grouped by file into pipelines. Each pipeline shows the file name, total proposed steps, and a summary of the change types (null fills, format fixes, deduplication, etc.).
2

Inspect a step

Click any step to see the before/after preview, the engine that produced it, and the confidence score.
3

Approve high-confidence fixes in bulk

Click Approve all high-confidence to one-shot accept every step above your trust threshold across all pipelines. This is the fastest way to clear the queue.
4

Select and act on a subset

Use the checkboxes to multi-select steps. The bulk action bar lets you approve, reject, or apply just the selected ones.
5

Apply

Approving a step marks it active. To actually execute the changes against your file, click Apply on the pipeline. ORCA generates a remediated copy — your original file is never modified.

​
Rejecting and editing

Rejection is just as important as approval — every reject teaches the trust system what you don’t want.
  • Reject removes the step from the queue and records negative feedback.
  • Edit lets you tweak a step’s parameters before approving (e.g. change the imputation value, adjust a regex).
Both actions feed the feedback API so future suggestions become more accurate.

​
API access

Every action in the inbox is available via the REST API: 
# List all pipelines with their steps
GET /api/v1/corrections

# List rule suggestions across files
GET /api/v1/suggestions?min_confidence=0.9

# Approve all high-confidence steps in a pipeline
POST /api/v1/corrections/{pipeline_id}/approve-all-high-confidence

# Apply a pipeline
POST /api/v1/corrections/{pipeline_id}/apply

# Submit feedback on a single suggestion
POST /api/v1/suggestions/{suggestion_id}/feedback

​
Tips

  • Start conservative. Set your auto-apply threshold to the strictest setting for the first week. Watch what gets approved. Then loosen it if you trust the system.
  • Use rejection as teaching. A rejected suggestion isn’t lost work — it’s training data that improves the next round.
  • Pair with contracts. If a fix repeatedly clears the same violation, promote it to a contract rule so the issue never reaches the inbox in the first place.

​
What’s next?