Patterns

Feedback & Correction

Patterns for collecting reviewer feedback on agent responses — thumbs rating, inline corrections, numeric scales, behavioral consequences, and feedback history.

Rating

Thumbs Feedback

Quick signal from the reviewer on response quality. Thumbs up confirms accuracy. Thumbs down opens a correction flow for structured remediation.

Controls

The project brief defines 23 requirements across 5 classes. implementation notes summarize the workflow behavior and open dependencies. I've mapped each requirement to its corresponding test case in the review plan.

State	Trigger	Behavior
Neutral	Default	Thumbs visible at half opacity, no selection
Positive	Thumbs up clicked	Brief green flash confirms, selection persists
Negative	Thumbs down clicked	Correction textarea expands below the message
With Correction	Submit button or Cmd/Ctrl+Enter	Textarea pre-filled, submit records the correction

Thumbs appear at reduced opacity until hover, keeping the reading experience clean. A thumbs-down always opens the correction flow — negative signal without context is less useful than a directed correction.

Editing

Inline Correction

Direct editing of agent prose when a specific factual error is identified — typically an incorrect launch plan reference or misquoted launch requirement.

Controls

The launch readiness plan requires dedicated support plan for issue triage procedures. This pattern ensures that the team has an explicit process to track and correct issues reported by users.

State	Visual	Behavior
Original	Error text with subtle highlight + edit button	Clicking the edit button transitions to corrected state
Corrected	Replacement text with green-tinted highlight	Applied tag appears with before → after summary

Inline corrections are scoped to specific factual errors — typically launch plan misreferences or incorrect readiness level claims. The correction is applied in-place so the reviewer sees the fix in context rather than receiving a full regeneration.

Assessment

Rating Scale

Numeric quality rating for agent responses, used to build a calibration dataset over time. More granular than thumbs, less intrusive than a full correction.

Controls

Based on the launch policy, the product must implement Export workflow for CSV and JSON exports and Retention setting for account retention. Both are covered by the ACME export service's internal platform approval.

Rate this response

Rating	Meaning	Agent Response
1	Incorrect or harmful	Flagged for immediate review, excluded from training
2	Mostly wrong	Queued for regeneration with guided corrections
3	Partially correct	Logged as neutral — no behavioral change
4	Good with minor issues	Reinforced with small adjustments noted
5	Excellent	Reinforced as-is, used as positive calibration example

Numbered buttons rather than stars — the scale is intentionally utilitarian. Reviewers are accustomed to readiness labels and respond well to explicit ordinal scales. The brief confirmation message auto-dismisses to avoid interrupting workflow.

Adaptation

Behavioral Consequence

How the agent's behavior visibly changes after receiving feedback. Demonstrates the feedback loop closing — the reviewer sees the before and after side by side.

Controls

The launch readiness plan requires dedicated support plan for issue triage procedures. This pattern ensures that the team has an explicit process to track and correct issues reported by users of the product.

State	Content	Purpose
Before	Original agent response with error	Establishes baseline for comparison
After	Corrected response with annotation	Shows the loop closing — feedback produces visible change

Closing the feedback loop visibly builds trust. The annotation card links back to the original feedback entry so the reviewer can trace exactly which correction produced the behavioral change.

Record

Feedback History

A running log of all reviewer feedback — positive signals, negative signals, and corrections. Provides an audit trail and training reference for the agent.

Controls

Column	Content	Notes
Type	Positive / Negative / Correction	Icon-coded, monochrome
Timestamp	Date and time of feedback	Sorted most recent first
Message	Excerpt of the rated agent response	Truncated to one line, full text on expand
Detail	Reviewer's note or correction text	Visible only when row is expanded

The feedback history doubles as an audit trail — useful for compliance reviews where traceability of reviewer decisions matters. Expandable rows keep the list scannable while preserving full detail on demand.