What Makes a Correction Useful vs. Noise?

Overview

A correction is most valuable when it tells the system both what the right answer is and why. The outcome alone helps, but a short reason turns a simple label change into a strong training signal.

---

What makes a correction useful

A useful correction includes:

1. Outcome: What the correct score or tier should be.
2. Reason (ideally one sentence): What specific factor the system missed.

This lets the system:

• Learn from the corrected label (the new tier/score), and
• Associate that change with concrete signals (e.g., relationship, engagement, role).

Examples of useful corrections

• Detailed reason:

• “Tier 2 → Tier 1. The contact is the decision-maker’s direct report and has attended two of our events in the last 90 days. That relationship context drives the upgrade.”

• Minimal but still useful reason:

• “Tier 2 → Tier 1. Direct relationship with decision-maker.”

Both are valid. The second is faster to write and still far better than changing the tier with no explanation.

---

Why the reason matters

• With outcome only, the system knows the previous score was wrong and adjusts pattern weights in a general way.
• With outcome + reason, the system can:
  • Tie the correction to specific signals (e.g., role, reporting line, event attendance), and
  • Adjust those particular patterns more precisely.

This makes each correction more informative and accelerates accuracy gains over time.

---

What makes a correction noisy

A correction is noisy when it changes a score without a clear, concrete reason the system can learn from.

Typical noisy pattern:

• You have a vague positive feeling about a record.
• You bump the score or tier up without pointing to a specific signal (e.g., no mention of role, relationship, engagement, or deal context).

The system can still learn from the outcome, but:

• It has no explicit signal to associate with the change.
• Those pattern-less corrections tend to wash out over time instead of compounding.

You should still correct if your instinct says the score is wrong. But whenever possible, add even a partial reason.

Better than nothing: partial reasons

Even a short, partial explanation is valuable, for example:

• “Tier 3 → Tier 2. Strong engagement over last 60 days.”
• “Tier 2 → Tier 1. Direct report to primary decision-maker.”

These give the system something concrete to latch onto.

---

Corrections you should not make

Do not correct records you are genuinely uncertain about.

• Marking an unknown record as Tier 1 (or any tier) teaches the system something you don’t actually believe.
• This introduces label noise, which slows learning and can reduce accuracy.

When you are uncertain:

• Skip the record.
• A skip is not a failure; it is accurate data about the limits of what you know.

Over time, this preserves the quality of the training data and improves the system’s ability to generalize from the corrections you are confident in.

---

Practical guidelines

• Always:
  • Change the tier/score when you believe it is wrong.
  • Add at least a short reason whenever you can.
• Prefer:
  • Specific, observable signals (role, relationship, engagement, deal history) over vague impressions.
• Avoid:
  • Correcting when you are unsure.
  • Vague reasons like “just feels like Tier 1” without any concrete basis.

---

Related: The correction cadence · How long corrections take · How accuracy improves