Adyakshar Labs

Adyakshar Labs / Blog

Engineering

99% Is a Failing Grade (When You Run It 10,000 Times)

Would you hire a team that gets it wrong 1% of the time?

It sounds fine. 99% is an A+ in school. Now run it through a real system.

The arithmetic nobody puts on the dashboard

A platform that makes 10,000 decisions a day at 99% accuracy gets 100 of them wrong. Every day. And it gets them wrong quietly, with the same confidence as the 9,900 it got right.

On a dashboard, that is a good number. In payments, logistics or healthcare, each of those 100 can be someone's money, someone's shipment or someone's care.

The percentage is not the problem. Scale is. A small error rate multiplied by a large volume becomes a steady stream of real consequences, and because it is steady, it stops looking like an incident and starts looking like normal operation.

Why averages hide the problem

  • Errors are not evenly spread. A 1% average can be 0.1% for most cases and 20% for one important group. The average tells you nothing about who gets hurt.
  • Silent failures do not alert. A crash is loud. A wrong answer delivered confidently is invisible until someone downstream notices.
  • The cost of a mistake varies. One wrong recommendation might be harmless. Another might be expensive or irreversible. A single rate treats them as equal.

What to do about it

Measure at the level of the decision. Count outcomes, not just uptime. Ask how many decisions were wrong, and how many of those mattered.

Break the error down. Look at error rates by segment, by input type and by consequence. The worst cases are almost never where the average points.

Make uncertainty visible. When a system is not sure, it should say so and route the case to a person or a safer default, rather than guess with confidence.

Test on the rare cases. The difficult 1% rarely appears in a sample of typical inputs. Build tests that go looking for it.

Plan for rollback. Assume some changes will make things worse. Have a way to undo them quickly, decided before the release rather than during it.

Watch the trend. A slow rise from 1% to 1.5% is a 50% increase in mistakes. Alert on drift, not just on thresholds.

A better question

Instead of asking "how accurate is it?", ask "how many wrong decisions does this make a day, who do they affect, and how would we know?"

If you can answer that, you are measuring the right thing.

Work with us

At Adyakshar Labs, we build AI, software and embedded systems for clients with clear communication and honest timelines. If you are building something where the last 1% matters, we would like to hear about it.

Visit adyaksharlabs.com or write to info@adyaksharlabs.com.

  • software engineering
  • reliability
  • machine learning
  • systems design
  • quality

Keep reading

From across the family

More reading from the other Adyakshar Labs products.

Sampad AI
What to Do in the First Two Days After a Tax Notice ArrivesA tax notice rarely gets worse in two days. It gets worse when it sits. A short checklist for capturing the facts, updating the client and planning the reply.
Quantum Labs
Harvest Now, Decrypt Later: How to Decide What to Protect FirstEncrypted data can be recorded today and read later. You cannot protect everything at once, so sort it by shelf life. A plain test and a three-tier list for deciding what comes first.
Bindus
How Long Did That Job Really Take? Using Time Sheets and Milestones HonestlyMost small teams quote from memory and learn the real number after the job is done. A simple habit with time sheets and milestones makes the next quote better than the last.