Skip to main content

Answer quality

Trust comes from evidence, not a confident answer. SixDegree measures whether the numbers in an answer can be traced to records your team can inspect.

Answer quality has two views because they answer different questions.

Live grounding​

The top of the page, and the one that matters.

Every answer is assessed on one question: do its numeric claims trace back to records that were retrieved? The measure is not whether the answer sounds right or matches an expected result. It is whether the numbers came from somewhere identifiable.

Two properties make this worth reading:

It needs no answer key. The check is reference-free, so it runs on real questions nobody wrote in advance.

It covers everything. Not a sample, not a nightly subset. Every answer, which means the rate on this page is the actual rate rather than an estimate from a test set.

The trend line shows the claim-grounded rate over time. Its scale starts at 0.6 rather than 0 so ordinary day-to-day movement near the top stays legible. A genuinely bad day still drops to the floor and looks like one.

A dip is worth investigating the same day. It can mean a source stopped syncing and answers are relying on incomplete information. That is the failure to catch before someone quotes a number in a meeting.

The release-gate suite​

Below, collapsed, because it's the checklist rather than the headline.

A curated set of golden questions with known-good answers, run as a point-in-time gate. It's how you check that a change didn't break something that used to work.

It is useful but narrow. It only covers questions someone thought to write down. Live grounding covers the questions people actually ask, which is why it comes first.

Voting on answers​

Every answer takes a thumbs up or down. A downvote asks why, and the reason is the useful part:

  • Wrong facts: the numbers or names are incorrect
  • Out of date: right once, not now
  • Did not answer: responded without addressing the question
  • Unsupported claim: cited or asserted something with no basis
  • Hard to read: right, but badly presented
  • Something else

The reasons separate problems with different fixes. "Out of date" points at a source that stopped syncing. "Made something up" points at grounding. "Hard to read" points at presentation. An undifferentiated pile of downvotes tells you people are unhappy; these tell you what to go fix.

Downvote reasons are summarized on this page, so patterns surface without anyone reading every vote.

Empty is not broken​

If the page can't load its data, it says so rather than showing zeros.

This distinction matters on a page about proof. "No runs yet" and "we could not find out" are different states. A clean zero should never hide a failure to retrieve the result.

Using it​

Read the trend weekly, not daily. Individual answers vary; the line is the signal.

Chase dips the same day. The usual cause is a source that went quiet.

Read the downvote reasons before the counts. Six "hard to read" votes and six "made something up" votes are the same number and completely different problems.

Treat the gate suite as a regression check, not a quality measure. It tells you nothing broke. Live grounding tells you how you're doing.

See Trust and control for the customer-facing explanation of the evidence behind an answer.