How TZ Score Works

Every result, channel, and project in Testzilla carries a TZ Score: a single number from 0 to 100 that tells you, at a glance, how healthy that thing is. This guide explains what the number actually measures so you can trust it.

The score on a single result

When the LLM judge grades a run, it does not just say pass or fail. It scores your test’s criteria and combines two things into the result’s TZ Score:

• Test score — the weighted percentage of your pass/fail criteria that were met. If the judge could not tell whether a criterion was met (an uncertain point), that criterion is left out of the calculation rather than counted against you.
• Quality score — a holistic read of the interaction’s quality across several dimensions.

The result’s TZ Score is a blend of the two, weighted toward the test score. So a per-result score is a graded percentage, not simply 100 for a pass and 0 for a fail.

Verdict and score are separate

The pass / fail / uncertain verdict and the numeric score are two different things. A run can read “FAIL” on a blocker while still showing a high percentage, because the blocker hard-fails the verdict without rewriting the score.

How the score rolls up

Channels and projects show a TZ Score too. It rolls up the hierarchy:

• A test’s standing score comes from its latest run only. Re-running a test replaces its contribution rather than adding to it, so the score always reflects the current state of your agent, not its history. (Running the same test many times in a batch produces its own average/high/low summary, separate from this standing score.)
• A channel’s score is the average of its tests’ scores.
• A project’s score is the average across all of its channels’ tests, weighted by how many tests each channel has — so a channel with two hundred tests moves the project number more than one with two tests.

TZ Score is not the same as pass rate

TZ Score and pass rate are different metrics that live side by side. Pass rate is the simple share of runs that passed (pass / (pass + fail)). TZ Score is the graded percentage described above. Use pass rate for “how often does it work,” and TZ Score for “how well.”

Fail versus not-run

Testzilla is careful to distinguish two things that look similar but mean opposite things:

• Fail — the test ran and scored low. The score ring shows a number (for example 0% or 45%) in red.
• Not run / no data — the test never produced a scorable result (it errored, was never executed, or was cut off). It has no score, and the ring shows a grey --, not a zero.

This matters for roll-ups: a not-run test is excluded from the average entirely rather than dragged down to zero. A channel with one pass and one errored test reads 100%, because only one test was actually scorable, not 50%.

The colour bands

The same thresholds colour every ring, chip, and strip in the app:

Score	Band	Colour
70 and above	Healthy	Green
60 to 69	Watch	Yellow
40 to 59	At risk	Orange
Below 40	Failing	Red
No score	Not run	Grey (--)

The pass/fail threshold sits at 70 — that is the line the app treats as “good.”

Share a TZ Score publicly (badge + embed)

You can turn a channel’s or project’s TZ Score into a public badge — an at-a-glance score you can embed in a README, a status page, or a client-facing site. Open the badge dialog from a channel or project and Testzilla mints an opaque, non-guessable public link and gives you ready-to-paste snippets:

• an SVG image URL for the badge itself,
• a short public share page URL,
• an HTML embed snippet, and
• a Markdown embed snippet.

The badge is derived server-side and shows only the headline score, band, and colour — never raw results, transcripts, or any PII. You can revoke the link at any time from the same dialog, which immediately invalidates the badge everywhere it is embedded.

Next steps

How testing works

See where the score fits in the Project -> Channel -> Test -> Run -> Result model.

Reports

Turn a channel or project’s scores into an AI analysis with issues and fixes.

Reading results

Walk a full Web Chat result: transcript, screenshots, and verdict.