Production Readiness Score (PRS)
The Production Readiness Score is a 0 to 100 number that says how much of a Python project's edge-case surface is covered by verified tests, weighted by how much each test can be trusted. It is not line coverage: only tests that passed the five gates earn points. Use it to track progress over time, as a README badge, or in CI with quell score --json.
What is the formula?
PRS = sum(confidence of WRITTEN tests) / (total edge cases x 100) x 100
Each WRITTEN test contributes its confidence (0 to 100). SCAFFOLDED and FLAGGED items contribute nothing but still count in the total. The result is clamped to 0 to 100.
Two modifiers apply when source files are provided:
| Modifier | Effect |
|---|---|
Every FLAGGED item is justified with a # quell: flagged comment in source | +5 |
A HIGH-confidence test is disabled with @pytest.mark.skip, @pytest.mark.xfail or pytest.skip() | -10 |
The code also defines a -5 penalty per SCAFFOLDED stub left unfinished for more than 14 days, tracked with git blame.
What are the tiers?
| Score | Tier | Label |
|---|---|---|
| 80 and above | green | Production Ready |
| 60 to 79 | yellow | Review Needed |
| below 60 | red | Edge Cases Uncovered |
How is a test's confidence computed?
Confidence has four weighted factors:
| Factor | Weight |
|---|---|
| Spec source quality | 35% |
| Violation specificity | 25% |
| Assertion strength | 25% |
| Coverage uniqueness | 15% |
Spec source quality starts from a base score per source: type constraints (Pydantic Field) 95, docstring Raises 88, guard clauses 80, PySpark schemas 80, survived mutants 70, bug reports 55; an LLM-generated test uses 55. Tiers: HIGH is 85 and above (ship without review), MEDIUM is 60 to 84 (review recommended), LOW is below 60 (review required).
Why can two commands print different scores?
- At the end of
quell find, the summary uses a simple calculation: the number of written tests times a default confidence of 80, over the number of gaps. With nothing written that is 0. quell scorereads a cached bucketedquell-report.jsonif present, and otherwise runs a quick scan that counts every gap the rule engine can write a test for as WRITTEN at confidence 80, without running the gates. It can therefore print a higher number thanquell find.quell cidoes not print the PRS: it prints the fraction of docstring and type requirements that have a covering test, as a percentage. See quell ci.
How do I improve it?
quell find src/ --fix # write verified tests for the gaps it can prove
quell score src/ # see the current score
Functions with explicit Raises: sections, Pydantic models with Field constraints and functions with guard clauses at the top give the most extractable requirements. See writing docstrings and CI integration.
The npm package has its own score, defined as the share of candidates that your tests execute according to coverage data: see the npm CLI reference.