The quality baseline

tests/quality-baseline.json freezes what the engine does: on a set of games, the witness the solver finds and the digest of the state after every input, the proof's verdict and state count, every golden save played to its end, the public surface, the number of tests and the first visit's JavaScript. A refactoring keeps it identical; a behaviour change rewrites it on purpose and says so.

What you get

Either ✔ same behaviour as tests/quality-baseline.json for a change meant to keep behaviour, or a rewritten baseline whose every moved line you know and own in the CHANGELOG fragment.

Context and versions

The engine repository. The baseline dates from 4.1.0 "Clarity" (frozen before the engine was split into modules); the "ratcheted" report from 4.1.3; the READMEs' figures from 4.1.8; the README's install section from 4.1.20; heroes in the measured games from 4.4.0.

What it measures, as of 4.4.0 (tools/quality-baseline.ts):

Part	What is frozen	Rule
games (16)	demo, reference, heroes, the five classics, three partial-order fixtures, and the world, cast, scale, picture, mini fixtures: the witness (status, finished, number of inputs, a hash of the digests after every input) and the proof (status, states, softlocks)	identical
saves (43 as of 4.4.0)	each golden save of tests/fixtures/saves/demo-*.json, loaded and played to its end: ended, inputs, last digest	identical
surface.api	a hash of tests/api-surface.json (the public API and the MCP tools)	identical
tests	test files and test declarations (it(, test(, .each, .skipIf, .runIf)	may only grow
bundle.initialJsKB	the first visit's JavaScript, gzipped, from verify:dist (with --dist)	may only shrink

It also keeps two places of the READMEs in step: the figures between <!-- metric:tests -->, metric:initialJsKB and metric:referenceStates markers (3257 declarations, 136 KB and 288 states at 4.4.0), and the archive URL between <!-- install:begin --> and <!-- install:end -->.

Prerequisites

A prepared repository (Developer setup). For --dist, a production build of the sample game in dist/ (npm run build:game).

Where to run the commands

In the engine repository, on your machine. CI's check job runs npm run quality:baseline -- --check --dist on every change.

Steps

1. Compare.

   # in the engine repository
   npm run quality:baseline -- --check
   npm run build:game && npm run quality:baseline -- --check --dist   # with the first visit, as CI

   copy

   It solves and replays every game (quality baseline: solving and replaying…), then prints ✔ same behaviour…, or one ✖ line per difference, for instance demo witness.digests: … → … (a state after some input differs: npm run replay on the witness names the entry), golden save demo-3.4.0.json last: … → …, tests/api-surface.json: the public API or the MCP tools changed, or README.md: a figure is behind the baseline.

2. Decide what a difference is.
   • Your change was meant to keep behaviour (a refactoring, a rename, a speed-up): a difference is a regression. Find it; npm run replay on the witness names the first entry whose state differs.
   • Your change alters behaviour on purpose (a rule, a command, a solver change, a new test, a smaller bundle): go to step 3.

3. Rewrite it, with the first visit if your change can move it:

   # in the engine repository
   npm run quality:baseline
   npm run build:game && npm run quality:baseline -- --dist

   copy

   A rewrite says what it moves, before writing: ⚠ behaviour moved: … for a digest, a verdict, a save or the surface, and ✔ ratcheted: … for more tests or a smaller first visit. It writes tests/quality-baseline.json, then the figures and the install URL of README.md and README.fr.md (✔ README.md: the figures written). Without --dist, the previous bundle is kept.

4. Own it. Commit the JSON and the READMEs with the change; say in the CHANGELOG fragment what moved and why ("the reference chapter's witness changes", "the first visit goes from 125 to 132 KB"). tests/quality-baseline.json is a file the maintainer reviews (CODEOWNERS).

Expected result, and how to check it

npm run quality:baseline -- --check --dist ends with ✔ same behaviour as tests/quality-baseline.json: 16 games, n golden saves, n test declarations, first visit n KB and exit code 0.

Common errors

Symptom	Cause	Fix
README.md: a figure is behind the baseline (npm run quality:baseline writes it)	You added a test (the test count moved) without rewriting	npm run quality:baseline, commit the READMEs
⚠ better than the baseline, ratchet it	More tests, or a smaller first visit: not a failure by itself	Rewrite it so the gain is kept
tests: n declarations → m (a test was removed)	A test was deleted or merged	Restore it, or own the removal in the pull request and rewrite
first visit's JavaScript: n KB → m KB gzipped	The player's first chunk grew	Load the new code on demand (Import boundaries), or own the growth
<game>: new game, not in the baseline	A game was added to the list	Rewrite it
README.md: the install section names another version	package.json's version changed (a release)	Rewrite it: the release branch does

Files or data changed

tests/quality-baseline.json, README.md, README.fr.md. --check writes nothing.

Recovery or going back

git restore tests/quality-baseline.json README.md README.fr.md undoes a rewrite you did not mean.

References

• tools/quality-baseline.ts, tests/quality-baseline.json.
• Before a pull request, The dev loop, Benchmarks.