diff --git a/README.md b/README.md
index 6064f355..477d0126 100644
--- a/README.md
+++ b/README.md
@@ -8,6 +8,74 @@ This repository contains scripts and progresm for Ray CI/CD on buildkite.
 This repository also contains scripts that are used on legacy CI/CD pipelines
 and KubeRay pipelines, in the `ray_ci` directory and `ecosystem_ci` directory.
 
+## Tag Rules and Test Selection
+
+`rayci` decides which pipeline steps to run on a pull request by mapping the
+PR's changed files to a set of *tags*, then selecting the steps whose
+declarations carry those tags. The mapping is driven by the `*.rules.txt` files
+in the Buildkite config directories. The evaluator lives in
+`raycicmd/tag_rule*.go`, with `RunTagAnalysis` (`raycicmd/tag_rule_run.go`) as
+the entry point.
+
+### The algorithm
+
+Tag selection runs only for pull-request builds. On the master branch, a
+`releases/*` branch, any non-PR build, or when `RAYCI_RUN_ALL_TESTS=1` is set,
+`RunTagAnalysis` short-circuits to `["*"]` — every tag, the full pipeline. This
+is why a merge to master runs everything regardless of what changed.
+
+For a PR build, given the changed-file set:
+
+1. Each `*.rules.txt` file loads into its **own** `TagRuleSet`. Multiple rules
+   files are evaluated independently and never concatenated. Ray, for example,
+   keeps `test.rules.txt` and `always.rules.txt` as separate rulesets.
+1. For each changed file, each ruleset returns the tags of its **first matching
+   rule** — first-match-wins, in file order (`TagRuleSet.MatchTags`). A rule
+   with no `@` tags is a skip rule: it matches, ends that ruleset's search for
+   the file, and contributes no tags.
+1. The per-file results union across every ruleset and every changed file into
+   one tag set (`tagsForChangedFiles`).
+
+A file that matches no rule in any ruleset is logged as `unhandled file (no
+matching rule)` and contributes nothing, so a rules file meant to cover
+everything ends with a `*` catch-all.
+
+### Rule file grammar
+
+A `*.rules.txt` file is a sequence of tag declarations followed by rules:
+
+- `! tag ...` declares one or more tags. All declarations must precede the first
+  rule, and every tag a rule emits must be declared or loading the config errors
+  (`ValidateRules`).
+- A token ending in `/` is a directory-prefix match: the path must equal the
+  directory or sit under it.
+- A token with no glob character is an exact file-path match.
+- A token containing `*` or `?` is a glob, compiled to an anchored regex where
+  `*` becomes `.*` and `?` becomes `.`, both crossing `/`. Every other regex
+  metacharacter is escaped to a literal (`globToRegexp`), so a bracket
+  expression such as `[abc]` matches those literal characters, not a character
+  class.
+- `@ tag ...` sets the tags the current rule emits. A rule with no `@` line is a
+  skip rule.
+- `;` ends a rule.
+
+### The `test-rules` conformance check
+
+`rayci test-rules` validates that the rules compute the tags a repo expects. It
+discovers each `*.rules.txt` file and runs its companion `*.rules.test.txt`
+fixture when one exists. The companion name is the rules filename with `.txt`
+replaced by `.test.txt` (`go.rules.txt` is tested by `go.rules.test.txt`); a
+rules file with no companion is not checked. Auto-discovery is overridable with
+`test_rules_files` in config or the `RAYCI_TEST_RULE_FILES` environment
+variable.
+
+Each fixture line is `path: tag1 tag2 ...`, asserting the tag set the evaluator
+computes for that path. The comparison is **exact set equality**, not a subset
+check, so a line must enumerate *every* tag the path emits — asserting a single
+tag is a positive claim that no other tag fires. `test-rules` runs the same
+`tagsForChangedFiles` matcher that live per-PR selection uses, so a green run
+validates the exact algorithm that selects steps.
+
 ## Development Setup
 
 ### Pre-commit Hooks