diff --git a/CLAUDE.md b/CLAUDE.md
index 7157f18..17ea177 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -23,6 +23,7 @@ Non-negotiables:
 - `docs/rules/cross-platform.md`
 - `docs/rules/testing.md`
 - `docs/rules/verification.md`
+- `docs/production-readiness.md` (the bar for shipping this in an IDE; read every conformance iteration)
 
 Confirm rule loading with `rules-loaded`.
 
diff --git a/docs/production-readiness.md b/docs/production-readiness.md
new file mode 100644
index 0000000..d4f3cc1
--- /dev/null
+++ b/docs/production-readiness.md
@@ -0,0 +1,57 @@
+# Production readiness standard (read this every iteration)
+
+PureXML is going into an IDE as the authority on whether a developer's XML and
+XSD is correct. The bar is not "passes our test suite." It is correctness,
+total and characterized, with the false-positive direction held at zero. Read
+this before and after every schema-validity (or any conformance) iteration and
+hold the change to it.
+
+## The four stoppers (each one blocks production on its own)
+
+1. **Rejecting valid input (false positives).** The tool must never reject a
+   correct schema or document. `valid-schemas-rejected` (#148) and
+   `valid-instances-rejected` (#146) must reach 0. In an editing loop the first
+   wrong red mark on work the developer knows is correct destroys trust in every
+   other diagnostic, and they disable the validator. This is the worst-felt
+   failure and the one to guard hardest: under-rejection is recoverable,
+   over-rejection is a non-starter.
+
+2. **Silently accepting faulty input (false negatives).** `invalid-schemas-accepted`
+   (#145) and `invalid-instances-accepted` (#147) must reach 0. The tool exists
+   to catch the developer's mistakes; a missed error behind a green light ships
+   broken work, and a false negative the user cannot see is worse than a crash.
+
+3. **Uncharacterized correctness.** Every count is measured against one finite
+   corpus (the XSTS subset). Real XML and XSD are unbounded. "Passes the suite"
+   is not "is correct." Production needs a differential harness against a trusted
+   reference validator over a large real-world corpus, plus fuzzing for hangs and
+   crashes. And no silent debt: every case the implementation excludes must be
+   named and bounded, never a "documented under-rejection" that emits nothing.
+
+4. **Interactive safety.** An IDE revalidates on keystrokes against possibly
+   hostile input. There must be a proven worst-case time and memory bound (a cap
+   is a band-aid, not a bound), every diagnostic must be located (line and
+   column), and validation must be recoverable (report all errors, not the
+   first).
+
+## Per-iteration checklist (apply on EVERY change)
+
+- Does this change reject any valid input? If yes, that is a false positive:
+  stop and fix before merge. The valid-rejected buckets (#146, #148) may never
+  rise, and their target is 0, not "held at a baseline."
+- Does it leave a silent gap? If it under-rejects, the exclusion must be NAMED
+  and bounded in code and changelog, and tracked as debt to close, never shipped
+  silently.
+- Is it verified against the XSTS differential AND an adversarial over-rejection
+  critic before merge?
+- Are any new diagnostics located (line/col) and recoverable?
+- Is it durable: committed, fast-forward merged, pushed to the remote, and the
+  memory mirror committed and pushed? "Merged locally" is not done.
+
+## Definition of done (the whole goal, not one iteration)
+
+All four buckets at 0 on the FULL XSTS (not the subset); a differential + fuzz
+harness so correctness is bounded, not sampled; every diagnostic located and
+recoverable; a proven resource bound so interactive use cannot hang; and every
+deliverable durable. Until at least false positives are 0, PureXML is a strong
+engine but not an authority to put in front of a developer.
diff --git a/docs/schema-validity-burndown.md b/docs/schema-validity-burndown.md
index 4423bd7..e425bd8 100644
--- a/docs/schema-validity-burndown.md
+++ b/docs/schema-validity-burndown.md
@@ -78,6 +78,11 @@ Each iteration targets one self-contained rule family with a clean root cause.
 
 ## Per-iteration protocol (the PR-critic-loop)
 
+**Read `docs/production-readiness.md` first, every iteration, and hold the change
+to its four stoppers and checklist.** The target for all four XSTS buckets is 0,
+both directions; valid-rejected is not an acceptable baseline. Durability
+(commit, ff-merge, push, mirror) is part of done.
+
 For each cluster, on a fresh branch off `main`:
 
 1. **Scope & measure.** Identify the cluster's cases in