diff --git a/docs/architecture/rtl-audit-evidence.md b/docs/architecture/rtl-audit-evidence.md
new file mode 100644
index 0000000000000..e0fe91223249d
--- /dev/null
+++ b/docs/architecture/rtl-audit-evidence.md
@@ -0,0 +1,224 @@
+---
+schema_version: 1
+template_version: 1
+kind: architecture
+id: architecture:rtl-audit-evidence
+authority: draft
+archive_reason: null
+superseded_by: null
+approved_by: null
+approved_at: null
+owners: [cixzhang, imdreamrunner, nynexman4464]
+applies_to:
+ [
+ apps/storybook/rtl-audit/,
+ .github/workflows/ci.yml,
+ .github/workflows/rtl-weekly.yml,
+ .github/scripts/rtl-audit-coverage.test.mjs,
+ .github/scripts/weekly-rtl-summary.js,
+ internal/eslint-plugin-astryx/no-physical-properties.js,
+ ]
+verified_by:
+ [
+ .github/scripts/rtl-audit-coverage.test.mjs,
+ .github/scripts/weekly-rtl-summary.test.mjs,
+ internal/eslint-plugin-astryx/no-physical-properties.test.mjs,
+ ]
+deciding_specs: []
+---
+
+# RTL audit evidence architecture
+
+## Purpose
+
+Astryx needs review evidence that distinguishes three claims:
+
+1. a component's direction-sensitive behavior was actually measured;
+2. RTL genuinely does not apply to the component; or
+3. RTL applicability remains unmeasured debt.
+
+A green lint rule, an all-N/A browser report, or a checked-in exception must not be
+credited with proving more than it observed. This draft records the current audit
+system and the policy choices still requiring owner approval.
+
+## System model
+
+```text
+changed component or full-library sweep
+ → discover the live Core/Lab component roster
+ → render relevant stories LTR and RTL
+ → run automatic relationship checks
+ → run curated component checks
+ → combine results per component
+ measured
+ verified N/A
+ unexplained coverage gap
+ optional known debt (unresolved proposal)
+ → report behavior failures, stale evidence, and applicability state
+```
+
+Lint and rendered audit are complementary:
+
+- lint rejects source patterns whose directional failure can be proved from the
+ declarations alone;
+- browser checks prove computed geometry, visibility, ordering, hit targets,
+ scroll/drag behavior, and composed transforms;
+- verified-N/A evidence, if admitted, proves only that reviewed source and stories
+ currently expose no direction-sensitive relationship.
+
+## Boundaries and invariants
+
+- **INV1 — Evidence follows the claim.** Source lint MUST NOT stand in for
+ computed geometry or interaction. Rendered screenshots MUST NOT stand in for
+ keyboard, drag, scroll, or hit-test behavior. A browser relationship check MUST
+ NOT claim that every component state was exercised when its stories did not
+ render those states.
+- **INV2 — RTL is graded as a relationship.** Directional evidence compares LTR
+ and RTL outcomes from the same built source. Pixel identity is not the contract;
+ semantic mirroring, order, side, geometry, and behavior are.
+- **INV3 — Behavior failures have no allowlist.** When an applicable automatic or
+ curated dimension shows incorrect mirroring, ordering, placement, geometry, or
+ behavior, the result is a finding. Known history does not turn incorrect RTL
+ behavior into pass or N/A.
+- **INV4 — All-N/A is not measured.** A component for which every dimension is N/A
+ has no positive RTL evidence. It is either verified not applicable under an
+ admitted evidence policy or remains a coverage gap/debt.
+- **INV5 — The live source roster owns coverage.** A component cannot disappear
+ from review merely because it has no story or because its story title changed.
+ Missing renderability is itself a coverage gap when the component is in scope.
+- **INV6 — Automatic and curated coverage compose.** Generic checks cover broadly
+ discoverable icon, decoration, and positional relationships. Component-specific
+ selectors are used only for semantics generic discovery cannot infer, such as
+ order, scroll direction, overlay side, or touch-target ownership.
+- **INV7 — CI scope and weekly scope answer different questions.** PR CI checks
+ components affected by the change. The scheduled unfiltered audit checks the
+ full live roster and shared regressions. Neither silently replaces the other.
+- **INV8 — Applicable evidence invalidates N/A.** A new applicable automatic or
+ curated dimension MUST make an existing verified-N/A declaration stale. Source-
+ and story-byte invalidation is a proposed additional closure requirement, not a
+ capability of the current reason-only registry.
+- **INV9 — Evidence mutations prove the gate.** A new audit rule or dimension MUST
+ demonstrate that the relevant defect passes unnoticed without the rule and is
+ detected after the rule. A passing happy path alone does not prove enforcement.
+- **INV10 — Touch-target evidence includes size and alignment.** For a coarse-
+ pointer target, proving that the target center matches the visible control is
+ insufficient. Browser evidence MUST also measure the effective/native target
+ against the owning component's stated minimum size. Current D7 enforcement
+ checks centering and hit-testing only; size enforcement is an open gap.
+- **INV11 — Audit state is not product intent.** This architecture owns evidence
+ admission and freshness. The human meaning of correct bidirectional behavior
+ belongs in `design:bidirectionality`; component contracts own local adoption and
+ exceptions.
+
+## Change coupling
+
+Changes to audit dimensions, component discovery, curated targets, applicability
+registries, proposed evidence hashes, workflow scope, summaries, or RTL lint must
+review this record.
+
+A new direction-sensitive component contract must map its requirement to executable
+evidence or remain an explicit gap. If owners admit verified N/A or known debt with
+byte-bound closure, the implementation MUST bind that evidence to the canonical
+source, directly owned helpers/styles, and owned stories so relevant changes fail
+stale rather than silently retaining their old classification.
+
+Promoting this record to `current` requires owner decisions on verified-N/A
+admission and pre-existing debt. Until then, implementation PRs may improve
+mechanics but cannot claim the unresolved policy as settled.
+
+## Owning code
+
+- `apps/storybook/rtl-audit/rtl-audit.mjs` — renders and compares LTR/RTL behavior.
+- `apps/storybook/rtl-audit/rtl-audit-coverage.mjs` — discovers the live roster and
+ combines measured, N/A, and gap states.
+- `apps/storybook/rtl-audit/targets.json` — curated D2/D3/D4/D7 behavior targets.
+- `apps/storybook/rtl-audit/verified-not-applicable.json` — candidate verified-N/A
+ registry; its admission policy is unresolved in this draft.
+- `apps/storybook/rtl-audit/known-coverage-gaps.json` — proposed by PR #5988
+ for pre-existing debt; this file is not present on current `main`, and its
+ admission policy is unresolved in this draft.
+- `.github/workflows/ci.yml` — affected-component PR audit.
+- `.github/workflows/rtl-weekly.yml` — full-roster scheduled audit.
+- `internal/eslint-plugin-astryx/no-physical-properties.js` — source-provable
+ physical/logical declaration constraints.
+- `design:bidirectionality` — candidate human product intent, owned separately.
+
+The shipped RTL mechanics themselves still lack one canonical architecture owner.
+A follow-up `architecture:bidirectional-layout` record should own CSS logical
+properties, `rtlStyles`, lazy DOM direction reads, centering helpers, and
+first-paint/hydration invariants. This evidence record MUST NOT be treated as that
+missing implementation contract.
+
+## Deciding specs
+
+No current system spec settles verified-N/A or known-debt admission. The current
+knowledge architecture requires evidence freshness and explicit human decisions;
+this draft narrows the remaining decision to the options below.
+
+## Verification
+
+| Invariant | Evidence | Failure signal |
+| ---------------- | ----------------------------------------------------------- | ------------------------------------------------------------------------------------------- |
+| INV1, INV2, INV6 | focused audit fixtures plus real built Storybook mutations | A source-only rule claims rendered correctness, or a relationship defect stays green |
+| INV3 | automatic/curated failure tests and summary tests | A known bad RTL outcome is suppressed as expected or N/A |
+| INV4, INV5 | coverage fixtures with missing/no-applicable stories | A component disappears or all-N/A reports as RTL-ready |
+| INV7 | PR workflow scope tests and weekly-summary tests | A changed component is skipped or full-roster debt is mistaken for PR-only evidence |
+| INV8 | automatic/curated applicability tests | A newly applicable dimension leaves a verified-N/A declaration current |
+| INV9 | red/green mutation receipts for each new dimension | A new dimension has only passing self-confirming tests |
+| INV10 | none for size today; D7 covers center/hit in LTR/RTL, sm/md | Removing the component's minimum native hit area remains green because D7 never checks size |
+
+### Candidate evidence states
+
+| State | Minimum evidence | May clear a new/changed component? | Current decision state |
+| ----------------------- | -------------------------------------------------------------------------------------------------- | ---------------------------------- | --------------------------------- |
+| Measured | At least one applicable relationship or behavior check passes; every applicable check passes | yes | settled by current audit behavior |
+| Verified N/A | Today: human-reviewed reason; proposed: source/helper/story evidence with fail-closed invalidation | unresolved | owner decision required |
+| Coverage gap | No applicable measurement and no admitted current N/A evidence | no | settled |
+| Known pre-existing debt | Proposed removal-only baseline entry that remains visible and non-failing while unchanged | unresolved | owner decision required |
+| Behavior failure | An applicable dimension fails | no | settled; no suppression |
+| Stale evidence | Today: an applicable dimension appears; proposed: recorded evidence bytes change | no | partially implemented |
+
+### Current enforcement gaps
+
+- D7 currently measures hit-testing and center alignment in LTR/RTL at sm/md under
+ a coarse pointer, but does not fail when the native target falls below the owning
+ component's minimum size. The #5177 review demonstrated this false-green by
+ removing the coarse native-input expansion. Add the size assertion and report
+ the native input box before crediting INV10 as enforced.
+- The reason-only verified-N/A registry becomes stale when an automatic/curated
+ dimension becomes applicable, but does not yet bind evidence to source/helper/
+ story bytes.
+- CI summary code still contains an unused `knownNotRtl` label path even though the
+ report emits no such allowlist. Remove it so the implementation cannot be read as
+ permitting known behavior failures.
+- Checkbox, Radio, and Switch currently implement a ×ばつ24 coarse native target,
+ but their component contracts do not yet own that number. Record target-size
+ requirements under the relevant component/composition contract and AST-020.
+
+## Open decisions
+
+### OQ1 — Is verified N/A an admitted closure state?
+
+- **Option A:** Admit verified N/A when one declaration carries a specific reason.
+ Before this option can become current, add evidence over the canonical source,
+ directly owned helpers/styles and owned stories, plus fail-closed invalidation.
+ This recognizes truly direction-neutral components without manufacturing
+ meaningless browser tests.
+- **Option B:** Require every component to have executable rendered measurement.
+ This removes judgment-based N/A but forces artificial checks for components with
+ no directional relationship.
+
+### OQ2 — May pre-existing unexplained all-N/A enter removal-only known debt?
+
+- **Option A:** Snapshot existing gaps as visible, non-failing debt while new gaps,
+ stale entries, behavior failures, and removals that do not become measured/N/A
+ continue to fail. This allows gradual rollout without hiding the inventory.
+- **Option B:** Keep every unexplained gap failing. This is simpler and stricter,
+ but leaves the audit permanently red until the full historical roster is
+ measured or reviewed.
+
+### OQ3 — When does RTL become a required merge check?
+
+The current PR audit is soft-gated. Owners must choose the prerequisite clean
+states and stability window before making it required. Regardless of gate status,
+reports must remain truthful and behavior failures must not be relabeled.