diff --git a/docs/adr/0003-confined-shell-stub-and-the-fail-open-allowance.md b/docs/adr/0003-confined-shell-stub-and-the-fail-open-allowance.md
new file mode 100644
index 0000000..1545989
--- /dev/null
+++ b/docs/adr/0003-confined-shell-stub-and-the-fail-open-allowance.md
@@ -0,0 +1,85 @@
+# ADR 0003 — The confined shell is a stub today: honest exec-confinement state (and the embedder's fail-open allowance)
+
+- Status: Accepted
+- Date: 2026-06-21
+- Context: `agent-bridle-tool-shell` (the stub release), the `SandboxKind`
+  honesty path (ADR 0002 I9) and the `LandlockSandbox` backstop (I10), and the
+  relationship to embedders that fail closed downstream (newt-agent).
+- **Extends ADR 0002** (the hard invariants) by recording the *current
+  operational state* of the exec-confinement path — in the same "never
+  overclaim" spirit (I9 + the status legend). It changes no invariant.
+- Companion (the downstream view): `newt-agent/docs/decisions/ocap_confinement_model.md`.
+  Prose: *The Age of the Confused Deputy*; `docs/DESIGN.md`.
+
+## Question
+
+ADR 0002 states the invariant **contract** (what MUST hold). This ADR answers a
+narrower, time-sensitive question so no one — human or agent — reads the contract
+as a description of *today*: **with the brush `CommandInterceptor` not yet
+upstream (reubeno/brush#1184), what does the confined shell actually do right
+now, what does that mean for embedders, and how do we keep the story honest?**
+
+## The current state (honest)
+
+- **The brush-backed confined shell is a STUB.** `agent-bridle-tool-shell` ships
+  with the brush deps removed (`Cargo.toml`: "STUB RELEASE — full brush-backed
+  shell is temporarily disabled pending reubeno/brush#1184"). `ShellTool`
+  registers but **fails closed on every invocation** — `invoke` returns
+  *"shell tool is temporarily unavailable in this build"* (tested:
+  `invoke_returns_unavailable_error`). This is **I5 (fail closed) doing its job**:
+  absent the real interceptor, exec authority is denied, not waved through.
+- **`sandbox_kind` honestly reports `None`** (I9). Nothing is labeled
+  kernel-confined, because nothing runs through a kernel sandbox. The
+  `LandlockSandbox` backstop (I10) confines a *permitted external program's
+  interior* — but a permitted program only runs through the real shell, and the
+  real shell is stubbed, so I10 is **not exercised today**.
+- Therefore the **operative exec floor in a stub build is: fail-closed (deny
+  everything).** The usable confined shell + its Landlock backstop are the
+  **target**, gated on reubeno/brush#1184 + agent-bridle#20.
+
+This does not weaken ADR 0002. I1–I3 (the lattice, no-amplify), I5 (fail-closed),
+I6 (canonical path containment), I9 (`sandbox_kind` honesty) remain enforced.
+What is *pending* is the interceptor that makes confined exec **usable** rather
+than merely fail-closed — exactly the state ADR 0002's status legend and the
+`Cargo.toml` notes already anticipate. No invariant moves down the ladder here;
+this ADR only timestamps where the shell sits on it.
+
+## What this means for embedders — and the fail-open allowance
+
+agent-bridle fails closed by construction (I5). That is correct, but in a stub
+build it means a downstream harness's `run_command` denies **everything**, which
+can make the agent unusable for legitimate work the leash hasn't yet been taught.
+agent-bridle does not paper over this and does not silently open — the policy
+choice belongs to the **embedder**, and must be disclosed:
+
+- **newt-agent** embeds the stub, so it fails closed on exec by default, and
+  **deliberately provides a fail-OPEN allowance: `--yolo` / `--disable-ocap`
+  (`NEWT_DISABLE_OCAP=1`).** It routes `run_command` *around* bridle to the plain
+  host shell (`sandbox_kind = none`) for that invocation — the human's explicit,
+  per-invocation choice to **unbridle** the agent. It is not insistence; it is an
+  *allowance*: the bridle is on by default, and the human may take it off. (newt
+  keeps its `web_fetch` net leash and its native-fs fence on under `--yolo`; only
+  the spawned subprocess is unconfined.) See
+  `newt-agent/docs/decisions/ocap_confinement_model.md`.
+
+The division of labor: **bridle's job is to fail closed; whether to offer a
+fail-open escape — and how loudly — is the embedder's**, and it must be explicit
+and disclosed. agent-bridle never opens silently.
+
+## Honesty rule (reaffirmed for this surface)
+
+No agent-bridle doc — and no embedder doc — may present the brush-backed confined
+shell or its Landlock backstop as **active** while the shell is a stub.
+`sandbox_kind` (here) and `verify_b1()` (downstream, in newt-agent) are the source
+of truth: if they say `none` / `Absent`, no prose may say "confined." When
+brush#1184 lands and the real shell is wired, **this ADR is revisited** and the
+relevant ADR 0002 invariants move up the ladder — never silently.
+
+## Consequences
+
+- Reviewers hold PRs to: the stub shell fails closed (I5); `sandbox_kind` never
+  overclaims (I9); the reubeno/brush#1184 + agent-bridle#20 dependency is named,
+  not hidden.
+- Embedders document their own fail-open allowance (if any) honestly and
+  cross-link this ADR, so bridle and its hosts tell one consistent story about
+  what confinement buys *today* versus at target.