diff --git a/docs/roadmap.md b/docs/roadmap.md
index e1d5467..7198348 100644
--- a/docs/roadmap.md
+++ b/docs/roadmap.md
@@ -265,6 +265,117 @@ clean. Pick the next milestone from P2.)_
   exercising the LLM-natural Show-body shape.
   - context: per-iter journal 2026-05-13-iter-str-concat.md.
 
+- [ ] **\[milestone\]** Stateful islands — bounded mutation for
+  streaming workloads (`Stateful a b` + `!Mut` effect + `mut`
+  syntactic block). Adds a sealed mutable-state layer on top of
+  the pure core: a `Stateful a b` first-class type whose interior
+  is mutable, whose exterior is a typed callable with `!Mut` in
+  its effect set, and whose state non-aliasing is enforced by
+  uniqueness inference at the block boundary. Targets the
+  online / streaming workload class — rolling indicators, IIR
+  filters, online aggregates, sensor fusion, online learning —
+  whose mathematics is "new sample + old state → new state +
+  output" per step, and which today's pure-functional state-
+  threading makes ergonomically expensive at scale (multi-record
+  explicit threading for the canonical sliding-window SMA, no
+  zero-allocation pipe combinator, growing tuple-state types as
+  pipelines lengthen).
+
+  **Motivation.** AILang's signature-as-contract thesis is *better*
+  served by an explicit `Stateful a b` + `!Mut` annotation than by
+  the implicit-closure-mutation idiom of myc / Lua / JS factory
+  patterns: the signature tells the truth about time-identity
+  without the body needing to be read — exactly the LLM-author
+  correctness affordance AILang exists to deliver, extended to a
+  workload class it does not yet serve. Decision 10's constraint #3
+  forbids *shared* mutable refs (DESIGN.md:1082); it does not
+  forbid uniqueness-bounded mutation. Lean 4 (`ST`), Roc (`Task`),
+  Haskell (`ST`/`IO`), and Koka (effects) all use a layered design
+  of this shape to host exactly this workload. AILang's existing
+  `own` / `borrow` mode machinery plus its effect-slot architecture
+  are the foundation; this milestone supplies the layer that sits
+  on top.
+
+  **Scope (subject to brainstorm refinement).**
+  - `Stateful a b` as a first-class type — a sealed callable with a
+    hidden mutable env, externally a typed callable whose effect
+    set includes `!Mut`.
+  - `!Mut` effect, the first non-IO / non-Diverge effect; forces
+    the effect-handler infrastructure forward.
+  - `mut` block as the syntactic boundary in Form A — outside,
+    AILang's pure self; inside, `var` / `assign` / mutable arrays /
+    while-loops legal.
+  - `var x = expr` + `assign x expr` AST nodes, legal only inside
+    `mut`.
+  - Mutable array primitive (`MutArray a`, O(1) index/update under
+    uniqueness) — co-developed because the ring-buffer use case
+    that motivates the whole effort has no carrier today.
+  - `pipe : Stateful a b → Stateful b c → Stateful a c` as a
+    built-in combinator with zero-allocation lowering — composition
+    is fusion at codegen, not closure-pair layering at runtime.
+  - Decision 12 (or amendment to Decision 10) that names
+    uniqueness-bounded mutation as the legitimate exception to
+    "no shared mutable refs", and the pure / mutable-island layering
+    as the architectural shape.
+
+  **Blockers (hard prerequisites + open design questions).**
+  - *Effect handlers absent.* DESIGN.md:2663 — "No effect
+    handlers — only the built-in IO and Diverge ops." `!Mut` is the
+    first non-trivial effect and forces handler infrastructure to
+    ship. May warrant lifting out as its own prerequisite milestone;
+    brainstorm decides.
+  - *Uniqueness inference through `var` captures.* Current
+    inference operates over the immutable RC graph; mutable
+    bindings captured by closures need a more powerful pass. Lean 4
+    has a known algorithm; AILang does not implement it.
+  - *No mutable-array primitive.* No `Array a`, no slab container
+    of any kind in the language today. Separate spec needed inside
+    this milestone (representation, `own`-only or `borrow`-able,
+    bounds-checking discipline).
+  - *`runST`-equivalent escape discharge.* Producing a `Stateful`
+    that legitimately escapes its `mut` block while proving the
+    inner state doesn't leak. Haskell's rank-2 trick
+    (`runST :: (forall s. ST s a) -> a`) is unavailable —
+    DESIGN.md:1930 confirms higher-rank polymorphism is not
+    supported. Need an alternative — likely sealed-by-construction
+    at the factory boundary, or an effect-mode tag that gates
+    escape.
+  - *Form A surface design.* Decision 1 forbids macros (source =
+    data, not text), so `mut` / `var` / `assign` are genuine AST
+    nodes with round-trip-invariant coverage. Surface needs LLM-
+    utility validation (does the author reach for `var` / `mut`
+    unprompted?) before implementation.
+  - *Codegen for in-place struct-field writes.* `Term::ReuseAs`
+    today lowers ADT in-place rewrite; no direct mutable struct-
+    field-write path exists. `crates/ailang-codegen/src/escape.rs`
+    plus the lowering passes need extension.
+  - *Boundary escape analysis.* "Interior state doesn't leak" is an
+    escape analysis specifically over `var` / `ref` values; the
+    existing pass covers allocations, not mutable cells.
+  - *Decision-10 status.* Whether to amend Decision 10 inline or
+    add a Decision 12 that names the mutable-island layer alongside
+    the pure layer. The latter is probably cleaner — Decision 10
+    stays load-bearing for the pure layer, Decision 12 names the
+    extension and its discipline.
+  - *Streaming bench corpus.* Current corpus (list_sum, tree_walk,
+    closure_chain, hof_pipeline) is all pure. Streaming-specific
+    benches (SMA pipeline, IIR filter, online stats) plus an
+    external baseline (myc, hand-C, Python/NumPy) need to exist to
+    validate that the layered design hits the zero-alloc
+    performance target that myc demonstrates.
+  - *LLM-utility test.* DESIGN.md §"Feature-acceptance criterion"
+    is the gate: the surface must produce code an LLM author
+    naturally writes. Fieldtest after spec, before any code
+    commits.
+
+  - context: 2026-05-15 chat on the `~/sma_factory.myc` analysis —
+    myc's stateful-closure-plus-pipe idiom delivers a streaming
+    workload pattern (3-line SMA factory, 2-line pipeline, zero
+    runtime allocation per tick) that AILang's pure-only model
+    cannot match without this layered extension. Pending brainstorm
+    will produce `docs/specs/<date>-stateful-islands.md` and decide
+    the effect-handler-prerequisite question.
+
 ## P3 — Ideas
 
 - [x] **\[todo\]** `compare_primitives_smoke.ail` counterpart.