diff --git a/docs/images/doc-images/actor-topology.png b/docs/images/doc-images/actor-topology.png
new file mode 100644
index 0000000000..3d94050da0
Binary files /dev/null and b/docs/images/doc-images/actor-topology.png differ
diff --git a/docs/internal/threat-model.md b/docs/internal/threat-model.md
new file mode 100644
index 0000000000..b45552a941
--- /dev/null
+++ b/docs/internal/threat-model.md
@@ -0,0 +1,265 @@
+# Lind-Wasm Threat Model
+
+## 1. Overview
+
+Lind-Wasm runs multiple mutually untrusting programs, called *cages*, inside one
+unprivileged Linux process. Two layers keep them apart:
+
+1. **WebAssembly SFI.** Each cage runs in its own Wasmtime `Store` with its own
+   linear memory, and every memory access is bounds-checked by JIT-emitted code.
+   A cage cannot touch memory outside its own region no matter what it runs.
+2. **RawPOSIX.** A user-space microvisor that handles every syscall a cage makes.
+   A cage cannot issue a raw Linux syscall. Calls go `make_syscall()` → 3i →
+   (grate stack) → RawPOSIX → kernel, and RawPOSIX checks each one before
+   forwarding.
+
+In between, **3i** routes a cage's syscalls through optional policy components
+called *grates* before they reach RawPOSIX. Affecting the host or another cage
+means defeating both layers.
+
+## 2. Methodology
+
+We do not sort components into trusted and untrusted, because that hides the fact
+that a grate is trusted by the cages it isolates and by nothing else. Instead we
+ask:
+
+- What are the actors?
+- What is isolated from what?
+- For each attacker position, what can it do to each other actor?
+
+We also do not assume the components we rely on are bug-free. For each one we
+state what a bug in it would cost. A bug in RawPOSIX and a bug in the host kernel
+are both bad, but their blast radius is not the same.
+
+## 3. Actors
+
+The instructive topology is small:
+
+![Lind-Wasm actor topology](../images/doc-images/actor-topology.png)
+
+*Figure 1. Color marks trust position: red cages (untrusted), amber grates
+(untrusted), green the Lind runtime and TCB, slate the kernel and hardware. The
+dashed line is the Wasm SFI boundary, guest above and host below. Cage A routes
+through Grate B, its ancestor grate. Cage C and Grate D each sit in their own
+separate stack, with their partner boxes drawn lighter, unrelated to A and B and
+to each other. RawPOSIX, Wasmtime, and 3i form one band because nothing isolates
+them, and the paper's convention puts them at the bottom.*
+
+| Actor | Role in the model |
+| --- | --- |
+| **Cage A** | The reference application cage. Fully adversarial. |
+| **Grate B** | The ancestor grate of Cage A. Mediates Cage A's syscalls. |
+| **Cage C** | An unrelated cage. No routing relationship to A or B. |
+| **Grate D** | An unrelated grate. No routing relationship to anything in A's stack. |
+| **RawPOSIX + Wasmtime + 3i** | The Lind runtime (TCB). One actor because the three share fate (Section 4). |
+| **Host kernel + hardware** | The bottom of the trust stack. |
+
+Grates are themselves cages under the same SFI. Cage versus grate is a question
+of position in the stack, not of isolation mechanism.
+
+## 4. Trust relationships
+
+Stated as isolation rather than a label:
+
+- **Host kernel + hardware.** Assumed correct. Out of scope to defend.
+- **Lind runtime (Wasmtime + 3i + RawPOSIX).** The TCB. All three are host Rust
+  code outside any sandbox, with no isolation between them, so a bug in one is a
+  bug in all. Their state (handler tables, per-cage vmmap, fdtables, signals) is
+  host memory the guest cannot address. Compromise it and the attacker owns every
+  cage and grate, still capped at the unprivileged-process ceiling (Section 7,
+  Case 4).
+- **Grates.** Untrusted code under the same SFI as a cage, never in the TCB. A
+  grate is trusted only by its own descendant cages, and only to enforce their
+  policy. See below.
+- **Application cages.** Untrusted and fully adversarial, buggy or malicious.
+
+### TCB versus trusted policy surface
+
+The TCB is the set of components that break the system's guarantees if
+compromised. It is fixed across deployments and small enough to analyze.
+
+The trusted policy surface is the grates that must be correct for a *given*
+deployment's policy to hold. A bad grate can subvert the policy on the cages it
+mediates: skip a namespace check, leak data it pulled via
+`copy_data_between_cages`, or fake a return value instead of doing the call. It
+still cannot touch the TCB, the host, or unrelated cages.
+
+So a deployment audits two things separately. The TCB once, and each grate on its
+own. Grates are small and single-purpose, so one can be checked without reading
+the others.
+
+## 5. Attacker model
+
+### Compromised cage
+
+The attacker fully controls one or more cages and can:
+
+- Run arbitrary Wasm, including exploits against bugs in its own linear memory.
+- Read and write any byte of its own linear memory.
+- Call any host import with chosen arguments, including `register_handler`,
+  `make_syscall`, and `copy_handler_table_to_cage`, subject to 3i's interposition
+  and scoping.
+- Call the syscall API with arbitrary inputs.
+- Rewrite its own handler table, try to copy in a different one, or try to drive
+  3i operations against peer or ancestor cages (scoping denies these).
+
+It cannot, by the Wasm memory model rather than any policy check, address host
+memory, issue raw Linux syscalls, alter JIT output, or reach another cage's
+memory.
+
+Out of the attacker's control: the kernel and hardware, the Wasmtime JIT and its
+output, the 3i logic and handler tables, RawPOSIX state, and other cages' memory.
+
+### Compromised grate
+
+A grate that is compromised can observe, change, drop, or fake results for any
+syscall it mediates, for the cages below it. It cannot read or write cages that
+do not route through it (except through `copy_data_between_cages`), corrupt
+RawPOSIX, the handler tables, or the runtime, or reach cages above it or beside
+it. It runs under the same SFI as any cage. Its reach is the set of cages it
+mediates and the 3i API, which is the same API any cage has.
+
+## 6. The matrix
+
+Columns are the **attacker**, rows the **victim**. Each cell is what the attacker
+can do to the victim.
+
+| Victim \ Attacker | Cage A | Grate B | Cage C | Grate D | Runtime | Kernel+HW |
+| --- | --- | --- | --- | --- | --- | --- |
+| **Cage A** | N/A | CC | RE | RE | KO | KO |
+| **Grate B** | FE | N/A | RE | RE | KO | KO |
+| **Cage C** | RE | RE | N/A | RE | KO | KO |
+| **Grate D** | RE | RE | RE | N/A | KO | KO |
+| **Runtime** (RawPOSIX/Wasmtime/3i) | U | U | U | U | N/A | KO |
+| **Kernel + HW** | U | U | U | U | U | N/A |
+
+Legend:
+
+- **CC** Complete control: read and write the victim's memory, rewrite its
+  syscalls, change its policy.
+- **FE** Denial of service on the grate, plus a direct input channel into it. The
+  cage can flood the grate that mediates it and hand it malformed calls. Against a
+  correct grate this is only DoS: memory isolation holds and nothing is
+  compromised. It differs from RE because the cage routes through this grate, so
+  this is the one place a grate's code takes cage-controlled input directly, where
+  a grate bug would surface (Case 2).
+- **RE** Resource exhaustion and performance interference. Memory isolation holds.
+- **KO** Complete compromise, capped at the unprivileged-process ceiling
+  (Section 7).
+- **U** Unaffected. No path from attacker to victim.
+- **N/A** Same actor.
+
+## 7. Cell reasoning
+
+Six cases cover the whole matrix.
+
+**Case 1: ancestor grate against its descendant cage (CC).** Grate B mediates
+Cage A, so a compromised Grate B controls Cage A completely: read and write its
+memory, rewrite its syscalls (`open` of A becomes `open` of B, `fork` becomes
+`exit`), change its policy. This is a grate's normal authority over the cages it
+mediates, turned hostile, and it is still capped at the unprivileged-process
+ceiling (Case 4).
+
+**Case 2: descendant cage against its ancestor grate (FE).** Cage A's calls run
+through Grate B, so it has reach an outsider does not. It can flood Grate B
+through 3i and feed it crafted calls hoping to trip a bug in the grate. Against a
+correct grate this amounts to denial of service and nothing more: memory
+isolation holds and the grate is not taken over. A buggy grate is the exception,
+since this is the one place a cage's input reaches a grate's code directly. That
+direct channel is what separates FE from RE, where an unrelated cage has no path
+to the grate at all.
+
+Isolation holds here because of how pointer arguments cross the boundary. A
+pointer a cage hands to a grate is never dereferenced blind. Either it travels
+down tagged with its originating cage and RawPOSIX faults if it falls outside
+that cage's address space, or the grate pulls the bytes in with the
+copy-from-cage call, which checks the same provenance before copying. A cage
+cannot turn a bad pointer into a memory-safety bug in a grate.
+
+**Case 3: any unrelated pair (RE).** Two parties with no routing relationship, in
+either direction, can only compete for shared resources: CPU, file descriptors,
+and the like. Memory isolation holds, and neither can read the other's memory or
+touch its policy.
+
+**Case 4: the runtime as attacker (KO).** A memory-safety bug in Wasmtime, 3i, or
+RawPOSIX that an attacker rides into the runtime compromises every cage and grate
+(the three share fate, Section 4). Even then the attacker is just an unprivileged
+Linux process. Nothing Lind adds sits in the kernel or hardware, so the worst its
+own code can cost you is whatever an unprivileged process can do. That ceiling is
+the main resilience argument for the design.
+
+**Case 5: cage or grate against the runtime or kernel (U).** Neither has a path
+downward. The runtime's state is host memory it cannot address (Section 4), and
+the only way up is a runtime bug, which is Case 4. So these cells are unaffected.
+
+**Case 6: kernel or hardware compromised (KO, out of scope).** If either falls,
+everything is lost and Lind cannot prevent it. It sits in the matrix only to keep
+the assumption visible.
+
+## 8. Security properties
+
+**P1: cross-cage memory isolation.** No cage can read or write another cage's
+linear memory. Each cage's memory is a separate mapped region, and every load and
+store is bounds-checked by JIT code against that cage's own `VMMemoryDefinition`.
+This holds against arbitrary hostile code.
+
+**P2: syscall mediation.** No cage issues a raw Linux syscall. Everything goes
+`make_syscall()` → 3i → grate stack → RawPOSIX → kernel, and RawPOSIX validates
+pointer arguments against the calling cage's vmmap before forwarding.
+
+**P3: directional confinement.** Syscalls run through the grate stack toward
+RawPOSIX, and a cage cannot reach past, around, or below the grates that mediate
+it.
+
+**P4: handler-table integrity.** The 3i logic and per-cage handler tables are
+host memory, unreachable from any cage (Section 4), so routing cannot be tampered
+with even if the 3i policy logic has a bug. A cage influences routing only
+through the declared 3i import API, which is itself open to ancestor-grate
+interposition.
+
+**P5: fork inheritance.** A forked child inherits its parent's handler table
+through `copy_handler_table_to_cage`, so a confined cage cannot spawn a child that
+escapes its policy. An ancestor grate can intercept the fork to constrain the
+child further.
+
+**P6: control-flow integrity.** Wasm's structured bytecode rules out arbitrary
+native jumps, indirect calls are type-checked and table-bounded, and return
+addresses live on Wasmtime's native stack rather than in cage-addressable memory,
+so a cage's loads and stores cannot reach them. RawPOSIX also refuses
+`PROT_EXEC` on any `mmap` or `mprotect` that also asks for `PROT_WRITE`, closing
+the one legitimate path to a writable-executable region a cage could run injected
+code from.
+
+### How 3i enforces P3
+
+Three properties of 3i close three different escapes:
+
+1. **Handler changes are interposable.** `register_handler` and
+   `copy_handler_table_to_cage` go through 3i themselves, so an ancestor grate can
+   see and deny a cage's attempt to rewrite the routing over it. A cage cannot
+   quietly change the rules it runs under.
+2. **3i operations are scoped to descendants.** A grate can act only on cages
+   below it, never a parent, sibling, or stranger. 3i enforces this in host code,
+   so it holds even if a compromised grate learns another cage's ID.
+3. **Dispatch itself is not interposable.** `make_syscall`, the primitive
+   everything else builds on, runs in trusted host code and cannot be hijacked.
+   Interposition changes what a handler does, not which handler runs.
+
+## 9. Out of scope
+
+Not addressed here:
+
+- **Speculative execution and other microarchitectural side channels.**
+- **Memory safety inside a cage.** Corruption confined to a cage's own memory
+  breaks no cross-cage guarantee. We lean on the underlying SFI model for in-cage
+  safety and do not reprove it.
+- **Wasmtime correctness.** The JIT and its output are assumed correct, and P1
+  and P6 rest on this.
+- **Compiler and glibc correctness.**
+
+### Resource exhaustion
+
+Resource exhaustion is **modeled but not mitigated**: it is the residual in every
+cross-party cell of the matrix. The claim is only that isolation and policy
+confinement survive it. Fair sharing of CPU, file descriptors, and the like
+across cages is future work, not a current property.
diff --git a/mkdocs.yml b/mkdocs.yml
index 6890b91332..e62fc49c66 100644
--- a/mkdocs.yml
+++ b/mkdocs.yml
@@ -25,6 +25,7 @@ nav:
     - Multi-Processing: internal/multiprocess-support.md
     - Signals: internal/signals.md
     - Memory: internal/memory.md
+    - Threat Model: internal/threat-model.md
   - Contributing:
     - contribute/index.md
     - Development setup: contribute/dev-container.md