diff --git a/development/HARDENING.md b/development/HARDENING.md
new file mode 100644
index 0000000..8aca7af
--- /dev/null
+++ b/development/HARDENING.md
@@ -0,0 +1,217 @@
+# OS Fingerprint Spoofing — Hardening Roadmap
+
+This document describes the current state of OS fingerprint spoofing in DECNET
+and the planned improvements to make `nmap -O`, `p0f`, and similar passive/active
+scanners see the intended OS rather than a generic Linux kernel.
+
+---
+
+## Current State
+
+OS spoofing is partially implemented. Each archetype declares an `nmap_os` slug
+(e.g. `"windows"`, `"linux"`, `"embedded"`). The **composer** resolves that slug
+via `os_fingerprint.get_os_sysctls()` and injects the resulting kernel parameters
+into the **base container** as Docker `sysctls`. Service containers inherit the
+same network namespace via `network_mode: "service:<base>"` and therefore appear
+identical to outside scanners.
+
+### Currently tuned knobs
+
+| Sysctl | Purpose |
+|---|---|
+| `net.ipv4.ip_default_ttl` | Primary TTL discriminator (64 = Linux, 128 = Windows, 255 = Embedded) |
+| `net.ipv4.tcp_syn_retries` | SYN retransmit count before giving up |
+
+### What this fools
+
+| Scanner probe | Status |
+|---|---|
+| ping TTL | ✅ Fully spoofed |
+| TCP SYN retry count | ✅ Tuned |
+| `nmap -O` OS family (Win vs Linux) | ⚠️ Partial — likely correct family, wrong version |
+| `p0f` passive fingerprint | ⚠️ Partial — TTL correct, window/options wrong |
+| Full `nmap -O` version/build match | ❌ Not achievable without deeper tuning |
+
+---
+
+## Improvement Phases
+
+### Phase 1 — Extended Sysctls (Low effort, High impact)
+
+Several additional sysctls are **network-namespace-scoped** and can be safely set
+per-container without `--privileged`. These directly affect nmap's SEQ, OPS, and
+WIN probe groups.
+
+**Changes required:** extend `OS_SYSCTLS` in `decnet/os_fingerprint.py`.
+
+| Sysctl | nmap probe group | Windows | Linux | Embedded |
+|---|---|---|---|---|
+| `net.ipv4.tcp_timestamps` | SEQ/OPS — timestamp option presence | `0` | `1` | `0` |
+| `net.ipv4.tcp_window_scaling` | WIN — window scale option | `1` | `1` | `0` |
+| `net.ipv4.tcp_sack` | OPS — SACK permitted option | `1` | `1` | `0` |
+| `net.ipv4.tcp_ecn` | ECN probe — explicit congestion notification | `0` | `2` | `0` |
+| `net.ipv4.ip_no_pmtu_disc` | IE — DF bit copying in ICMP replies | `0` | `0` | `1` |
+| `net.ipv4.tcp_fin_timeout` | T2–T6 — FIN_WAIT duration | `30` | `60` | `15` |
+
+> **Highest single-value impact:** setting `net.ipv4.tcp_timestamps = 0` for
+> Windows is the strongest signal. nmap's OPS probes explicitly look for the TCP
+> timestamp option; its absence is a definitive Windows discriminator.
+
+**Expected result after Phase 1:** `nmap -O` correctly identifies OS family in
+the vast majority of scans. `p0f` passive fingerprinting becomes significantly
+more convincing.
+
+---
+
+### Phase 2 — TCP Window Size Mangling (Medium effort, Very high impact)
+
+nmap's WIN probes record the raw **TCP window size** in SYN-ACK replies. This
+is the single most discriminating feature after TTL. It cannot be set with
+per-namespace sysctls because `net.core.rmem_default` is global.
+
+The fix is an **iptables rule applied at base container startup** via a custom
+entrypoint script.
+
+#### Target window sizes by OS
+
+| OS | TCP Window Size | Notes |
+|---|---|---|
+| Windows 10 / 11 | `64240` | Most common modern value |
+| Windows 7 / Server 2008 | `8192` | Classic Windows signature |
+| Linux 5.x / 6.x | `29200` | Default `tcp_rmem` min/4 |
+| Linux 4.x | `43690` | Older default |
+| FreeBSD / macOS | `65535` | BSD signature |
+| Embedded / Cisco | `4128`–`8760` | Varies widely |
+
+#### Implementation sketch
+
+Add a parameterized entrypoint script (`templates/base/entrypoint.sh`) that
+receives the target window size as an environment variable and applies an
+`iptables` MANGLE rule before yielding to `sleep infinity`:
+
+```bash
+#!/bin/sh
+# Apply TCP window size spoofing via iptables mangle
+if [ -n "$SPOOF_TCP_WINDOW" ]; then
+    iptables -t mangle -A POSTROUTING -p tcp \
+        -j TCPMSS --set-mss 1460
+    # Clamp outgoing window to the target value
+    # Requires xt_TCPMSS kernel module on the host
+fi
+exec sleep infinity
+```
+
+The composer would inject `SPOOF_TCP_WINDOW` as an environment variable on the
+base container, sourced from the OS fingerprint profile.
+
+**Required changes:**
+- `os_fingerprint.py` — add `tcp_window` field to each OS profile.
+- `composer.py` — pass `SPOOF_TCP_WINDOW` env var to base container.
+- `templates/base/entrypoint.sh` — new file, applies the iptables rule.
+- `templates/base/Dockerfile` — new file, minimal image with `iptables`.
+
+> **Note:** requires `NET_ADMIN` capability (already granted) and the
+> `xt_TCPMSS` and `xt_mangle` kernel modules loaded on the host. Both are
+> present in any standard Linux distribution kernel.
+
+---
+
+### Phase 3 — ICMP Response Tuning (Medium effort, Medium impact)
+
+nmap's `IE` probe group sends two ICMP echo requests with specific ToS values,
+code fields, and payload sizes and inspects what the target returns. Currently
+nothing in DECNET controls ICMP echo reply behavior.
+
+**Namespace-scoped sysctls to add per-OS:**
+
+| Sysctl | Effect | Windows | Linux |
+|---|---|---|---|
+| `net.ipv4.icmp_ratelimit` | Packets/sec rate limit on ICMP errors | `0` (none) | `100` |
+| `net.ipv4.icmp_ratemask` | Which ICMP types are rate-limited | `0` | `6168` |
+
+**Expected result:** nmap's `IE` response classification improves from
+"no response / filtered" to a correctly typed ICMP echo reply with OS-correct
+rate limiting behavior.
+
+---
+
+### Phase 4 — IP ID Sequence Behavior (Hard, Medium impact)
+
+nmap's SEQ probe group fires 6 TCP SYN packets in rapid succession and measures
+the **IP ID increment pattern** across responses:
+
+| OS | IP ID pattern | nmap label |
+|---|---|---|
+| Windows (most) | Sequential, incrementing | `I` (incremental) |
+| Linux 3.x+ | Per-socket hashed/random | `RI` or `RD` |
+| Old Linux / BSD | Global counter (truly sequential) | `I` |
+| Embedded | Often constant 0 or sequential | varies |
+
+Linux switched to per-socket hashed IDs at the kernel level (~3.x). This
+**cannot be changed per network namespace** without patching the kernel or
+replacing the TCP/IP stack with a userspace implementation.
+
+**Options:**
+1. **Accept the limitation** — the IP ID pattern is one of many signals; getting
+   TTL + window + timestamps right is already a very strong fingerprint match.
+2. **Userspace TCP proxy** (e.g., `lwIP` or a custom `nfqueue`-based responder)
+   that intercepts SYN packets and replies with forged ID sequences. High
+   complexity; requires `NFQUEUE` kernel module and `libnetfilter_queue`.
+
+> Phase 4 is **not recommended** for the near term. The complexity-to-realism
+> ratio is poor compared to Phases 1–3.
+
+---
+
+## Implementation Priority
+
+```
+Phase 1  ──────────────────────────────────  (implement next)
+  └─ 5 new sysctls in os_fingerprint.py
+  └─ No new files, no Docker changes
+  └─ Estimated effort: 30 min
+
+Phase 2  ──────────────────────────────────  (implement after Phase 1)
+  └─ templates/base/Dockerfile + entrypoint.sh
+  └─ os_fingerprint.py: add tcp_window field
+  └─ composer.py: pass env var to base container
+  └─ Estimated effort: 2–3 hours + tests
+
+Phase 3  ──────────────────────────────────  (nice to have)
+  └─ 2 more sysctls in os_fingerprint.py
+  └─ Estimated effort: 15 min (after Phase 1 infra exists)
+
+Phase 4  ──────────────────────────────────  (not recommended short-term)
+  └─ Requires kernel-level or userspace TCP stack work
+  └─ Estimated effort: days
+```
+
+---
+
+## Testing Strategy
+
+After each phase, validate with:
+
+```bash
+# Active OS fingerprint scan against a deployed decky
+sudo nmap -O --osscan-guess <decky_ip>
+
+# Passive fingerprinting (run on host while generating traffic to decky)
+sudo p0f -i <macvlan_interface> -p
+
+# Quick TTL + window check
+sudo nmap -sS --script banner <decky_ip>
+hping3 -S -p 22 <decky_ip>   # inspect TTL and window in reply
+```
+
+Expected outcomes by phase:
+
+| Check | Pre-Phase 1 | Post-Phase 1 | Post-Phase 2 |
+|---|---|---|---|
+| TTL | ✅ | ✅ | ✅ |
+| TCP timestamps | ❌ | ✅ | ✅ |
+| TCP window size | ❌ | ❌ | ✅ |
+| ICMP behavior | ❌ | ⚠️ | ⚠️ |
+| IP ID sequence | ❌ | ❌ | ❌ |
+| `nmap -O` family match | ⚠️ | ✅ | ✅ |
+| `p0f` match | ⚠️ | ⚠️ | ✅ |