diff --git a/.gitignore b/.gitignore
index ea412d7..fd1b471 100644
--- a/.gitignore
+++ b/.gitignore
@@ -17,4 +17,6 @@ dist/
 *:Zone.Identifier
 .playwright-mcp/
 # stray hardware/UI test screenshots
-/*.png
\ No newline at end of file
+/*.png
+# Vendor device SDKs (reference only — protocol captured in wiki, not committed)
+/dingtian/
diff --git a/wiki/concepts/threat-model.md b/wiki/concepts/threat-model.md
index 8164749..0f08486 100644
--- a/wiki/concepts/threat-model.md
+++ b/wiki/concepts/threat-model.md
@@ -35,3 +35,9 @@ The controls that actually address insider/operator fraud are different in kind:
 The same reframing recurs at the device layer: the [[uhppote-controller]]'s real problem is
 unauthenticated commands ([[uhppote-udp-protocol]]), addressed by detection
 ([[event-log-ingestion]]) or prevention ([[esp32-custom-controller]]).
+
+> **Direction shift:** the system is heading toward **fully unmanned operation** — no operator, no
+> booth ([[autonomous-direction]]). That removes the booth-operator as the *primary* adversary, but
+> swaps in **unattended-machine threats** (tailgating, plate spoofing, physical tampering, forced
+> entry). The append-only signed log + reconciliation controls carry over; the emphasis moves from
+> "catch the cashier" to "trust the automated record and detect tampering."
diff --git a/wiki/decisions/autonomous-direction.md b/wiki/decisions/autonomous-direction.md
new file mode 100644
index 0000000..ad00b64
--- /dev/null
+++ b/wiki/decisions/autonomous-direction.md
@@ -0,0 +1,45 @@
+---
+type: decision
+tags: [parking, direction, roadmap]
+sources: []
+updated: 2026-06-15
+status: open
+---
+
+# Project Direction: Toward Fully Autonomous (Unmanned)
+
+> Stated goal (2026-06-15): the system will evolve to **fully automatic operation — no human
+> operator, no booth at all**. Recorded because "unmanned" is an architectural force that shapes
+> several existing decisions, not just a feature.
+
+## What "unmanned" changes
+
+- **Threat model shift.** The original primary adversary was *"the legitimate operator at the
+  booth"* ([[threat-model]]). Remove the operator and that specific fraud vector (take cash → void
+  the record) largely disappears — but it's replaced by **unattended-machine threats**: tailgating,
+  plate spoofing/obscuring, physical tampering with a box nobody is watching, and forced entry.
+  The [[append-only-event-chain]] + [[reconciliation]] controls still apply; the emphasis moves
+  from "catch the cashier" to "trust the automated record + detect tampering."
+- **Host-in-the-loop entry becomes mandatory, not optional.** With no person to hand over a ticket
+  or wave a car through, the machine must own the whole flow: detect arrival → issue ticket / read
+  plate → open. This is exactly why the [[access-controller-button-flow]] blocker matters and why
+  a controller whose input does **not** auto-fire the relay (see [[dingtian-relay]]) is required.
+- **Reliability / fail-state get more critical** ([[fail-state-safety]]). No operator to recover a
+  stuck barrier or a trapped car ⇒ watchdogs, **exit-fails-open**, and hardware manual override
+  stop being nice-to-haves. Unattended uptime is a hard requirement.
+- **Payment goes unmanned.** Pay-station / pay-on-foot or in-lane unmanned terminal rather than a
+  booth P2PE + cash drawer — sharpens [[open-questions]] #3 toward the unmanned option (PCI scope
+  still kept out of the app via a certified terminal).
+- **Identity leans on automation.** Plate recognition ([[lpr-camera]]) and permit reads become the
+  primary identity sources, since there's no one to issue/inspect a paper ticket by hand.
+
+## Near-term stance
+
+Build for the unmanned target but don't over-engineer ahead of it. Current concrete step: the
+**[[dingtian-relay]]** controller over **HTTP** (device pushes input events to the host; host
+commands relays) — see [[dingtian-vs-mqtt]] for why HTTP over a message bus for now.
+
+## Open
+
+Lane topology, payment subsystem, and reconciliation channel ([[open-questions]]) should all be
+(re)evaluated through the **unmanned** lens before procurement.
diff --git a/wiki/decisions/dingtian-vs-mqtt.md b/wiki/decisions/dingtian-vs-mqtt.md
new file mode 100644
index 0000000..a29f74f
--- /dev/null
+++ b/wiki/decisions/dingtian-vs-mqtt.md
@@ -0,0 +1,46 @@
+---
+type: decision
+tags: [parking, decision, devices, transport]
+sources: []
+updated: 2026-06-15
+status: settled
+---
+
+# Transport for the Relay Controller: HTTP/UDP now, MQTT parked
+
+**Decision (2026-06-15): use direct HTTP + UDP for the [[dingtian-relay]] controller now. MQTT is
+deliberately skipped, but kept on the radar** for when the system scales.
+
+## Options the device supports
+
+The Dingtian relay board speaks several protocols: Dingtian string (UDP/TCP), Dingtian binary
+(UDP, optional multicast/password), **HTTP CGI**, **HTTP input-link push** (`input_link_url`),
+**Modbus** (RTU/TCP/ASCII), and **MQTT**.
+
+## Why not MQTT (yet)
+
+- **A broker is new infrastructure** on a deliberately **single-purpose hardened appliance**
+  ([[disk-os-hardening]]) — another service to install, secure, supervise, and keep alive.
+- **Extra failure mode on the critical path.** Today host→UDP→relay. MQTT inserts a broker on both
+  control and event paths; if it stalls, the lane stalls — and there are 3 processes to debug, not 2.
+- **Doesn't fit [[offline-first]] for this scale.** MQTT earns its keep with *many* devices/consumers
+  and intermittent links. Here it's **one host + a few devices on one isolated LAN, metres apart** —
+  request/response control + a single input event, no fleet.
+- **The device's MQTT input publish is periodic** ("default every 30 s"), so it's not even a clean
+  on-press event without relying on unverified on-change behaviour.
+
+## Why HTTP/UDP fits
+
+- **Relay control:** direct **UDP string protocol** (port 60001) — `11`=relay1 on, `21`=off,
+  `T1`=toggle, `11*`=jog/pulse. No deps, no broker.
+- **Input/button events:** the device's **`input_link_url`** can **HTTP POST to the host backend
+  when an input fires** — real push, device calls our existing Fastify server directly, no broker.
+  (Polling `00` status over UDP every ~50 ms is the self-contained fallback.)
+- Fewest moving parts; matches the local same-origin model already in use.
+
+## When to revisit MQTT
+
+If the system grows to **many lanes / many controllers**, or multiple subsystems (LPR, payment,
+signage) all need to share events, a broker becomes a worthwhile central event bus. That aligns
+with the [[autonomous-direction|unmanned]] roadmap at multi-lane scale — re-evaluate then. Until
+then, direct HTTP/UDP wins on simplicity and reliability.
diff --git a/wiki/entities/dingtian-relay.md b/wiki/entities/dingtian-relay.md
new file mode 100644
index 0000000..81b65bb
--- /dev/null
+++ b/wiki/entities/dingtian-relay.md
@@ -0,0 +1,46 @@
+---
+type: entity
+tags: [parking, hardware, access-control, relay]
+sources: []
+updated: 2026-06-15
+---
+
+# Dingtian Relay Controller
+
+A network **relay + input** board (the unit on hand is the **4-channel** variant: 4 relays + 4
+inputs). Chosen to drive the entry/exit lane because — unlike the [[uhppote-controller]] — its
+**inputs are independent of its relays**, which solves the [[access-controller-button-flow]]
+blocker (a button on an input does not auto-open a relay; the host decides).
+
+SDK: `dingtian/4ch/sdk_v2_0_0/` (programming manual, examples). MIT-compatible use; no vendor
+runtime needed.
+
+## ⚠️ The one gotcha: `input_link_relay`
+
+By **default the device links each input to auto-fire its matching relay** (`input_link_relay: 1`,
+`on_action_on: [[0],[1],…]` in the config) — i.e. the *same* auto-open problem as the UHPPOTE.
+The difference: **it is configurable.** Set `input_link_relay: 0` (or clear the action mappings)
+so an input only *reports* and the host commands the relay. **This config step is mandatory** for
+the ticket-first entry flow. See [[autonomous-direction]].
+
+## Protocol (Dingtian string — what we use)
+
+Transport options: UDP/TCP string, UDP binary, HTTP CGI, Modbus, MQTT. We use **HTTP + UDP** —
+see [[dingtian-vs-mqtt]].
+
+- **Relay control — UDP ASCII, port 60001:** `1`+ch = ON, `2`+ch = OFF, `T`+ch = toggle.
+  Pulse/jog `11*` (default 500 ms), delay `11:30` (30 s then off), flash `11F5`. `X` = all relays.
+  Intent-only pulse for a barrier = `pulseOpen` ([[barrier-not-a-door]]).
+- **Status / inputs — send `00`** → `「relays」:「inputs」:「count」`, e.g. **`0000:1111:4`** (4ch:
+  relays off, inputs high). `0` = OFF/Low, `1` = ON/High. Poll-based.
+- **Input push — `input_link_url`:** device **HTTP POSTs to a host URL on input change** — the
+  push path for button events without a broker.
+- **Discovery:** UDP multicast `224.0.2.11:60000`, send `\x05\xAA` (devices reply). Defaults:
+  IP `192.168.1.100`, UDP `60000` (binary) / `60001` (string).
+- Binary protocol (port 60000) adds optional **password** + multicast; bitmask relay/input maps.
+
+## Status
+
+Protocol understood from the SDK; **driver + on-hardware test not built yet**. Next: a `dingtian`
+relay driver in [[device-registry]] (UDP control + status parse) and the input-push endpoint,
+with `input_link_relay` disabled on the device. Needs the device IP + LAN to test.
diff --git a/wiki/index.md b/wiki/index.md
index f949e8e..454c827 100644
--- a/wiki/index.md
+++ b/wiki/index.md
@@ -38,6 +38,7 @@ Counts: 1 source · 14 entities · 10 concepts · 2 decision records.
 - [[wiegand]] — reader standard feeding the controller directly (autonomous permit-holder path).
 - [[lpr-camera]] — edge-AI plate recognition; host-side casual-identity source.
 - [[zkteco-controller]] — C3/inBio controller; aux-input path may enable host-in-the-loop (driver TBD).
+- [[dingtian-relay]] — relay+input board; inputs decoupled from relays → solves the button blocker (driver TBD).
 - [[bom]] — reference bill of materials (barrier, loops, controller, readers, payment, host, network).
 
 ## Concepts — foundational forces
@@ -74,3 +75,5 @@ Counts: 1 source · 14 entities · 10 concepts · 2 decision records.
 - [[standing-decisions]] — settled decisions (stack, platform, integrity, access control, readers).
 - [[open-questions]] — 7 open items (6 procurement + JWT key choice); ESP32 device auth deferred.
 - [[access-controller-button-flow]] — ⚠️ BLOCKER: UHPPOTE/ZKTeco on hand can't do ticket-first entry as wired.
+- [[autonomous-direction]] — roadmap: toward fully unmanned (no booth); reshapes threat model + fail-state.
+- [[dingtian-vs-mqtt]] — transport choice: direct HTTP/UDP now, MQTT parked until multi-lane scale.
diff --git a/wiki/log.md b/wiki/log.md
index c89dad6..fa36717 100644
--- a/wiki/log.md
+++ b/wiki/log.md
@@ -94,3 +94,14 @@ earlier "broadcast permission (EACCES)" note in [[device-discovery]] — the rea
 cause was the lib not enabling SO_BROADCAST for the global 255.255.255.255;
 documented the three verified broadcast gotchas + serialization. Added a `reference`
 page type to the schema; new "Dev environment" index section.
+
+## [2026-06-15] decision | Dingtian relay chosen; HTTP over MQTT; unmanned direction
+New relay+input controller on hand (Dingtian 4ch). Its inputs are decoupled from
+relays (configurable via input_link_relay) — solves the [[access-controller-button-flow]]
+blocker the UHPPOTE couldn't. Transport decision [[dingtian-vs-mqtt]]: direct
+HTTP/UDP now (UDP string for relay control on :60001; device input_link_url HTTP
+push for button events), MQTT skipped (broker = infra + failure mode + overkill at
+this scale) but kept for later multi-lane scale. Recorded the stated roadmap to
+**fully unmanned, no-booth** operation in [[autonomous-direction]] and its threat-model
+shift (operator-fraud → unattended-machine threats). New stub [[dingtian-relay]]
+with the full protocol from the SDK. Driver + on-hardware test still to build.