diff --git a/ARCHITECTURE.md b/ARCHITECTURE.md
new file mode 100644
index 000000000..c0ff6b217
--- /dev/null
+++ b/ARCHITECTURE.md
@@ -0,0 +1,203 @@
+## d3
+
+[d3](http://d3js.org/) is the primary library used by iD. It is used for
+rendering the map data as well as many sorts of general DOM manipulation tasks
+for which jQuery would often be used.
+
+Notable features of d3 that are used by iD include
+[d3.xhr](https://github.com/mbostock/d3/wiki/Requests#wiki-d3_xhr), which is
+used to make the API requests to download data from openstreetmap.org and save
+changes;
+[d3.dispatch](https://github.com/mbostock/d3/wiki/Internals#wiki-d3_dispatch),
+which provides a callback-based [Observer
+pattern](http://en.wikipedia.org/wiki/Observer_pattern) between different
+parts of iD;
+[d3.geo.path](https://github.com/mbostock/d3/wiki/Geo-Paths#wiki-path), which
+generates SVG paths for lines and areas; and
+[d3.behavior.zoom](https://github.com/mbostock/d3/wiki/Zoom-Behavior#wiki-zoom),
+which implements map panning and zooming.
+
+## Core
+
+The iD *core* implements the OSM data types, a graph of OSM object's
+relationships to each other, and an undo/redo history for changes made during
+editing. It aims to be generic enough to be used by other JavaScript-based
+tools for OpenStreetMap.
+
+Briefly, the OSM data model includes three basic data types: nodes, ways, and
+relations. A _node_ is a point type, having a single geographic coordinate. A
+_way_ is an ordered list of nodes. And a _relation_ groups together nodes,
+ways, and other relations to provide free-form higher-level structures. Each
+of these three types has _tags_: an associative array of key-value pairs which
+describe the object.
+
+In iD, these three types are implemented by `iD.Node`, `iD.Way` and
+`iD.Relation`. These three classes inherit from a common base, `iD.Entity`
+(the only use of classical inheritance in iD). Generically, we refer to a
+node, way or relation as an _entity_.
+
+Every entity has an _ID_ either assigned by the OSM database, or, for an
+entity that is newly created, constructed as a proxy consisting of a negative
+numeral. IDs from the OSM database as treated as opaque strings; no
+[assumptions](http://lists.openstreetmap.org/pipermail/dev/2013-February/026495.html)
+are made of them other than that they can be compared for identity and do not
+begin with a minus sign (and thus will not conflict with proxy IDs). In fact,
+in the OSM database the three types of entities have separate ID spaces; a
+node can have the same ID as a way, for instance. Because it is useful to
+store heterogeneous entities in the same datastructure, iD ensures that every
+entity has a fully-unique ID by prefixing each OSM ID with the first letter of
+the entity type. For example, a way with OSM ID 123456 is represented as
+'w123456' within iD.
+
+iD entities are *immutable*: once constructed, an `Entity` object cannot
+change. Tags cannot be updated; nodes cannot be added or removed from ways,
+and so on. Immutability makes it easier to reason about the behavior of an
+entity: if your code has a reference to one, it is safe to store it and use it
+later, knowing that it cannot have been changed outside of your control. It
+also makes it possible to implement the entity graph (described below) as an
+efficient [persistent data
+structure](http://en.wikipedia.org/wiki/Persistent_data_structure). But
+obviously, iD is an editor, and must allow entities to change somehow. The
+solution is that all edits produce new copies of anything that changes. At the
+entity level, this takes the form of methods such as `iD.Node#move`, which
+returns a new node object that has the same ID and tags as the original, but a
+different coordinate. More generically, `iD.Entity#update` returns a new
+entity of the same type and ID as the original but with specified properties
+such as `nodes`, `tags`, or `members` replaced.
+
+Entities are related to one another: ways have many nodes and relations have
+many members. In order to render a map of a certain area, iD needs a
+datastructure to hold all the entities in that area and traverse these
+relationships. `iD.Graph` provides this functionality. The core of a graph is
+a map between IDs and the associated entities; given an ID, the graph can give
+you the entity. Like entities, a graph is immutable: adding, replacing, or
+removing an entity produces a new graph, and the original is unchanged.
+Because entities are immutable, the original and new graphs can share
+references to entities that have not changed, keeping memory use to a minimum.
+If you are familiar with how git works internally, this persistent data
+structure approach is very similar.
+
+The final component of the core is comprised of `iD.History` and
+`iD.Difference`, which track the changes made in an editing session and
+provide undo/redo capabilities. Here, the immutable nature of the core types
+really pays off: the history is a simple stack of graphs, each representing
+the state of the data at a particular point in editing. The graph at the top
+of the stack is the current state, off which all rendering is based. To undo
+the last change, this graph is popped off the stack, and the map is
+re-rendered based on the new top of the stack. Contrast this to a mutable
+graph as used in JOSM and Potlatch: every command that changes the graph must
+implement an equal and opposite undo command that restores the graph to the
+previous state.
+
+## Actions
+
+In iD, an _action_ is a function that accepts a graph as input and returns a
+modified graph as output. Actions typically need other inputs as well; for
+example, `iD.actions.DeleteNode` also requires the ID of a node to delete. The
+additional input is passed to the action's constructor:
+
+``` var action = iD.actions.DeleteNode('n123456'); // construct the action var
+newGraph = action(oldGraph); // apply the action ```
+
+iD provides actions for all the typical things an editor needs to do: add a
+new entity, split a way in two, connect the vertices of two ways together, and
+so on. In addition to performing the basic work needed to accomplish these
+things, an action typically contains a significant amount of logic for keeping
+the relationships between entities logical and consistent. For example, an
+action as apparently simple as `DeleteNode`, in addition to removing the node
+from the graph, needs to do two other things: remove the node from any ways in
+which it is a member (which in turn requires deleting parent ways that are
+left with just a single node), and removing it from any relations of which it
+is a member.
+
+As you can imagine, implementing all these details requires an expert
+knowledge of the OpenStreetMap data model. It is our hope that JavaScript
+based tools for OpenStreetMap can reuse the implementations provided by iD in
+other contexts, significantly reducing the work necessary to create a robust
+tool.
+
+## Modes
+
+With _modes_, we shift gears from abstract data types and algorithms to the
+parts of the architecture that implement the user interface for iD. Modes are
+manifested in the interface by the four buttons at the top left:
+
+![Mode buttons](img/modes.png)
+
+The modality of existing OSM editors runs the gamut from Potlatch 2, which is
+almost entirely modeless, to JOSM, which sports half a dozen modes out of the
+box and has many more provided by plugins. iD seeks a middle ground: too few
+modes can leave new users unsure where to start, while too many can be
+overwhelming.
+
+iD's user-facing modes consist of a base "Browse" mode, in which you can move
+around the map and select and edit entities, and three geometrically-oriented
+drawing modes: Point, Line, and Area. In the code, these are broken down a
+little bit more. There are separate modes for when an entity is selected
+(`iD.modes.Select`) versus when nothing is selected (`iD.modes.Browse`), and
+each of the geometric modes is split into one mode for starting to draw an
+object and one mode for continuing an existing object (with the exception of
+`iD.modes.AddPoint`, which is a single-step operation for obvious reasons).
+
+The code interface for each mode consists of a pair of methods: `enter` and
+`exit`. In the `enter` method, a mode sets up all the behavior that should be
+present when that mode is active. This typically means binding callbacks to
+DOM events that will be triggered on map elements, installing keybindings, and
+showing certain parts of the interface like the inspector in `Select` mode.
+The `exit` mode does the opposite, removing the behavior installed by the
+`enter` method. Together the two methods ensure that modes are self-contained
+and exclusive: each mode knows exactly the behavior that is specific to that
+mode, and exactly one mode's behavior is active at any time.
+
+## Behavior
+
+Certain behaviors are common to more than one mode. For example, iD indicates
+interactive map elements by drawing a halo around them when you hover over
+them, and this behavior is common to both the browse and draw modes. Instead
+of duplicating the code to implement this behavior in all these modes, we
+extract it to `iD.behavior.Hover`.
+
+_Behaviors_ take their inspiration from [d3's
+behaviors](https://github.com/mbostock/d3/wiki/Behaviors). Like d3's `zoom`
+and `drag`, each iD behavior is a function that takes as input a d3 selection
+(assumed to consist of a single element) and installs the DOM event bindings
+necessary to implement the behavior. The `Hover` behavior, for example,
+installs bindings for the `mouseover` and `mouseout` events that add and
+remove a `hover` class from map elements.
+
+Because certain behaviors are appropriate to some but not all modes, we need
+the ability to remove a behavior when entering a mode where it is not
+appropriate. (This is functionality [not yet
+provided](https://github.com/mbostock/d3/issues/894) by d3's own behaviors.)
+Each behavior implements an `off` function that "uninstalls" the behavior.
+This is very similar to the `exit` method of a mode, and in fact many modes do
+little else but uninstall behaviors in their `exit` methods.
+
+## Operations
+
+_Operations_ wrap actions, providing their user-interface: tooltips, key
+bindings, and the logic that determines whether an action can be validly
+performed given the current map state and selection. Each operation is
+constructed with the list of IDs which are currently selected and a `context`
+object which provides access to the history and other important parts of iD's
+internal state. After being constructed, an operation can be queried as to
+whether or not it should be made available (i.e., show up in the context menu)
+and if so, if it should be enabled.
+
+![Operations menu](img/operations.png)
+
+We make a distinction between availability and enabled state for the sake of
+learnability: most operations are available so long as an entity of the
+appropriate type is selected. Even if it remains disabled for other reasons
+(e.g. because you can't split a way on its start or end vertex), a new user
+can still learn that "this is something I can do to this type of thing", and a
+tooltip can provide an explanation of what that operation does and the
+conditions under which it is enabled.
+
+To execute an operation, call it as a function, with no arguments. The typical
+operation will perform the appropriate action, creating a new undo state in
+the history, and then enter the appropriate mode. For example,
+`iD.operations.Split` performs `iD.actions.Split`, then enters
+`iD.modes.Select` with the resulting ways selected.
+
+## Rendering and other UI
diff --git a/img/modes.png b/img/modes.png
new file mode 100644
index 000000000..6b5ec8603
Binary files /dev/null and b/img/modes.png differ
diff --git a/img/operations.png b/img/operations.png
new file mode 100644
index 000000000..c1e516b29
Binary files /dev/null and b/img/operations.png differ