diff --git a/include/ircd/json.h b/include/ircd/json.h
index 9ab7ea0b1..2f4f0e6c5 100644
--- a/include/ircd/json.h
+++ b/include/ircd/json.h
@@ -22,6 +22,40 @@
 #pragma once
 #define HAVE_IRCD_JSON_H
 
+//
+// The IRCd JSON subsystem is meant to be a fast, safe, and extremely
+// lightweight interface. We have taken a somewhat non-traditional approach
+// and it's important for the developer to understand a few things.
+//
+// Most JSON interfaces are functions to convert some JSON input to and from
+// text into native-machine state like JSON.parse() for JS, boost::ptree, etc.
+// For a parsing operation, they make a pass recursing over the entire text,
+// allocating native structures, copying data into them, indexing their keys,
+// and perhaps performing native-type conversions and checks to present the
+// user with a final tree of machine-state usable in their language. The
+// original input is then discarded.
+//
+// Instead, we are interested in having the ability to *compute directly over
+// JSON text* itself, and perform the allocating, indexing, copying and
+// converting entirely at the time and place of our discretion -- if ever.
+//
+// The core of this system is a robust and efficient abstract formal grammar
+// built with boost::spirit. The formal grammar provides a *proof of robust-
+// ness*: security vulnerabilities are more easily spotted by vetting this
+// grammar rather than laboriously tracing the program flow of an informal
+// handwritten parser.
+//
+// Next we have taught boost::spirit how to parse into std::string_view rather
+// than std::string. Parsing is now a composition of pointers into the original
+// string of JSON. No dynamic allocation ever takes place. No copying of data
+// ever takes place. IRCd can service an entire request from the original
+// network input with absolutely minimal requisite cost.
+//
+// The output side is also ambitious but probably a little more friendly to
+// the developer. We leverage boost::spirit here also providing *formally
+// proven* output safety. In other words, the grammar prevents exploits like
+// injecting and terminating JSON as it composes the output.
+//
 namespace ircd {
 namespace json {
 
@@ -47,6 +81,9 @@ enum type
 enum type type(const string_view &);
 enum type type(const string_view &, std::nothrow_t);
 
+using path = std::initializer_list<string_view>;
+std::ostream &operator<<(std::ostream &, const path &);
+
 } // namespace json
 } // namespace ircd
 
@@ -54,4 +91,26 @@ enum type type(const string_view &, std::nothrow_t);
 #include "json/object.h"
 #include "json/value.h"
 #include "json/index.h"
-// #include "json/parse.h"
+#include "json/parse.h"
+
+namespace ircd {
+
+using json::operator<<;
+
+} // namespace ircd
+
+inline std::ostream &
+ircd::json::operator<<(std::ostream &s, const path &p)
+{
+	auto it(std::begin(p));
+	if(it != std::end(p))
+	{
+		s << *it;
+		++it;
+	}
+
+	for(; it != std::end(p); ++it)
+		s << '.' << *it;
+
+	return s;
+}
diff --git a/include/ircd/json/parse.h b/include/ircd/json/parse.h
index 5528ff4d0..3bad555ea 100644
--- a/include/ircd/json/parse.h
+++ b/include/ircd/json/parse.h
@@ -52,6 +52,7 @@ struct parse
 		return std::tuple_size<tuple_type>();
 	}
 
+	parse() = default;
 	template<class R> parse(R &r, const json::object &obj);
 };
 
diff --git a/include/ircd/json/value.h b/include/ircd/json/value.h
index 7046eedc8..c19205ee7 100644
--- a/include/ircd/json/value.h
+++ b/include/ircd/json/value.h
@@ -25,6 +25,12 @@
 namespace ircd {
 namespace json {
 
+// The ircd::json::value is used if we have to keep state in machine-form
+// rather than directly computing JSON strings. This class ends up being useful
+// for recursive initializer_lists to compose JSON from machine values. It
+// is lightweight, consuming the space of two pointers which is the same size
+// as a string_view.
+//
 struct value
 {
 	union // xxx std::variant