Internet Engineering Task Force (IETF)                         K. Davies
Request for Comments: 7940                                         ICANN
Category: Standards Track                                     A. Freytag
ISSN: 2070-1721                                              ASMUS, Inc.
                                                             August 2016

            Representing Label Generation Rulesets Using XML

Abstract

   This document describes a method of representing rules for validating
   identifier labels and alternate representations of those labels using
   Extensible Markup Language (XML).  These policies, known as "Label
   Generation Rulesets" (LGRs), are used for the implementation of
   Internationalized Domain Names (IDNs), for example.  The rulesets are
   used to implement and share that aspect of policy defining which
   labels and Unicode code points are permitted for registrations, which
   alternative code points are considered variants, and what actions may
   be performed on labels containing those variants.

Status of This Memo

   This is an Internet Standards Track document.

   This document is a product of the Internet Engineering Task Force
   (IETF).  It represents the consensus of the IETF community.  It has
   received public review and has been approved for publication by the
   Internet Engineering Steering Group (IESG).  Further information on
   Internet Standards is available in Section 2 of RFC 5741.

   Information about the current status of this document, any errata,
   and how to provide feedback on it may be obtained at
   http://www.rfc-editor.org/info/rfc7940.

Copyright Notice

   Copyright (c) 2016 IETF Trust and the persons identified as the
   document authors.  All rights reserved.

   This document is subject to BCP 78 and the IETF Trust's Legal
   Provisions Relating to IETF Documents
   (http://trustee.ietf.org/license-info) in effect on the date of
   publication of this document.  Please review these documents
   carefully, as they describe your rights and restrictions with respect
   to this document.  Code Components extracted from this document must
   include Simplified BSD License text as described in Section 4.e of
   the Trust Legal Provisions and are provided without warranty as
   described in the Simplified BSD License.

Table of Contents

   1.  Introduction  . . . . . . . . . . . . . . . . . . . . . . . .   4
   2.  Design Goals  . . . . . . . . . . . . . . . . . . . . . . . .   5
   3.  Normative Language  . . . . . . . . . . . . . . . . . . . . .   6
   4.  LGR Format  . . . . . . . . . . . . . . . . . . . . . . . . .   6
     4.1.  Namespace . . . . . . . . . . . . . . . . . . . . . . . .   6
     4.2.  Basic Structure . . . . . . . . . . . . . . . . . . . . .   6
     4.3.  Metadata  . . . . . . . . . . . . . . . . . . . . . . . .   7
       4.3.1.  The "version" Element . . . . . . . . . . . . . . . .   8
       4.3.2.  The "date" Element  . . . . . . . . . . . . . . . . .   8
       4.3.3.  The "language" Element  . . . . . . . . . . . . . . .   9   8
       4.3.4.  The "scope" Element . . . . . . . . . . . . . . . . .   9
       4.3.5.  The "description" Element . . . . . . . . . . . . . .  10
       4.3.6.  The "validity-start" and "validity-end" Elements  . .  10
       4.3.7.  The "unicode-version" Element . . . . . . . . . . . .  10
       4.3.8.  The "references" Element  . . . . . . . . . . . . . .  11
   5.  Code Points and Variants  . . . . . . . . . . . . . . . . . .  12
     5.1.  Sequences . . . . . . . . . . . . . . . . . . . . . . . .  13
     5.2.  Conditional Contexts  . . . . . . . . . . . . . . . . . .  14
     5.3.  Variants  . . . . . . . . . . . . . . . . . . . . . . . .  15
       5.3.1.  Basic Variants  . . . . . . . . . . . . . . . . . . .  15
       5.3.2.  The "type" Attribute  . . . . . . . . . . . . . . . .  17  16
       5.3.3.  Null Variants . . . . . . . . . . . . . . . . . . . .  18  17
       5.3.4.  Variants with Reflexive Mapping . . . . . . . . . . .  18
       5.3.5.  Conditional Variants  . . . . . . . . . . . . . . . .  19
     5.4.  Annotations . . . . . . . . . . . . . . . . . . . . . . .  21
       5.4.1.  The "ref" Attribute . . . . . . . . . . . . . . . . .  21
       5.4.2.  The "comment" Attribute . . . . . . . . . . . . . . .  22  21
     5.5.  Code Point Tagging  . . . . . . . . . . . . . . . . . . .  22
   6.  Whole Label and Context Evaluation  . . . . . . . . . . . . .  23  22
     6.1.  Basic Concepts  . . . . . . . . . . . . . . . . . . . . .  23  22
     6.2.  Character Classes . . . . . . . . . . . . . . . . . . . .  24  23
       6.2.1.  Declaring and Invoking Named Classes  . . . . . . . .  24
       6.2.2.  Tag-Based Classes . . . . . . . . . . . . . . . . . .  25  24
       6.2.3.  Unicode Property-Based Classes  . . . . . . . . . . .  25
       6.2.4.  Explicitly Declared Classes . . . . . . . . . . . . .  27  26
       6.2.5.  Combined Classes  . . . . . . . . . . . . . . . . . .  27
     6.3.  Whole Label and Context Rules . . . . . . . . . . . . . .  29
       6.3.1.  The "rule" Element  . . . . . . . . . . . . . . . . .  29
       6.3.2.  The Match Operators . . . . . . . . . . . . . . . . .  30
       6.3.3.  The "count" Attribute . . . . . . . . . . . . . . . .  31
       6.3.4.  The "name" and "by-ref" Attributes  . . . . . . . . .  32
       6.3.5.  The "choice" Element  . . . . . . . . . . . . . . . .  33
       6.3.6.  Literal Code Point Sequences  . . . . . . . . . . . .  33
       6.3.7.  The "any" Element . . . . . . . . . . . . . . . . . .  34
       6.3.8.  The "start" and "end" Elements  . . . . . . . . . . .  34
       6.3.9.  Example Context Rule from IDNA Specification  . . . .  35
     6.4.  Parameterized Context or When Rules . . . . . . . . . . .  35
       6.4.1.  The "anchor" Element  . . . . . . . . . . . . . . . .  36
       6.4.2.  The "look-behind" and "look-ahead" Elements . . . . .  37
       6.4.3.  Omitting the "anchor" Element . . . . . . . . . . . .  38
   7.  The "action" Element  . . . . . . . . . . . . . . . . . . . .  38
     7.1.  The "match" and "not-match" Attributes  . . . . . . . . .  39
     7.2.  Actions with Variant Type Triggers  . . . . . . . . . . .  39
       7.2.1.  The "any-variant", "all-variants", and
               "only-variants" Attributes  . . . . . . . . . . . . .  40
       7.2.2.  Example from Tables in the Style of RFC 3743  . . . .  42
     7.3.  Recommended Disposition Values  . . . . . . . . . . . . .  43
     7.4.  Precedence  . . . . . . . . . . . . . . . . . . . . . . .  43
     7.5.  Implied Actions . . . . . . . . . . . . . . . . . . . . .  44
     7.6.  Default Actions . . . . . . . . . . . . . . . . . . . . .  44
   8.  Processing a Label against an LGR . . . . . . . . . . . . . .  45
     8.1.  Determining Eligibility for a Label . . . . . . . . . . .  45
       8.1.1.  Determining Eligibility Using Reflexive Variant
               Mappings  . . . . . . . . . . . . . . . . . . . . . .  46
     8.2.  Determining Variants for a Label  . . . . . . . . . . . .  46
     8.3.  Determining a Disposition for a Label or Variant Label  .  47
     8.4.  Duplicate Variant Labels  . . . . . . . . . . . . . . . .  48
     8.5.  Checking Labels for Collision . . . . . . . . . . . . . .  49
   9.  Conversion to and from Other Formats  . . . . . . . . . . . .  49
   10. Media Type  . . . . . . . . . . . . . . . . . . . . . . . . .  50
   11. IANA Considerations . . . . . . . . . . . . . . . . . . . . .  50
     11.1.  Media Type Registration  . . . . . . . . . . . . . . . .  50
     11.2.  URN Registration . . . . . . . . . . . . . . . . . . . .  51
     11.3.  Disposition Registry . . . . . . . . . . . . . . . . . .  51
   12. Security Considerations . . . . . . . . . . . . . . . . . . .  52
     12.1.  LGRs Are Only a Partial Remedy for Problem Space . . . .  52
     12.2.  Computational Expense of Complex Tables  . . . . . . . .  53
   13. References  . . . . . . . . . . . . . . . . . . . . . . . . .  53
     13.1.  Normative References . . . . . . . . . . . . . . . . . .  53
     13.2.  Informative References . . . . . . . . . . . . . . . . .  54
   Appendix A.  Example Tables . . . . . . . . . . . . . . . . . . .  56
   Appendix B.  How to Translate Tables Based on RFC 3743 into the
                XML Format . . . . . . . . . . . . . . . . . . . . .  60
   Appendix C.  Indic Syllable Structure Example . . . . . . . . . .  65
     C.1.  Reducing Complexity . . . . . . . . . . . . . . . . . . .  67
   Appendix D.  RELAX NG Compact Schema  . . . . . . . . . . . . . .  68
   Acknowledgements  . . . . . . . . . . . . . . . . . . . . . . . .  77
   Authors' Addresses  . . . . . . . . . . . . . . . . . . . . . . .  78

1.  Introduction

   This document specifies a method of using Extensible Markup Language
   (XML) to describe Label Generation Rulesets (LGRs).  LGRs are
   algorithms used to determine whether, and under what conditions, a
   given identifier label is permitted, based on the code points it
   contains and their context.  These algorithms comprise a list of
   permissible code points, variant code point mappings, and a set of
   rules that act on the code points and mappings.  LGRs form part of an
   administrator's policies.  In deploying Internationalized Domain
   Names (IDNs), they have also been known as IDN tables or variant
   tables.

   There are other kinds of policies relating to labels that are not
   normally covered by LGRs and are therefore not necessarily
   representable by the XML format described here.  These include, but
   are not limited to, policies around trademarks, or prohibition of
   fraudulent or objectionable words.

   Administrators of the zones for top-level domain registries have
   historically published their LGRs using ASCII text or HTML.  The
   formatting of these documents has been loosely based on the format
   used for the Language Variant Table described in [RFC3743].
   [RFC4290] also provides a "model table format" that describes a
   similar set of functionality.  Common to these formats is that the
   algorithms used to evaluate the data therein are implicit or
   specified elsewhere.

   Through the first decade of IDN deployment, experience has shown that
   LGRs derived from these formats are difficult to consistently
   implement and compare, due to their differing formats.  A universal
   format, such as one using a structured XML format, will assist by
   improving machine readability, consistency, reusability, and
   maintainability of LGRs.

   When used to represent a simple list of permitted code points, the
   format is quite straightforward.  At the cost of some complexity in
   the resulting file, it also allows for an implementation of more
   sophisticated handling of conditional variants that reflects the
   known requirements of current zone administrator policies.

   Another feature of this format is that it allows many of the
   algorithms to be made explicit and machine implementable.  A
   remaining small set of implicit algorithms is described in this
   document to allow commonality in implementation.

   While the predominant usage of this specification is to represent IDN
   label policy, the format is not limited to IDN usage and may also be
   used for describing ASCII domain name label rulesets, or other types
   of identifier labels beyond those used for domain names.

2.  Design Goals

   The following goals informed the design of this format:

   o  The format needs to be implementable in a reasonably
      straightforward manner in software.

   o  The format should be able to be automatically checked for
      formatting errors, so that common mistakes can be caught.

   o  An LGR needs to be able to express the set of valid code points
      that are allowed for registration under a specific administrator's
      policies.

   o  An LGR needs to be able to express computed alternatives to a
      given identifier based on mapping relationships between code
      points, whether one-to-one or many-to-many.  These computed
      alternatives are commonly known as "variants".

   o  Variant code points should be able to be tagged with explicit
      dispositions or categories that can be used to support registry
      policy (such as whether to allocate the computed variant or to
      merely block it from usage or registration).

   o  Variants and code points must be able to be stipulated based on
      contextual information.  For example, some variants may only be
      applicable when they follow a certain code point or when the code
      point is displayed in a specific presentation form.

   o  The data contained within an LGR must be able to be interpreted
      unambiguously, so that independent implementations that utilize
      the contents will arrive at the same results.

   o  To the largest extent possible, policy rules should be able to be
      specified in the XML format without relying on hidden or built-in
      algorithms in implementations.

   o  LGRs should be suitable for comparison and reuse, such that one
      could easily compare the contents of two or more to see the
      differences, to merge them, and so on.

   o  As many existing IDN tables as practicable should be able to be
      migrated to the LGR format with all applicable interpretation
      logic retained.

   These requirements are partly derived from reviewing the existing
   corpus of published IDN tables, plus the requirements of ICANN's work
   to implement an LGR for the DNS root zone [LGR-PROCEDURE].  In
   particular, Section B of that document identifies five specific
   requirements for an LGR methodology.

   The syntax and rules in [RFC5892] and [RFC3743] were also reviewed.

   It is explicitly not the goal of this format to stipulate what code
   points should be listed in an LGR by a zone administrator.  Which
   registration policies are used for a particular zone are outside the
   scope of this memo.

3.  Normative Language

   The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
   "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this
   document are to be interpreted as described in [RFC2119].

4.  LGR Format

   An LGR is expressed as a well-formed XML document [XML] that conforms
   to the schema defined in Appendix D.

   As XML is case sensitive, an LGR must be authored with the correct
   casing.  For example, the XML element names MUST be in lowercase as
   described in this specification, and matching of attribute values is
   only performed in a case-sensitive manner.

   A document that is not well-formed, is non-conforming, or violates
   other constraints specified in this specification MUST be rejected.

4.1.  Namespace

   The XML Namespace URI is "urn:ietf:params:xml:ns:lgr-1.0".

   See Section 11.2 for more information.

4.2.  Basic Structure

   The basic XML framework of the document is as follows:

       <?xml version="1.0"?>
       <lgr xmlns="urn:ietf:params:xml:ns:lgr-1.0">
           ...
       </lgr>
   The "lgr" element contains up to three sub-elements or sections.
   First is an optional "meta" element that contains all metadata
   associated with the LGR, such as its authorship, what it is used for,
   implementation notes, and references.  This is followed by a required
   "data" element that contains the substantive code point data.
   Finally, an optional "rules" element contains information on
   contextual and whole-label evaluation rules, rules
   for evaluating labels, if any, along with "action" elements providing
   for the disposition of labels and computed variant labels.

       <?xml version="1.0"?>
       <lgr xmlns="urn:ietf:params:xml:ns:lgr-1.0">
           <meta>
               ...
           </meta>
           <data>
               ...
           </data>
           <rules>
               ...
           </rules>
       </lgr>

   A document MUST contain exactly one "lgr" element.  Each "lgr"
   element MUST contain zero or one "meta" element, exactly one "data"
   element, and zero or one "rules" element; and these three elements
   MUST be in that order.

   Some elements that are direct or nested child elements of the "rules"
   element MUST be placed in a specific relative order to other elements
   for the LGR to be valid.  An LGR that violates these constraints MUST
   be rejected.  In other cases, changing the ordering would result in a
   valid, but different, specification.

   In the following descriptions, required, non-repeating elements or
   attributes are generally not called out explicitly, in contrast to
   "OPTIONAL" ones, or those that "MAY" be repeated.  For attributes
   that take lists as values, the elements MUST be space-separated.

4.3.  Metadata

   The "meta" element expresses metadata associated with the LGR, and
   the element SHOULD be included so that the associated metadata are
   available as part of the LGR and cannot become disassociated.  The
   following subsections describe elements that may appear within the
   "meta" element.

   The "meta" element can be used to identify the author or relevant
   contact person, explain the intended usage of the LGR, and provide
   implementation notes as well as references.  Detailed metadata allow
   the LGR document to become self-documenting -- for example, if
   rendered in a human-readable format by an appropriate tool.

   Providing metadata pertaining to the date and version of the LGR is
   particularly encouraged to make it easier for interoperating
   consumers to ensure that they are using the correct LGR.

   With the exception of the "unicode-version" element, the data
   contained within is not required by software consuming the LGR in
   order to calculate valid labels or to calculate variants.  If
   present, the "unicode-version" element MUST be used by a consumer of
   the table to identify that it has the correct Unicode property data
   to perform operations on the table.  This ensures that possible
   differences in code point properties between editions of the Unicode
   Standard do not impact the product of calculations utilizing an LGR.

4.3.1.  The "version" Element

   The "version" element is OPTIONAL.  It is used to uniquely identify
   each version of the LGR.  No specific format is required, but it is
   RECOMMENDED that it be the decimal representation of a single
   positive integer, which is incremented with each revision of the
   file.

   An example of a typical first edition of a document:

       <version>1</version>

   The "version" element may have an OPTIONAL "comment" attribute.

       <version comment="draft">1</version>

4.3.2.  The "date" Element

   The OPTIONAL "date" element is used to identify the date the LGR was
   posted.  The contents of this element MUST be a valid ISO 8601
   "full-date" string as described in [RFC3339].

   Example of a date:

       <date>2009-11-01</date>

4.3.3.  The "language" Element

   Each OPTIONAL "language" element identifies a language or script for
   which the LGR is intended.  The value of the "language" element MUST
   be a valid language tag as described in [RFC5646].  The tag may refer
   to a script plus undefined language if the LGR is not intended for a
   specific language.

   Example of an LGR for the English language:

       <language>en</language>

   If the LGR applies to a script rather than a specific language, the
   "und" language tag SHOULD be used followed by the relevant script
   subtag from [RFC5646].  For example, for a Cyrillic script LGR:

       <language>und-Cyrl</language>

   If the LGR covers a set of multiple languages or scripts, the
   "language" element MAY be repeated.  However, for cases of a script-
   specific LGR exhibiting insignificant admixture of code points from
   other scripts, it is RECOMMENDED to use a single "language" element
   identifying the predominant script.  In the exceptional case of a
   multi-script LGR where no script is predominant, use Zyyy (Common):

       <language>und-Zyyy</language>

4.3.4.  The "scope" Element

   This OPTIONAL element refers to a scope, such as a domain, to which
   this policy is applied.  The "type" attribute specifies the type of
   scope being defined.  A type of "domain" means that the scope is a
   domain that represents the apex of the DNS zone to which the LGR is
   applied.  For that type, the content of the "scope" element MUST be a
   domain name written relative to the root zone, in presentation format
   with no trailing dot.  However, in the unique case of the DNS root
   zone, it is represented as ".".

       <scope type="domain">example.com</scope>

   There may be multiple "scope" tags used -- for example, to reflect a
   list of domains to which the LGR is applied.

   No other values of the "type" attribute are defined by this
   specification; however, this specification can be used for
   applications other than domain names.  Implementers of LGRs for
   applications other than domain names SHOULD define the scope
   extension grammar in an IETF specification or use XML namespaces to
   distinguish their scoping mechanism distinctly from the base LGR
   namespace.  An explanation of any custom usage of the scope in the
   "description" element is RECOMMENDED.

       <scope xmlns="http://example.com/ns/scope/1.0">
           ... content per alternate namespace ...
       </scope>

4.3.5.  The "description" Element

   The "description" element is an OPTIONAL, free-form element that
   contains any additional relevant description that is useful for the
   user in its interpretation.  Typically, this field contains
   authorship information, as well as additional context on how the LGR
   was formulated and how it applies, such as citations and references
   that apply to the LGR as a whole.

   This field should not be relied upon for providing instructions on
   how to parse or utilize the data contained elsewhere in the
   specification.  Authors of tables should expect that software
   applications that parse and use LGRs will not use the "description"
   element to condition the application of the LGR's data and rules.

   The element has an OPTIONAL "type" attribute, which refers to the
   Internet media type [RFC2045] of the enclosed data.  Typical types
   would be "text/plain" or "text/html".  The attribute SHOULD be a
   valid media type.  If supplied, it will be assumed that the contents
   are of that media type.  If the description lacks a "type" value, it
   will be assumed to be plain text ("text/plain").

4.3.6.  The "validity-start" and "validity-end" Elements

   The "validity-start" and "validity-end" elements are OPTIONAL
   elements that describe the time period from which the contents of the
   LGR become valid (are used in registry policy) and time when the
   contents of the LGR cease to be used, respectively.

   The dates MUST conform to the "full-date" format described in
   Section 5.6 of [RFC3339].

       <validity-start>2014-03-12</validity-start>

4.3.7.  The "unicode-version" Element

   Whenever an LGR depends on character properties from a given version
   of the Unicode Standard, the version number used in creating the LGR
   MUST be listed in the form x.y.z, where x, y, and z are positive
   decimal integers (see [Unicode-Versions]).  If any software
   processing the table does not have access to character property data
   of the requisite version, it MUST NOT perform any operations relating
   to whole-label evaluation relying on Unicode character properties
   (Section 6.2.3).

   The value of a given Unicode character property may change between
   versions of the Unicode Character Database [UAX44], unless such
   change has been explicitly disallowed in [Unicode-Stability].  It is
   RECOMMENDED to only reference properties defined as stable or
   immutable.  As an alternative to referencing the property, the
   information can be presented explicitly in the LGR.

       <unicode-version>6.3.0</unicode-version>

   It is not necessary to include a "unicode-version" element for LGRs
   that do not make use of Unicode character properties; however, it is
   RECOMMENDED.

4.3.8.  The "references" Element

   An LGR may define a list of references that are used to associate
   various individual elements in the LGR to one or more normative
   references.  A common use for references is to annotate that code
   points belong to an externally defined collection or standard or to
   give normative references for rules.

   References are specified in an OPTIONAL "references" element
   containing one or more "reference" elements, each with a unique "id"
   attribute.  It is RECOMMENDED that the "id" attribute be a zero-based
   integer; however, in addition to digits 0-9, it MAY contain uppercase
   letters A-Z, as well as a period, hyphen, colon, or underscore.  The
   value of each "reference" element SHOULD be the citation of a
   standard, dictionary, or other specification in any suitable format.
   In addition to an "id" attribute, a "reference" element MAY have a
   "comment" attribute for an optional free-form annotation.

       <references>
         <reference id="0">The Unicode Consortium.  The Unicode
           Standard, Version 8.0.0, (Mountain View, CA: The Unicode
           Consortium, 2015.  ISBN 978-1-936213-10-8)
           http://www.unicode.org/versions/Unicode8.0.0/</reference>
         <reference id="1">Big-5: Computer Chinese Glyph and Character
            Code Mapping Table, Technical Report C-26, 1984</reference>
         <reference id="2" comment="synchronized with Unicode 6.1">
            ISO/IEC
            10646:2012 3rd edition</reference>
         ...
       </references>
       ...
       <data>
         <char cp="0620" ref="0 2" />
         ...
       </data>
   A reference is associated with an element by using its id as part of
   an optional "ref" attribute (see Section 5.4.1).  The "ref" attribute
   may be used with many kinds of elements in the "data" or "rules"
   sections of the LGR, most notably those defining code points,
   variants, and rules.  However, a "ref" attribute may not occur in
   certain kinds of elements, including references to named character
   classes or rules.  See below for the description of these elements.

5.  Code Points and Variants

   The bulk of an LGR is a description of which set of code points is
   eligible for a given label.  For rulesets that perform operations
   that result in potential variants, the code point-level relationships
   between variants need to also be described.

   The code point data is collected within the "data" element.  Within
   this element, a series of "char" and "range" elements describe
   eligible code points or ranges of code points, respectively.
   Collectively, these are known as the repertoire.

   Discrete permissible code points or code point sequences (see
   Section 5.1) are declared with a "char" element.  Here is a minimal
   example declaration for a single code point, with the code point
   value given in the "cp" attribute:

       <char cp="002D"/>

   As described below, a full declaration for a "char" element, whether
   or not it is used for a single code point or for a sequence (see
   Section 5.1), may have optional child elements defining variants.
   Both the "char" and "range" elements can take a number of optional
   attributes for conditional inclusion, commenting, cross-referencing,
   and character tagging, as described below.

   Ranges of permissible code points may be declared with a "range"
   element, as in this minimal example:

       <range first-cp="0030" last-cp="0039"/>

   The range is inclusive of the first and last code points.  Any
   additional attributes defined for a "range" element act as if applied
   to each code point within.  A "range" element has no child elements.

   It is always possible to substitute a list of individually specified
   code points for a "range" element.  The reverse is not necessarily
   the case.  Whenever such a substitution is possible, it makes no
   difference in processing the data.  Tools reading or writing the LGR
   format are free to aggregate sequences of consecutive code points of
   the same properties into "range" elements.

   Code points MUST be represented according to the standard Unicode
   convention but without the prefix "U+": they are expressed in
   uppercase hexadecimal and are zero-padded to a minimum of 4 digits.

   The rationale for not allowing other encoding formats, including
   native Unicode encoding in XML, is explored in [UAX42].  The XML
   conventions used in this format, such as element and attribute names,
   mirror this document where practical and reasonable to do so.  It is
   RECOMMENDED to list all "char" elements in ascending order of the
   "cp" attribute.  Not doing so makes it unnecessarily difficult for
   authors and reviewers to check for errors, such as duplications, or
   to review and compare against listing of code points in other
   documents and specifications.

   All "char" elements in the "data" section MUST have distinct "cp"
   attributes.  The "range" elements MUST NOT specify code point ranges
   that overlap either another range or any single code point "char"
   elements.  An LGR that defines the same code point more than once by
   any combination of "char" or "range" elements MUST be rejected.

5.1.  Sequences

   A sequence of two or more code points may be specified in an LGR --
   for example, when defining the source for n:m variant mappings.
   Another use of sequences would be in cases when the exact sequence of
   code points is required to occur in order for the constituent
   elements to be eligible, such as when some code point is only
   eligible when preceded or followed by a certain code point.  The
   following would define the eligibility of the MIDDLE DOT (U+00B7)
   only when both preceded and followed by the LATIN SMALL LETTER L
   (U+006C):

       <char cp="006C 00B7 006C" comment="Catalan middle dot"/>

   All sequences defined this way must be distinct, but sub-sequences
   may be defined.  Thus, the sequence defined here may coexist with
   single code point definitions such as:

       <char cp="006C" />

   As an alternative to using sequences to define a required context, a
   "char" or "range" element may specify a conditional context using an
   optional "when" attribute as described below in Section 5.2.  Using a
   conditional context is more flexible because a context is not limited
   to a specific sequence of code points.  In addition, using a context
   allows the choice of specifying either a prohibited or a required
   context.

5.2.  Conditional Contexts

   A conditional context is specified by a rule that must be satisfied
   (or, alternatively, must not be satisfied) for a code point in a
   given label, often at a particular location in a label.

   To specify a conditional context, either a "when" or "not-when"
   attribute may be used.  The value of each "when" or "not-when"
   attribute is a context rule as described below in Section 6.3.  This
   rule can be a rule evaluating the whole label or a parameterized
   context rule.  The context condition is met when the rule specified
   in the "when" attribute is matched or when the rule in the "not-when"
   attribute fails to match.  It is an error to reference a rule that is
   not actually defined in the "rules" element.

   A parameterized context rule (see Section 6.4) defines the context
   immediately surrounding a given code point; unlike a sequence, the
   context is not limited to a specific fixed code point but, for
   example, may designate any member of a certain character class or a
   code point that has a certain Unicode character property.

   Given a suitable definition of a parameterized context rule named
   "follows-virama", this example specifies that a ZERO WIDTH JOINER
   (U+200D) is restricted to immediately follow any of several code
   points classified as virama:

       <char cp="200D" when="follows-virama" />

   For a complete example, see Appendix A.

   In contrast, a whole label rule (see Section 6.3) specifies a
   condition to be met by the entire label -- for example, that it must
   contain at least one code point from a given script anywhere in the
   label.  In the following example, no digit from either range may
   occur in a label that mixes digits from both ranges:

       <data>
          <range first-cp="0660" last-cp="0669" not-when="mixed-digits"
                 tag="arabic-indic-digits" />
          <range first-cp="06F0" last-cp="06F9" not-when="mixed-digits"
                 tag="extended-arabic-indic-digits" />
       </data>

   (See Section 6.3.9 for an example of the "mixed-digits" rule.)
   The OPTIONAL "when" or "not-when" attributes are mutually exclusive.
   They MAY be applied to both "char" and "range" elements in the "data"
   element, including "char" elements defining sequences of code points,
   as well as to "var" elements (see Section 5.3.5).

   If a label contains one or more code points that fail to satisfy a
   conditional context, the label is invalid (see Section 7.5).  For
   variants, the conditional context restricts the definition of the
   variant to the case where the condition is met.  Outside the
   specified context, a variant is not defined.

5.3.  Variants

   Most LGRs typically only determine simple code point eligibility, and
   for them, the elements described so far would be the only ones
   required for their "data" section.  Others additionally specify a
   mapping of code points to other code points, known as "variants".
   What constitutes a variant code point is a matter of policy and
   varies for each implementation.  The following examples are intended
   to demonstrate the syntax; they are not necessarily typical.

5.3.1.  Basic Variants

   Variant code points are specified using one of more "var" elements as
   children of a "char" element.  The target mapping is specified using
   the "cp" attribute.  Other, optional attributes for the "var" element
   are described below.

   For example, to map LATIN SMALL LETTER V (U+0076) as a variant of
   LATIN SMALL LETTER U (U+0075):

       <char cp="0075">
           <var cp="0076"/>
       </char>

   A sequence of multiple code points can be specified as a variant of a
   single code point.  For example, the sequence of LATIN SMALL LETTER O
   (U+006F) then LATIN SMALL LETTER E (U+0065) might hypothetically be
   specified as a variant for a LATIN SMALL LETTER O WITH DIAERESIS
   (U+00F6) as follows:

       <char cp="00F6">
           <var cp="006F 0065"/>
       </char>

   The source and target of a variant mapping may both be sequences but
   not ranges.

   If the source of one mapping is a prefix sequence of the source for
   another, both variant mappings will be considered at the same
   location in the input label when generating permuted variant labels.
   If poorly designed, an LGR containing such an instance of a prefix
   relation could generate multiple instances of the same variant label
   for the same original label, but with potentially different
   dispositions.  Any duplicate variant labels encountered MUST be
   treated as an error (see Section 8.4).

   The "var" element specifies variant mappings in only one direction,
   even though the variant relation is usually considered symmetric;
   that is, if A is a variant of B, then B should also be a variant of
   A.  The format requires that the inverse of the variant be given
   explicitly to fully specify symmetric variant relations in the LGR.
   This has the beneficial side effect of making the symmetry explicit:

       <char cp="006F 0065">
           <var cp="00F6"/>
       </char>

   Variant relations are normally not only symmetric but also
   transitive.  If A is a variant of B and B is a variant of C, then A
   is also a variant of C.  As with symmetry, these transitive relations
   are only part of the LGR if spelled out explicitly.  Implementations
   that require an LGR to be symmetric and transitive should verify this
   mechanically.

   All variant mappings are unique.  For a given "char" element, all
   "var" elements MUST have a unique combination of "cp", "when", and
   "not-when" attributes.  It is RECOMMENDED to list the "var" elements
   in ascending order of their target code point sequence.  (For "when"
   and "not-when" attributes, see Section 5.3.5.)

5.3.2.  The "type" Attribute

   Variants may be tagged with an OPTIONAL "type" attribute.  The value
   of the "type" attribute may be any non-empty value not starting with
   an underscore and not containing spaces.  This value is used to
   resolve the disposition of any variant labels created using a given
   variant.  (See Section 7.2.)

   By default, the values of the "type" attribute directly describe the
   target policy status (disposition) for a variant label that was
   generated using a particular variant, with any variant label being
   assigned a disposition corresponding to the most restrictive variant
   type.  Several conventional disposition values are predefined below
   in Section 7.  Whenever these values can represent the desired
   policy, they SHOULD be used.

       <char cp="767C">
           <var cp="53D1" type="allocatable"/>
           <var cp="5F42" type="blocked"/>
           <var cp="9AEA" type="blocked"/>
           <var cp="9AEE" type="blocked"/>
       </char>

   By default, if a variant label contains any instance of one of the
   variants of type "blocked", the label would be blocked, but if it
   contained only instances of variants to be allocated, it could be
   allocated.  See the discussion about implied actions in Section 7.6.

   The XML format for the LGR makes the relation between the values of
   the "type" attribute on variants and the resulting disposition of
   variant labels fully explicit.  See the discussion in Section 7.2.
   Making this relation explicit allows a generalization of the "type"
   attribute from directly reflecting dispositions to a more
   differentiated intermediate value that is then used in the resolution
   of label disposition.  Instead of the default action of applying the
   most restrictive disposition to the entire label, such a generalized
   resolution can be used to achieve additional goals, such as limiting
   the set of allocatable variant labels or implementing other policies
   found in existing LGRs (see, for example, Appendix B).

   Because variant mappings MUST be unique, it is not possible to define
   the same variant for the same "char" element with different "type"
   attributes (however, see Section 5.3.5).

5.3.3.  Null Variants

   A null variant is a variant string that maps to no code point.  This
   is used when a particular code point sequence is considered
   discretionary in the context of a whole label.  To specify a null
   variant, use an empty "cp" attribute.  For example, to mark a string
   with a ZERO WIDTH NON-JOINER (U+200C) to the same string without the
   ZERO WIDTH NON-JOINER:

       <char cp="200C">
           <var cp=""/>
       </char>

   This is useful in expressing the intent that some code points in a
   label are to be mapped away when generating a canonical variant of
   the label.  However, in tables that are designed to have symmetric
   variant mappings, this could lead to combinatorial explosion if not
   handled carefully.

   The symmetric form of a null variant is expressed as follows:

       <char cp="">
           <var cp="200C" type="invalid" />
       </char>

   A "char" element with an empty "cp" attribute MUST specify at least
   one variant mapping.  It is strongly RECOMMENDED to use a type of
   "invalid" or equivalent when defining variant mappings from null
   sequences, so that variant mappings from null sequences are removed
   in variant label generation (see Section 5.3.2).

5.3.4.  Variants with Reflexive Mapping

   At first glance, there seems to be no call for adding variant
   mappings for which source and target code points are the same -- that
   is, for which the mapping is reflexive, or, in other words, an
   identity mapping.  Yet, such reflexive mappings occur frequently in
   LGRs that follow [RFC3743].

   Adding a "var" element allows both a type and a reference id to be
   specified for it.  While the reference id is not used in processing,
   the type of the variant can be used to trigger actions.  In permuting
   the label to generate all possible variants, the type associated with
   a reflexive variant mapping is applied to any of the permuted labels
   containing the original code point.

   In the following example, let's assume that the goal is to allocate
   only those labels that contain a variant that is considered
   "preferred" in some way.  As defined in the example, the code point
   U+3473 exists both as a variant of U+3447 and as a variant of itself
   (reflexive mapping).  Assuming an original label of "U+3473 U+3447",
   the permuted variant "U+3473 U+3473" would consist of the reflexive
   variant of U+3473 followed by a variant of U+3447.  Given the variant
   mappings as defined here, the types for both of the variant mappings
   used to generate that particular permutation would have the value
   "preferred":

       <char cp="3447" ref="0">
         <var cp="3473" type="preferred" ref="1 3" />
       </char>
       <char cp="3473" ref="0">
         <var cp="3447" type="blocked" ref="1 3" />
         <var cp="3473" type="preferred" ref="0" />
       </char>

   Having established the variant types in this way, a set of actions
   could be defined that return a disposition of "allocatable" or
   "activated" for a label consisting exclusively of variants with type
   "preferred", for example.  (For details on how to define actions
   based on variant types, see Section 7.2.1.)

   In general, using reflexive variant mappings in this manner makes it
   possible to calculate disposition values using a uniform approach for
   all labels, whether they consist of mapped variant code points,
   original code points, or a mixture of both.  In particular, the
   dispositions for two otherwise identical labels may differ based on
   which variant mappings were executed in order to generate each of
   them.  (For details on how to generate variants and evaluate
   dispositions, see Section 8.)

   Another useful convention that uses reflexive variants is described
   below in Section 7.2.1.

5.3.5.  Conditional Variants

   Fundamentally, variants are mappings between two sequences of code
   points.  However, in some instances, for a variant relationship to
   exist, some context external to the code point sequence must also be
   considered.  For example, a positional context may determine whether
   two code point sequences are variants of each other.

   An example of that are Arabic code points, which can have different
   forms based on position, with some code points sharing forms, thus
   making them variants in the positions corresponding to those forms.
   Such positional context cannot be solely derived from the code point
   by itself, as the code point would be the same for the various forms.

   As described in Section 5.2, an OPTIONAL "when" or "not-when"
   attribute may be given for any "var" element to specify required or
   prohibited contextual conditions under which the variant is defined.

   Assuming that the "rules" element contains suitably defined rules for
   "arabic-isolated" and "arabic-final", the following example shows how
   to mark ARABIC LETTER ALEF WITH WAVY HAMZA BELOW (U+0673) as a
   variant of ARABIC LETTER ALEF WITH HAMZA BELOW (U+0625), but only
   when it appears in its isolated or final forms:

       <char cp="0625">
           <var cp="0673" when="arabic-isolated"/>
           <var cp="0673" when="arabic-final"/>
       </char>

   While a "var" element MUST NOT contain multiple conditions (it is
   only allowed a single "when" or "not-when" attribute), multiple "var"
   elements using the same mapping MAY be specified with different
   "when" or "not-when" attributes.  The combination of mapping and
   conditional context defines a unique variant.

   For each variant label, care must be taken to ensure that at most one
   of the contextual conditions is met for variants with the same
   mapping; otherwise, duplicate variant labels would be created for the
   same input label.  Any such duplicate variant labels MUST be treated
   as an error; see Section 8.4.

   Two contexts may be complementary, as in the following example, which
   shows ARABIC LETTER TEH MARBUTA (U+0629) as a variant of ARABIC
   LETTER HEH (U+0647), but with two different types.

       <char cp="0647" >
         <var cp="0629" not-when="arabic-final" type="blocked" />
         <var cp="0629" when="arabic-final" type="allocatable" />
       </char>

   The intent is that a label that uses U+0629 instead of U+0647 in a
   final position should be considered essentially the same label and,
   therefore, allocatable to the same entity, while the same
   substitution in a non-final position leads to labels that are
   different, but considered confusable, so that either one, but not
   both, should be delegatable.

   For symmetry, the reverse mappings must exist and must agree in their
   "when" or "not-when" attributes.  However, symmetry does not apply to
   the other attributes.  For example, these are potential reverse
   mappings for the above:

       <char cp="0629" >
         <var cp="0647" not-when="arabic-final" type="allocatable" />
         <var cp="0647" when="arabic-final" type="allocatable" />
       </char>

   Here, both variants have the same "type" attribute.  While it is
   tempting to recognize that, in this instance, the "when" and "not-
   when" attributes are complementary; therefore, between them they
   cover every single possible context, it is strongly RECOMMENDED to
   use the format shown in the example that makes the symmetry easily
   verifiable by parsers and tools.  (The same applies to entries
   created for transitivity.)

   Arabic is an example of a script for which such conditional variants
   have been implemented based on the joining contexts for Arabic code
   points.  The mechanism defined here supports other forms of
   conditional variants that may be required by other scripts.

5.4.  Annotations

   Two attributes, the "ref" and "comment" attributes, can be used to
   annotate individual elements in the LGR.  They are ignored in
   machine-processing of the LGR.  The "ref" attribute is intended for
   formal annotations and the "comment" attribute for free-form
   annotations.  The latter can be applied more widely.

5.4.1.  The "ref" Attribute

   Reference information MAY optionally be specified by a "ref"
   attribute consisting of a space-delimited sequence of reference
   identifiers (see Section 4.3.8).

       <char cp="5220" ref="0">
           <var cp="5220" ref="5"/>
           <var cp="522A" ref="2 3"/>
       </char>

   This facility is typically used to give source information for code
   points or variant relations.  This information is ignored when
   machine-processing an LGR.  If applied to a range, the "ref"
   attribute applies to every code point in the range.  All reference
   identifiers MUST be from the set declared in the "references" element
   (see Section 4.3.8).  It is an error to repeat a reference identifier
   in the same "ref" attribute.  It is RECOMMENDED that identifiers be
   listed in ascending order.

   In addition to "char", "range", and "var" elements in the "data"
   section, a "ref" attribute may be present for a number of element
   types contained in the "rules" element as described below: actions
   and literals ("char" inside a rule), as well as for definitions of
   rules and classes, but not for references to named character classes
   or rules using the "by-ref" attribute defined below.  (The use of the
   "by-ref" and "ref" attributes is mutually exclusive.)  None of the
   elements in the metadata take a "ref" attribute; to provide
   additional information, use the "description" element instead.

5.4.2.  The "comment" Attribute

   Any "char", "range", or "variant" element in the "data" section may
   contain an OPTIONAL "comment" attribute.  The contents of a "comment"
   attribute are free-form plain text.  Comments are ignored in machine
   processing of the table.  "comment" attributes MAY also be placed on
   all elements in the "rules" section of the document, such as actions
   and match operators, as well as definitions of classes and rules, but
   not on child elements of the "class" element.  Finally, in the
   metadata, only the "version" and "reference" elements MAY have
   "comment" attributes (to match the syntax in [RFC3743]).

5.5.  Code Point Tagging

   Typically, LGRs are used to explicitly designate allowable code
   points, where any label that contains a code point not explicitly
   listed in the LGR is considered an ineligible label according to the
   ruleset.

   For more-complex registry rules, there may be a need to discern one
   or more subsets of code points.  This can be accomplished by applying
   an OPTIONAL "tag" attribute to "char" or "range" elements that are
   child elements of the "data" element.  By collecting code points that
   share the same tag value, character classes may be defined (see
   Section 6.2.2) that can then be used in parameterized context or
   whole label rules (see Section 6.3.2).

   Each "tag" attribute MAY contain multiple values separated by
   white space.  A tag value is an identifier that may also include
   certain punctuation marks, such as a colon.  Formally, it MUST
   correspond to the XML 1.0 Nmtoken (Name token) production (see [XML]
   Section 2.3).  It is an error to duplicate a value within the same
   "tag" attribute.  A "tag" attribute for a "range" element applies to
   all code points in the range.  Because code point sequences are not
   proper members of a set of code points, a "tag" attribute MUST NOT be
   present in a "char" element defining a code point sequence.

6.  Whole Label and Context Evaluation

6.1.  Basic Concepts

   The "rules" element contains the specification of both context-based
   and whole label rules.  Collectively, these are known as Whole Label
   Evaluation (WLE) rules (Section 6.3).  The "rules" element also
   contains the character classes (Section 6.2) that they depend on, and
   any actions (Section 7) that assign dispositions to labels based on
   rules or variant mappings.

   A whole label rule is applied to the whole label.  It is used to
   validate both original labels and any variant labels computed from
   them.

   A rule implementing a conditional context as discussed in Section 5.2
   does not necessarily apply to the whole label but may be specific to
   the context around a single code point or code point sequence.
   Certain code points in a label sometimes need to satisfy context-
   based rules -- for example, for the label to be considered valid, or
   to satisfy the context for a variant mapping (see the description of
   the "when" attribute in Section 6.4).

   For example, if a rule is referenced in the "when" attribute of a
   variant mapping, it is used to describe the conditional context under
   which the particular variant mapping is defined to exist.

   Each rule is defined in a "rule" element.  A rule may contain the
   following as child elements:

   o  literal code points or code point sequences

   o  character classes, which define sets of code points to be used for
      context comparisons

   o  context operators, which define when character classes and
      literals may appear

   o  nested rules, whether defined in place or invoked by reference

   Collectively, these are called "match operators" and are listed in
   Section 6.3.2.  An LGR containing rules or match operators that

   1.  are incorrectly defined or nested,

   2.  have invalid attributes, or

   3.  have invalid or undefined attribute values

   MUST be rejected.  Note that not all of the constraints defined here
   are validated by the schema.

6.2.  Character Classes

   Character classes are sets of characters that often share a
   particular property.  While they function like sets in every way,
   even supporting the usual set operators, they are called "character
   classes" here in a nod to the use of that term in regular expression
   syntax.  (This also avoids confusion with the term "character set" in
   the sense of character encoding.)

   Character classes can be specified in several ways:

   o  by defining the class via matching a tag in the code point data.
      All characters with the same "tag" attribute are part of the same
      class;
   o  by referencing a value of one of the Unicode character properties
      defined in the Unicode Character Database;

   o  by explicitly listing all the code points in the class; or

   o  by defining the class as a set combination of any number of other
      classes.

6.2.1.  Declaring and Invoking Named Classes

   A character class has an OPTIONAL "name" attribute consisting of a
   single identifier not containing spaces.  All names for classes must
   be unique.  If the "name" attribute is omitted, the class is
   anonymous and exists only inside the rule or combined class where it
   is defined.  A named character class is defined independently and can
   be referenced by name from within any rules or as part of other
   character class definitions.

       <class name="example" comment="an example class definition">
           0061 4E00
       </class>
       ...
       <rule>
           <class by-ref="example" />
       </rule>

   An empty "class" element with a "by-ref" attribute is a reference to
   an existing named class.  The "by-ref" attribute MUST NOT be used in
   the same "class" element with any of these attributes: "name",
   "from-tag", "property", or "ref".  The "name" attribute MUST be
   present if and only if the class is a direct child element of the
   "rules" element.  It is an error to reference a named class for which
   the definition has not been seen.

6.2.2.  Tag-Based Classes

   The "char" or "range" elements that are child elements of the "data"
   element MAY contain a "tag" attribute that consists of one or more
   space-separated tag values; for example:

       <char cp="0061" tag="letter lower"/>
       <char cp="4E00" tag="letter"/>

   This defines two tags for use with code point U+0061, the tag
   "letter" and the tag "lower".  Use

       <class name="letter" from-tag="letter" />
       <class name="lower" from-tag="lower" />
   to define two named character classes, "letter" and "lower",
   containing all code points with the respective tags, the first with
   0061 and 4E00 as elements, and the latter with 0061 but not 4E00 as
   an element.  The "name" attribute may be omitted for an anonymous
   in-place definition of a nested, tag-based class.

   Tag values are typically identifiers, with the addition of a few
   punctuation symbols, such as a colon.  Formally, they MUST correspond
   to the XML 1.0 Nmtoken production.  While a "tag" attribute may
   contain a list of tag values, the "from-tag" attribute MUST always
   contain a single tag value.

   If the document contains no "char" or "range" elements with a
   corresponding tag, the character class represents the empty set.
   This is valid, to allow a common "rules" element to be shared across
   files.  However, it is RECOMMENDED that implementations allow for a
   warning to ensure that referring to an undefined tag in this way is
   intentional.

6.2.3.  Unicode Property-Based Classes

   A class is defined in terms of Unicode properties by giving the
   Unicode property alias and the property value or property value
   alias, separated by a colon.

       <class name="virama" property="ccc:9" />

   The example above selects all code points for which the Unicode
   Canonical Combining Class (ccc) value is 9.  This value of the ccc is
   assigned to all code points that encode viramas.

   Unicode property values MUST be designated via a composite of the
   attribute name and value as defined for the property value in
   [UAX42], separated by a colon.  Loose matching of property values and
   names as described in [UAX44] is not appropriate for an XML schema
   and is not supported; it is likewise not supported in the XML
   representation [UAX42] of the Unicode Character Database itself.

   A property-based class MAY be anonymous, or, when defined as an
   immediate child of the "rules" element, it MAY be named to relate a
   formal property definition to its usage, such as the use of the value
   9 for ccc to designate a virama (or halant) in various scripts.

   Unicode properties may, in principle, change between versions of the
   Unicode Standard.  However, the values assigned for a given version
   are fixed.  If Unicode properties are used, a Unicode version MUST be
   declared in the "unicode-version" element in the header.  (Note: Some
   Unicode properties are by definition stable across versions and do
   not change once assigned; see [Unicode-Stability].)

   All implementations processing LGR files SHOULD provide support for
   the following minimal set of Unicode properties:

   o  General Category (gc)

   o  Script (sc)

   o  Canonical Combining Class (ccc)

   o  Bidi Class (bc)

   o  Arabic Joining Type (jt)

   o  Indic Syllabic Category (InSC)

   o  Deprecated (Dep)

   The short name for each property is given in parentheses.

   If a program that is using an LGR to determine the validity of a
   label encounters a property that it does not support, it MUST abort
   with an error.

6.2.4.  Explicitly Declared Classes

   A class of code points may also be declared by listing all code
   points that are members of the class.  This is useful when tagging
   cannot be used because code points are not listed individually as
   part of the eligible set of code points for the given LGR -- for
   example, because they only occur in code point sequences.

   To define a class in terms of an explicit list of code points, use a
   space-separated list of hexadecimal code point values:

       <class name="abcd">0061 0062 0063 0064</class>

   This defines a class named "abcd" containing the code points for
   characters "a", "b", "c", and "d".  The ordering of the code points
   is not material, but it is RECOMMENDED to list them in ascending
   order; not doing so makes it unnecessarily difficult for users to
   detect errors such as duplicates or to compare and review these
   classes against other specifications.

   In a class definition, ranges of code points are represented by a
   hexadecimal start and end value separated by a hyphen.  The following
   declaration is equivalent to the preceding:

       <class name="abcd">0061-0064</class>

   Range and code point declarations can be freely intermixed:

       <class name="abcd">0061 0062-0063 0064</class>

   The contents of a class differ from a repertoire in that the latter
   MAY contain sequences as elements, while the former MUST NOT.
   Instead, they closely resemble character classes as found in regular
   expressions.

6.2.5.  Combined Classes

   Classes may be combined using operators for set complement, union,
   intersection, difference (elements of the first class that are not in
   the second), and symmetric difference (elements in either class but
   not both).  Because classes fundamentally function like sets, the
   union of several character classes is itself a class, for example.

   +-------------------+----------------------------------------------+
   | Logical Operation | Example                                      |
   +-------------------+----------------------------------------------+
   | Complement        | <complement><class by-ref="xxx"></complement>|
   +-------------------+----------------------------------------------+
   | Union             | <union>                                      |
   |                   |    <class by-ref="class-1"/>                 |
   |                   |    <class by-ref="class-2"/>                 |
   |                   |    <class by-ref="class-3"/>                 |
   |                   | </union>                                     |
   +-------------------+----------------------------------------------+
   | Intersection      | <intersection>                               |
   |                   |    <class by-ref="class-1"/>                 |
   |                   |    <class by-ref="class-2"/>                 |
   |                   | </intersection>                              |
   +-------------------+----------------------------------------------+
   | Difference        | <difference>                                 |
   |                   |    <class by-ref="class-1"/>                 |
   |                   |    <class by-ref="class-2"/>                 |
   |                   | </difference>                                |
   +-------------------+----------------------------------------------+
   | Symmetric         | <symmetric-difference>                       |
   | Difference        |    <class by-ref="class-1"/>                 |
   |                   |    <class by-ref="class-2"/>                 |
   |                   | </symmetric-difference>                      |
   +-------------------+----------------------------------------------+

                               Set Operators

   The elements from this table may be arbitrarily nested inside each
   other, subject to the following restriction: a "complement" element
   MUST contain precisely one "class" or one of the operator elements,
   while an "intersection", "symmetric-difference", or "difference"
   element MUST contain precisely two, and a "union" element MUST
   contain two or more of these elements.

   An anonymous combined class can be defined directly inside a rule or
   any of the match operator elements that allow child elements (see
   Section 6.3.2) by using the set combination as the outer element.

       <rule>
           <union>
               <class by-ref="xxx"/>
               <class by-ref="yyy"/>
           </union>
       </rule>
   The example shows the definition of an anonymous combined class that
   represents the union of classes "xxx" and "yyy".  There is no need to
   wrap this union inside another "class" element, and, in fact, set
   combination elements MUST NOT be nested inside a "class" element.

   Lastly, to create a named combined class that can be referenced in
   other classes or in rules as <class by-ref="xxxyyy"/>, add a "name"
   attribute to the set combination element -- for example,
   <union name="xxxyyy" /> -- and place it at the top level immediately
   below the "rules" element (see Section 6.2.1).

       <rules>
          <union name="xxxyyy">
              <class by-ref="xxx"/>
              <class by-ref="yyy"/>
          </union>
            ...
       </rules>

   Because (as for ordinary sets) a combination of classes is itself a
   class, no matter by what combinations of set operators a combined
   class is created, a reference to it always uses the "class" element
   as described in Section 6.2.1.  That is, a named class is always
   referenced via an empty "class" element using the "by-ref" attribute
   containing the name of the class to be referenced.

6.3.  Whole Label and Context Rules

   Each rule comprises a series of matching operators that must be
   satisfied in order to determine whether a label meets a given
   condition.  Rules may reference other rules or character classes
   defined elsewhere in the table.

6.3.1.  The "rule" Element

   A matching rule is defined by a "rule" element, the child elements of
   which are one of the match operators from Section 6.3.2.  In
   evaluating a rule, each child element is matched in order.  "rule"
   elements MAY be nested inside each other and inside certain match
   operators.

   A simple rule to match a label where all characters are members of
   some class called "preferred-codepoint":

       <rule name="preferred-label">
           <start />
           <class by-ref="preferred-codepoint" count="1+"/>
           <end />
       </rule>

   Rules are paired with explicit and implied actions, triggering these
   actions when a rule matches a label.  For example, a simple explicit
   action for the rule shown above would be:

       <action disp="allocatable" match="preferred-label" />

   The rule in this example would have the effect of setting the policy
   disposition for a label made up entirely of preferred code points to
   "allocatable".  Explicit actions are further discussed in Section 7
   and implicit actions in Section 7.5.  Another use of rules is in
   defining conditional contexts for code points and variants as
   discussed in Sections 5.2 and 5.3.5.

   A rule that is an immediate child element of the "rules" element MUST
   be named using a "name" attribute containing a single identifier
   string with no spaces.  A named rule may be incorporated into another
   rule by reference and may also be referenced by an "action" element,
   "when" attribute, or "not-when" attribute.  If the "name" attribute
   is omitted, the rule is anonymous and MUST be nested inside another
   rule or match operator.

6.3.2.  The Match Operators

   The child elements of a rule are a series of match operators, which
   are listed here by type and name and with a basic example or two.

   +------------+-------------+------------------------------------+
   | Type       | Operator    | Examples                           |
   +------------+-------------+------------------------------------+
   | logical    | any         | <any />                            |
   |            +-------------+------------------------------------+
   |            | choice      | <choice>                           |
   |            |             |  <rule by-ref="alternative1"/>     |
   |            |             |  <rule by-ref="alternative2"/>     |
   |            |             | </choice>                          |
   +--------------------------+------------------------------------+
   | positional | start       | <start />                          |
   |            +-------------+------------------------------------+
   |            | end         | <end />                            |
   +--------------------------+------------------------------------+
   | literal    | char        | <char cp="0061 0062 0063" />       |
   +--------------------------+------------------------------------+
   | set        | class       | <class by-ref="class1" />          |
   |            |             | <class>0061 0064-0065</class>      |
   +--------------------------+------------------------------------+
   | group      | rule        | <rule by-ref="rule1" />            |
   |            |             | <rule><any /></rule>               |
   +--------------------------+------------------------------------+
   | contextual | anchor      | <anchor />                         |
   |            +-------------+------------------------------------+
   |            | look-ahead  | <look-ahead><any /></look-ahead>   |
   |            +-------------+------------------------------------+
   |            | look-behind | <look-behind><any /></look-behind> |
   +--------------------------+------------------------------------+

                              Match Operators

   Any element defining an anonymous class can be used as a match
   operator, including any of the set combination operators (see
   Section 6.2.5) as well as references to named classes.

   All match operators shown as empty elements in the Examples column of
   the table above do not support child elements of their own;
   otherwise, match operators MAY be nested.  In particular, anonymous
   "rule" elements can be used for grouping.

6.3.3.  The "count" Attribute

   The OPTIONAL "count" attribute, when present, specifies the minimally
   required or maximal permitted number of times a match operator is
   used to match input.  If the "count" attribute is

   n    the match operator matches the input exactly n times, where n is
        1 or greater.

   n+   the match operator matches the input at least n times, where n
        is 0 or greater.

   n:m  the match operator matches the input at least n times, where n
        is 0 or greater, but matches the input up to m times in total,
        where m > n.  If m = n and n > 0, the match operator matches the
        input exactly n times.

   If there is no "count" attribute, the match operator matches the
   input exactly once.

   In matching, greedy evaluation is used in the sense defined for
   regular expressions: beyond the required number or times, the input
   is matched as many times as possible, but not so often as to prevent
   a match of the remainder of the rule.

   A "count" attribute MUST NOT be applied to any element that contains
   a "name" attribute but MAY be applied to operators such as "class"
   that declare anonymous classes (including combined classes) or invoke
   any predefined classes by reference.  The "count" attribute MUST NOT
   be applied to any "class" element, or element defining a combined
   class, when it is nested inside a combined class.

   A "count" attribute MUST NOT be applied to match operators of type
   "start", "end", "anchor", "look-ahead", or "look-behind" or to any
   operators, such as "rule" or "choice", that contain a nested instance
   of them.  This limitation applies recursively and irrespective of
   whether a "rule" element containing these nested instances is
   declared in place or used by reference.

   However, the "count" attribute MAY be applied to any other instances
   of either an anonymous "rule" element or a "choice" element,
   including those instances nested inside other match operators.  It
   MAY also be applied to the elements "any" and "char", when used as
   match operators.

6.3.4.  The "name" and "by-ref" Attributes

   Like classes (see Section 6.2.1), rules declared as immediate child
   elements of the "rules" element MUST be named using a unique "name"
   attribute, and all other instances MUST NOT be named.  Anonymous
   rules and classes or references to named rules and classes can be
   nested inside other match operators by reference.

   To reference a named rule or class inside a rule or match operator,
   use a "rule" or "class" element with an OPTIONAL "by-ref" attribute
   containing the name of the referenced element.  It is an error to
   reference a rule or class for which the complete definition has not
   been seen.  In other words, it is explicitly not possible to define
   recursive rules or class definitions.  The "by-ref" attribute
   MUST NOT appear in the same element as the "name" attribute or in an
   element that has any child elements.

   The example shows several named classes and a named rule referencing
   some of them by name.

       <class name="letter" property="gc:L"/>
       <class name="combining-mark" property="gc:M"/>
       <class name="digit" property="gc:Nd" />
       <rule name="letter-grapheme">
          <class by-ref="letter" count="1+"/>
          <class by-ref="combining-mark" count="0+"/>
       </rule>

6.3.5.  The "choice" Element

   The "choice" element is used to represent a list of two or more
   alternatives:

       <rule name="ldh">
          <choice count="1+">
              <class by-ref="letter"/>
              <class by-ref="digit"/>
              <char cp="002D" comment="literal HYPHEN"/>
          </choice>
       </rule>

   Each child element of a "choice" element represents one alternative.
   The first matching alternative determines the match for the "choice"
   element.  To express a choice where an alternative itself consists of
   a sequence of elements, the sequence must be wrapped in an anonymous
   rule.

6.3.6.  Literal Code Point Sequences

   A literal code point sequence matches a single code point or a
   sequence.  It is defined by a "char" element, with the code point or
   sequence to be matched given by the "cp" attribute.  When used as a
   literal, a "char" element MAY contain a "count" attribute in addition
   to the "cp" attribute and OPTIONAL "comment" or "ref" attributes.  No
   other attributes or child elements are permitted.

6.3.7.  The "any" Element

   The "any" element is an empty element that matches any single code
   point.  It MAY have a "count" attribute.  For an example, see
   Section 6.3.9.

   Unlike a literal, the "any" element MUST NOT have a "ref" attribute.

6.3.8.  The "start" and "end" Elements

   To match the beginning or end of a label, use the "start" or "end"
   element.  An empty label would match this rule:

       <rule name="empty-label">
           <start/>
           <end/>
       </rule>

   Conceptually, whole label rules evaluate the label as a whole, but in
   practice, many rules do not actually need to be specified to match
   the entire label.  For example, to express a requirement of not
   starting a label with a digit, a rule needs to describe only the
   initial part of a label.

   This example uses the previously defined rules, together with "start"
   and "end" elements, to define a rule that requires that an entire
   label be well-formed.  For this example, that means that it must
   start with a letter and that it contains no leading digits or
   combining marks nor combining marks placed on digits.

       <rule name="leading-letter" >
         <start />
         <rule by-ref="letter-grapheme" count="1"/>
         <choice count="0+">
           <rule by-ref="letter-grapheme" count="0+"/>
           <class by-ref="digit" count="0+"/>
         </choice>
         <end />
       </rule>

   Each "start" or "end" element occurs at most once in a rule, except
   if nested inside a "choice" element in such a way that in matching
   each alternative at most one occurrence of each is encountered.
   Otherwise, the result is an error, as is any case where a "start" or
   "end" element is not encountered as the first or last element to be
   matched, respectively, in matching a rule.  "start" and "end"
   elements are empty elements that do not have a "count" attribute or
   any other attribute other than "comment".  It is an error for any
   match operator enclosing a nested "start" or "end" element to have a
   "count" attribute.

6.3.9.  Example Context Rule from IDNA Specification

   This is an example of the WLE rule from [RFC5892] forbidding the
   mixture of the Arabic-Indic and extended Arabic-Indic digits in the
   same label.  It is implemented as a whole label rule associated with
   the code point ranges using the "not-when" attribute, which defines
   an impermissible context.  The example also demonstrates several
   instances of the use of anonymous rules for grouping.

       <data>
          <range first-cp="0660" last-cp="0669" not-when="mixed-digits"
                 tag="arabic-indic-digits" />
          <range first-cp="06F0" last-cp="06F9" not-when="mixed-digits"
                 tag="extended-arabic-indic-digits" />
       </data>
       <rules>
          <rule name="mixed-digits">
             <choice>
               <rule>
                   <class from-tag="arabic-indic-digits"/>
                   <any count="0+"/>
                   <class from-tag="extended-arabic-indic-digits"/>
                </rule>
                <rule>
                   <class from-tag="extended-arabic-indic-digits"/>
                   <any count="0+"/>
                   <class from-tag="arabic-indic-digits"/>
                </rule>
             </choice>
          </rule>
       </rules>

   As specified in the example, a label containing a code point from
   either of the two digit ranges is invalid for any label matching the
   "mixed-digits" rule, that is, any time that a code point from the
   other range is also present.  Note that invalidating the label is not
   the same as invalidating the definition of the "range" elements; in
   particular, the definition of the tag values does not depend on the
   "when" attribute.

6.4.  Parameterized Context or When Rules

   To recap: When a rule is intended to provide a context for evaluating
   the validity of a code point or variant mapping, it is invoked by the
   "when" or "not-when" attributes described in Section 5.2.  For "char"
   and "range" elements, an action implied by a context rule always has
   a disposition of "invalid" whenever the rule given by the "when"
   attribute is not matched (see Section 7.5).  Conversely, a "not-when"
   attribute results in a disposition of "invalid" whenever the rule is
   matched.  When a rule is used in this way, it is called a context or
   "when" rule.

   The example in the previous section shows a whole label rule used as
   a context rule, essentially making the whole label the context.  The
   next sections describe several match operators that can be used to
   provide a more specific specification of a context, allowing a
   parameterized context rule.  See Section 7 for an alternative method
   of defining an invalid disposition for a label not matching a whole
   label rule.

6.4.1.  The "anchor" Element

   Such parameterized context rules are rules that contain a special
   placeholder represented by an "anchor" element.  As each When Rule is
   evaluated, if an "anchor" element is present, it is replaced by a
   literal corresponding to the "cp" attribute of the element containing
   the "when" (or "not-when") attribute.  The match to the "anchor"
   element must be at the same position in the label as the code point
   or variant mapping triggering the When Rule.

   For example, the Greek lower numeral sign is invalid if not
   immediately preceding a character in the Greek script.  This is most
   naturally addressed with a parameterized When Rule using
   "look-ahead":

       <char cp="0375" when="preceding-greek"/>
       ...
       <class name="greek-script" property="sc:Grek"/>
       <rule name="preceding-greek">
           <anchor/>
           <look-ahead>
               <class by-ref="greek-script"/>
           </look-ahead>
       </rule>

   In evaluating this rule, the "anchor" element is treated as if it was
   replaced by a literal

       <char cp="0375"/>

   but only the instance of U+0375 at the given position is evaluated.
   If a label had two instances of U+0375 with the first one matching
   the rule and the second not, then evaluating the When Rule MUST
   succeed for the first instance and fail for the second.

   Unlike other rules, rules containing an "anchor" element MUST only be
   invoked via the "when" or "not-when" attributes on code points or
   variants; otherwise, their "anchor" elements cannot be evaluated.
   However, it is possible to invoke rules not containing an "anchor"
   element from a "when" or "not-when" attribute.  (See Section 6.4.3.)

   The "anchor" element is an empty element, with no attributes
   permitted except "comment".

6.4.2.  The "look-behind" and "look-ahead" Elements

   Context rules use the "look-behind" and "look-ahead" elements to
   define context before and after the code point sequence matched by
   the "anchor" element.  If the "anchor" element is omitted, neither
   the "look-behind" nor the "look-ahead" element may be present in a
   rule.

   Here is an example of a rule that defines an "initial" context for an
   Arabic code point:

       <class name="transparent" property="jt:T"/>
       <class name="right-joining" property="jt:R"/>
       <class name="left-joining" property="jt:L"/>
       <class name="dual-joining" property="jt:D"/>
       <class name="non-joining" property="jt:U"/>
       <rule name="Arabic-initial">
         <look-behind>
           <choice>
             <start/>
             <rule>
               <class by-ref="transparent" count="0+"/>
               <class by-ref="non-joining"/>
             </rule>
           </choice>
         </look-behind>
         <anchor/>
         <look-ahead>
           <class by-ref="transparent" count="0+" />
           <choice>
             <class by-ref="right-joining" />
             <class by-ref="dual-joining" />
           </choice>
         </look-ahead>
       </rule>
   A "when" rule (or context rule) is a named rule that contains any
   combination of "look-behind", "anchor", and "look-ahead" elements, in
   that order.  Each of these elements occurs at most once, except if
   nested inside a "choice" element in such a way that in matching each
   alternative at most one occurrence of each is encountered.
   Otherwise, the result is undefined.  None of these elements takes a
   "count" attribute, nor does any enclosing match operator; otherwise,
   the result is undefined.  If a context rule contains a "look-ahead"
   or "look-behind" element, it MUST contain an "anchor" element.  If,
   because of a "choice" element, a required anchor is not actually
   encountered, the results are undefined.

6.4.3.  Omitting the "anchor" Element

   If the "anchor" element is omitted, the evaluation of the context
   rule is not tied to the position of the code point or sequence
   associated with the "when" attribute.

   According to [RFC5892], the Katakana middle dot is invalid in any
   label not containing at least one Japanese character anywhere in the
   label.  Because this requirement is independent of the position of
   the middle dot, the rule does not require an "anchor" element.

       <char cp="30FB" when="japanese-in-label"/>
       <rule name="japanese-in-label">
           <union>
               <class property="sc:Hani"/>
               <class property="sc:Kata"/>
               <class property="sc:Hira"/>
           </union>
       </rule>

   The Katakana middle dot is used only with Han, Katakana, or Hiragana.
   The corresponding When Rule requires that at least one code point in
   the label be in one of these scripts, but the position of that code
   point is independent of the location of the middle dot; therefore, no
   anchor is required.  (Note that the Katakana middle dot itself is of
   script Common, that is, "sc:Zyyy".)

7.  The "action" Element

   The purpose of an action is to assign a disposition to a label in
   response to being triggered by the label meeting a specified
   condition.  Often, the action simply results in blocking or
   invalidating a label that does not match a rule.  An example of an
   action invalidating a label because it does not match a rule named
   "leading-letter" is as follows:

       <action disp="invalid" not-match="leading-letter"/>

   If an action is to be triggered on matching a rule, a "match"
   attribute is used instead.  Actions are evaluated in the order that
   they appear in the XML file.  Once an action is triggered by a label,
   the disposition defined in the "disp" attribute is assigned to the
   label and no other actions are evaluated for that label.

   The goal of the LGR is to identify all labels and variant labels and
   to assign them disposition values.  These dispositions are then fed
   into a further process that ultimately implements all aspects of
   policy.  To allow this specification to be used with the widest range
   of policies, the permissible values for the "disp" attribute are
   neither defined nor restricted.  Nevertheless, a set of commonly used
   disposition values is RECOMMENDED.  (See Section 7.3.)

7.1.  The "match" and "not-match" Attributes

   An OPTIONAL "match" or "not-match" attribute specifies a rule that
   must be matched or not matched as a condition for triggering an
   action.  Only a single rule may be named as the value of a "match" or
   "not-match" attribute.  Because rules may be composed of other rules,
   this restriction to a single attribute value does not impose any
   limitation on the contexts that can trigger an action.

   An action MUST NOT contain both a "match" and a "not-match"
   attribute, and the value of either attribute MUST be the name of a
   previously defined rule; otherwise, the document MUST be rejected.
   An action without any attributes is triggered by all labels
   unconditionally.  For a very simple LGR, the following action would
   allocate all labels that match the repertoire:

       <action disp="allocatable" />

   Since rules are evaluated for all labels, whether they are the
   original label or computed by permuting the defined and valid variant
   mappings for the label's code points, actions based on matching or
   not matching a rule may be triggered for both original and variant
   labels, but the rules are not affected by the disposition attributes
   of the variant mappings.  To trigger any actions based on these
   dispositions requires the use of additional optional attributes for
   actions described next.

7.2.  Actions with Variant Type Triggers

7.2.1.  The "any-variant", "all-variants", and "only-variants"
        Attributes

   An action may contain one of the OPTIONAL attributes "any-variant",
   "all-variants", or "only-variants" defining triggers based on variant
   types.  The permitted value for these attributes consists of one or
   more variant type values, separated by spaces.  These MAY include
   type values that are not used in any "var" element in the LGR.  When
   a variant label is generated, these variant type values are compared
   to the set of type values on the variant mappings used to generate
   the particular variant label (see Section 8).

   Any single match may trigger an action that contains an "any-variant"
   attribute, while for an "all-variants" or "only-variants" attribute,
   the variant type for all variant code points must match one or
   several of the type values specified in the attribute to trigger the
   action.  There is no requirement that the entire list of variant type
   values be matched, as long as all variant code points match at least
   one of the values.

   An "only-variants" attribute will trigger the action only if all code
   points of the variant label have variant mappings from the original
   code points.  In other words, the label contains no original code
   points other than those with a reflexive mapping (see Section 5.3.4).

       <char cp="0078" comment="x">
           <var cp="0078" type="allocatable" comment="reflexive" />
           <var cp="0079" type="blocked" />
       </char>
       <char cp="0079" comment="y">
           <var cp="0078" type="allocatable" />
       </char>
       ...
       <action disp="blocked" any-variant="blocked" />
       <action disp="allocatable" only-variants="allocatable" />
       <action disp="some-disp" any-variant="allocatable" />

   In the example above, the label "xx" would have variant labels "xx",
   "xy", "yx", and "yy".  The first action would result in blocking any
   variant label containing "y", because the variant mapping from "x" to
   "y" is of type "blocked", triggering the "any-variant" condition.
   Because in this example "x" has a reflexive variant mapping to itself
   of type "allocatable", the original label "xx" has a reflexive
   variant "xx" that would trigger the "only-variants" condition on the
   second action.

   A label "yy" would have the variants "xy", "yx", and "xx".  Because
   the variant mapping from "y" to "x" is of type "allocatable" and a
   mapping from "y" to "y" is not defined, the labels "xy" and "yx"
   trigger the "any-variant" condition on the third label.  The variant
   "xx", being generated using the mapping from "y" to "x" of type
   "allocatable", would trigger the "only-variants" condition on the
   section action.  As there is no reflexive variant "yy", the original
   label "yy" cannot trigger any variant type triggers.  However, it
   could still trigger an action defined as matching or not matching a
   rule.

   In each action, one variant type trigger may be present by itself or
   in conjunction with an attribute matching or not matching a rule.  If
   variant triggers and rule-matching triggers are used together, the
   label MUST "match" or respectively "not-match" the specified rule AND
   satisfy the conditions on the variant type values given by the
   "any-variant", "all-variants", or "only-variants" attribute.

   A useful convention combines the "any-variant" trigger with reflexive
   variant mappings (Section 5.3.4).  This convention is used, for
   example, when multiple LGRs are defined within the same registry and
   for overlapping repertoire.  In some cases, the delegation of a label
   from one LGR must prohibit the delegation of another label in some
   other LGR.  This can be done using a variant of type "blocked" as in
   this example from an Armenian LGR, where the Armenian, Latin, and
   Cyrillic letters all look identical:

       <char cp="0570" comment="ARMENIAN SMALL LETTER HO">
         <var cp="0068" type="blocked" comment="LATIN SMALL LETTER H" />
         <var cp="04BB" type="blocked"
              comment="CYRILLIC SMALL LETTER SHHA" />
       </char>

   The issue is that the target code points for these two variants are
   both outside the Armenian repertoire.  By using a reflexive variant
   with the following convention:

       <char cp="0068" comment="not part of repertoire">
         <var cp="0068" type="out-of-repertoire-var"
              comment="reflexive mapping" />
         <var cp="04BB" type="blocked" />
         <var cp="0570" type="blocked" />
       </char>
         ...

   and associating this with an action of the form:

       <action disp="invalid" any-variant="out-of-repertoire-var" />
   it is possible to list the symmetric and transitive variant mappings
   in the LGR even where they involve out-of-repertoire code points.  By
   associating the action shown with the special type for these
   reflexive mappings, any original labels containing one or more of the
   out-of-repertoire code points are filtered out, just as if these code
   points had not been listed in the LGR in the first place.
   Nevertheless, they do participate in the permutation of variant
   labels for n-repertoire labels (Armenian in the example), and these
   permuted variants can be used to detect collisions with out-of-
   repertoire labels (see Section 8).

7.2.2.  Example from Tables in the Style of RFC 3743

   This section gives an example of using variant type triggers,
   combined with variants with reflexive mappings (Section 5.3.4), to
   achieve LGRs that implement tables like those defined according to
   [RFC3743] where the goal is to allow as variants only labels that
   consist entirely of simplified or traditional variants, in addition
   to the original label.

   This example assumes an LGR where all variants have been given
   suitable "type" attributes of "blocked", "simplified", "traditional",
   or "both", similar to the ones discussed in Appendix B.  Given such
   an LGR, the following example actions evaluate the disposition for
   the variant label:

       <action disp="blocked" any-variant="blocked" />
       <action disp="allocatable" only-variants="simplified both" />
       <action disp="allocatable" only-variants="traditional both" />
       <action disp="blocked" all-variants="simplified traditional" />
       <action disp="allocatable" />

   The first action matches any variant label for which at least one of
   the code point variants is of type "blocked".  The second matches any
   variant label for which all of the code point variants are of type
   "simplified" or "both" -- in other words, an all-simplified label.
   The third matches any label for which all variants are of type
   "traditional" or "both" -- that is, all traditional.  These two
   actions are not triggered by any variant labels containing some
   original code points, unless each of those code points has a variant
   defined with a reflexive mapping (Section 5.3.4).

   The final two actions rely on the fact that actions are evaluated in
   sequence and that the first action triggered also defines the final
   disposition for a variant label (see Section 7.4).  They further rely
   on the assumption that the only variants with type "both" are also
   reflexive variants.

   Given these assumptions, any remaining simplified or traditional
   variants must then be part of a mixed label and so are blocked; all
   labels surviving to the last action are original code points only
   (that is, the original label).  The example assumes that an original
   label may be a mixed label; if that is not the case, the disposition
   for the last action would be set to "blocked".

   There are exceptions where the assumption on reflexive mappings made
   above does not hold, so this basic scheme needs some refinements to
   cover all cases.  For a more complete example, see Appendix B.

7.3.  Recommended Disposition Values

   The precise nature of the policy action taken in response to a
   disposition and the name of the corresponding "disp" attributes are
   only partially defined here.  It is strongly RECOMMENDED to use the
   following dispositions only in their conventional sense.

   invalid  The resulting string is not a valid label.  This disposition
        may be assigned implicitly; see Section 7.5.  No variant labels
        should be generated from a variant mapping with this type.

   blocked  The resulting string is a valid label but should be blocked
        from registration.  This would typically apply for a derived
        variant that is undesirable due to having no practical use or
        being confusingly similar to some other label.

   allocatable  The resulting string should be reserved for use by the
        same operator of the origin string but not automatically
        allocated for use.

   activated  The resulting string should be activated for use.  (This
        is the same as a Preferred Variant [RFC3743].)

   valid  The resultant string is a valid label.  (This is the typical
        default action if no dispositions are defined.)

7.4.  Precedence

   Actions are applied in the order of their appearance in the file.
   This defines their relative precedence.  The first action triggered
   by a label defines the disposition for that label.  To define the
   order of precedence, list the actions in the desired order.  The
   conventional order of precedence for the actions defined in
   Section 7.3 is "invalid", "blocked", "allocatable", "activated", and
   then "valid".  This default precedence is used for the default
   actions defined in Section 7.6.

7.5.  Implied Actions

   The context rules on code points ("not-when" or "when" rules) carry
   an implied action with a disposition of "invalid" (not eligible) if a
   "when" context is not satisfied or a "not-when" context is matched,
   respectively.  These rules are evaluated at the time the code points
   for a label or its variant labels are checked for validity (see
   Section 8).  In other words, they are evaluated before any of the
   whole-label evaluation rules
   actions are applied, and with higher precedence.  The context rules
   for variant mappings are evaluated when variants are generated and/or
   when variant tables are made symmetric and transitive.  They have an
   implied action with a disposition of "invalid", which means that a
   putative variant mapping does not exist whenever the given context
   matches a "not-when" rule or fails to match a "when" rule specified
   for that mapping.  The result of that disposition is that the variant
   mapping is ignored in generating variant labels and the value is
   therefore not accessible to trigger any explicit actions.

   Note that such non-existing variant mapping is different from a
   blocked variant, which is a variant code point mapping that exists
   but results in a label that may not be allocated.

7.6.  Default Actions

   If a label does not trigger any of the actions defined explicitly in
   the LGR, the following implicitly defined default actions are
   evaluated.  They are shown below in their relative order of
   precedence (see Section 7.4).  Default actions have a lower order of
   precedence than explicit actions (see Section 8.3).

   The default actions for variant labels are defined as follows.  The
   first set is triggered based on the standard variant type values of
   "invalid", "blocked", "allocatable", and "activated":

       <action disp="invalid" any-variant="invalid"/>
       <action disp="blocked" any-variant="blocked"/>
       <action disp="allocatable" any-variant="allocatable"/>
       <action disp="activated" all-variants="activated"/>

   A final default action sets the disposition to "valid" for any label
   matching the repertoire for which no other action has been triggered.
   This "catch-all" action also matches all remaining variant labels
   from variants that do not have a type value.

       <action disp="valid" comment="Catch-all if other rules not met"/>

   Conceptually, the implicitly defined default actions act just like a
   block of "action" elements that is added (virtually) beyond the last
   of the user-supplied actions.  Any label not processed by the
   user-supplied actions would thus be processed by the default actions
   as if they were present in the LGR.  As the last default action is a
   "catch-all", all processing is guaranteed to end with a definite
   disposition for the label.

8.  Processing a Label against an LGR

8.1.  Determining Eligibility for a Label

   In order to test a given label for membership in the LGR, a consumer
   of the LGR must iterate through each code point within a given label
   and test that each instance of a code point is a member of the LGR.
   If any instance of a code point is not a member of the LGR, the label
   shall be deemed invalid.

   An individual instance of a code point is deemed a member of the LGR
   when it is listed using a "char" element, or is part of a range
   defined with a "range" element, and all necessary conditions in any
   "when" or "not-when" attributes are correctly satisfied for that
   instance.

   Alternatively, an instance of a code point is also deemed a member of
   the LGR when it forms part of a sequence that corresponds to a
   sequence listed using a "char" element for which the "cp" attribute
   defines a sequence, and all necessary conditions in any "when" or
   "not-when" attributes are correctly satisfied for that instance of
   the sequence.

   In determining eligibility, at each position the longest possible
   sequence of code points is evaluated first.  If that sequence matches
   a sequence defined in the LGR and satisfies any required context at
   that position, the instances of its constituent code points are
   deemed members of the LGR and evaluation proceeds with the next code
   point following the sequence.  If the sequence does not match a
   defined sequence or does not satisfy the required context,
   successively shorter sequences are evaluated until only a single code
   point remains.  The eligibility of that code point is determined as
   described above for an individual code point instance.

   A label must also not trigger any action that results in a
   disposition of "invalid"; otherwise, it is deemed not eligible.
   (This step may need to be deferred until variant code point
   dispositions have been determined.)

8.1.1.  Determining Eligibility Using Reflexive Variant Mappings

   For LGRs that contain reflexive variant mappings (defined in
   Section 5.3.4), the final evaluation of eligibility for the label
   must be deferred until variants are generated.  In essence, LGRs that
   use this feature treat the original label as the (identity) variant
   of itself.  For such LGRs, the ordinary determination of eligibility
   described here is but a first step that generally excludes only a
   subset of invalid labels.

   To further check the validity of a label with reflexive mappings, it
   is not necessary to generate all variant labels.  Only a single
   variant needs to be created, where any reflexive variants are applied
   for each code point, and the label disposition is evaluated (as
   described in Section 8.3).  A disposition of "invalid" results in the
   label being not eligible.  (In the exceptional case where context
   rules are present on reflexive mappings, multiple reflexive variants
   may be defined, but for each original label, at most one of these can
   be valid at each code position.  However, see Section 8.4.)

8.2.  Determining Variants for a Label

   For a given eligible label, the set of variant labels is deemed to
   consist of each possible permutation of original code points and
   substituted code points or sequences defined in "var" elements,
   whereby all "when" and "not-when" attributes are correctly satisfied
   for each "char" or "var" element in the given permutation and all
   applicable whole label rules are satisfied as follows:

   1.  Create each possible permutation of a label by substituting each
       code point or code point sequence in turn by any defined variant
       mapping (including any reflexive mappings).

   2.  Apply variant mappings with "when" or "not-when" attributes only
       if the conditions are satisfied; otherwise, they are not defined.

   3.  Record each of the "type" values on the variant mappings used in
       creating a given variant label in a disposition set; for any
       unmapped code point, record the "type" value of any reflexive
       variant (see Section 5.3.4).

   4.  Determine the disposition for each variant label per Section 8.3.

   5.  If the disposition is "invalid", remove the label from the set.

   6.  If final evaluation of the disposition for the unpermuted label
       per Section 8.3 results in a disposition of "invalid", remove all
       associated variant labels from the set.

   The number of potential permutations can be very large.  In practice,
   implementations would use suitable optimizations to avoid having to
   actually create all permutations (see Section 8.5).

   In determining the permuted set of variant labels in step (1) above,
   all eligible partitions into sequences must be evaluated.  A label
   "ab" that matches a sequence "ab" defined in the LGR but also matches
   the sequence of individual code points "a" and "b" (both defined in
   the LGR) must be permuted using any defined variant mappings for both
   the sequence "ab" and the code points "a" and "b" individually.

8.3.  Determining a Disposition for a Label or Variant Label

   For a given label (variant or original), its disposition is
   determined by evaluating, in order of their appearance, all actions
   for which the label or variant label satisfies the conditions.

   1.  For any label that contains code points or sequences not defined
       in the repertoire, or does not satisfy the context rules on all
       of its code points and variants, the disposition is "invalid".

   2.  For all other labels, the disposition is given by the value of
       the "disp" attribute for the first action triggered by the label.
       An action is triggered if all of the following are true:

       *  the label matches the whole label rule given in the "match"
          attribute for that action;

       *  the label does not match the whole label rule given in the
          "not-match" attribute for that action;

       *  any of the recorded variant types for a variant label match
          the types given in the "any-variant" attribute for that
          action;

       *  all of the recorded variant types for a variant label match
          the types given in the "all-variants" or "only-variants"
          attribute given for that action;

       *  in case of an "only-variants" attribute, the label contains
          only code points that are the target of applied variant
          mappings;

       or

       *  the action does not contain any "match", "not-match",
          "any-variant", "all-variants", or "only-variants" attributes:
          catch-all.

   3.  For any remaining variant label, assign the variant label the
       disposition using the default actions defined in Section 7.6.
       For this step, variant types outside the predefined recommended
       set (see Section 7.3) are ignored.

   4.  For any remaining label, set the disposition to "valid".

8.4.  Duplicate Variant Labels

   For a poorly designed LGR, it is possible to generate duplicate
   variant labels from the same input label, but with different, and
   potentially conflicting, dispositions.  Implementations MUST treat
   any duplicate variant labels encountered as an error, irrespective of
   their dispositions.

   This situation can arise in two ways.  One is described in
   Section 5.3.5 and involves defining the same variant mapping with two
   context rules that are formally distinct but nevertheless overlap so
   that they are not mutually exclusive for the same label.

   The other case involves variants defined for sequences, where one
   sequence is a prefix of another (see Section 5.3.1).  The following
   shows such an example resulting in conflicting reflexive variants:

       <char cp="0061">
         <var cp="0061" type="allocatable"/>
       </char>
       <char cp="0062"/>
       <char cp="0061 0062">
         <var cp="0061 0062" type="blocked"/>
       </char>

   A label "ab" would generate the variant labels "{a}{b}" and "{ab}"
   where the curly braces show the sequence boundaries as they were
   applied during variant mapping.  The result is a duplicate variant
   label "ab", one based on a variant of type "allocatable" plus an
   original code point "b" that has no variant, and another one based on
   a single variant of type "blocked", thus creating two variant labels
   with conflicting dispositions.

   In the general case, it is difficult to impossible to prove by
   mechanical inspection of the LGR that duplicate variant labels will
   never occur, so implementations have to be prepared to detect this
   error during variant label generation.  The condition is easily
   avoided by careful design of context rules and special attention to
   the relation among code point sequences with variants.

8.5.  Checking Labels for Collision

   The obvious method for checking for collision between labels is to
   generate the fully permuted set of variants for one of them and see
   whether it contains the other label as a member.  As discussed above,
   this can be prohibitive and is not necessary.

   Because of symmetry and transitivity, all variant mappings form
   disjoint sets.  In each of these sets, the source and target of each
   mapping are also variants of the sources and targets of all the other
   mappings.  However, members of two different sets are never variants
   of each other.

   If two labels have code points at the same position that are members
   of two different variant mapping sets, any variant labels of one
   cannot be variant labels of the other: the sets of their variant
   labels are likewise disjoint.  Instead of generating all permutations
   to compare all possible variants, it is enough to find out whether
   code points at the same position belong to the same variant set
   or not.

   For that, it is sufficient to substitute an "index" mapping that
   identifies the set.  This index mapping could be, for example, the
   variant mapping for which the target code point (or sequence) comes
   first in some sorting order.  This index mapping would, in effect,
   identify the set of variant mappings for that position.

   To check for collision then means generating a single variant label
   from the original by substituting the respective "index" value for
   each code point.  This results in an "index label".  Two labels
   collide whenever the index labels for them are the same.

9.  Conversion to and from Other Formats

   Both [RFC3743] and [RFC4290] provide different grammars for IDN
   tables.  The formats in those documents are unable to fully support
   the increased requirements of contemporary IDN variant policies.

   This specification is a superset of functionality provided by the
   older IDN table formats; thus, any table expressed in those formats
   can be expressed in this new format.  Automated conversion can be
   conducted between tables conformant with the grammar specified in
   each document.

   For notes on how to translate a table in the style of RFC 3743, see
   Appendix B.

10.  Media Type

   Well-formed LGRs that comply with this specification SHOULD be
   transmitted with a media type of "application/lgr+xml".  This media
   type will signal to an LGR-aware client that the content is designed
   to be interpreted as an LGR.

11.  IANA Considerations

   IANA has completed the following actions:

11.1.  Media Type Registration

   The media type "application/lgr+xml" has been registered to denote
   transmission of LGRs that are compliant with this specification, in
   accordance with [RFC6838].

   Type name: application

   Subtype name: lgr+xml

   Required parameters: N/A

   Optional parameters: charset (as for application/xml per [RFC7303])

   Security considerations:  See the security considerations for
      application/xml in [RFC7303] and the specific security
      considerations for Label Generation Rulesets (LGRs) in RFC 7940

   Interoperability considerations:  As for application/xml per
      [RFC7303]

   Published specification:  See RFC 7940

   Applications that use this media type:  Software using LGRs for
      international identifiers, such as IDNs, including registry
      applications and client validators.

   Additional information:

      Deprecated alias names for this type: N/A

      Magic number(s): N/A

      File extension(s): .lgr

      Macintosh file type code(s): N/A
   Person & email address to contact for further information:

      Kim Davies <kim.davies@icann.org>

      Asmus Freytag <asmus@unicode.org>

   Intended usage: COMMON

   Restrictions on usage: N/A

   Author:

      Kim Davies <kim.davies@icann.org>

      Asmus Freytag <asmus@unicode.org>

   Change controller: IESG

   Provisional registration? (standards tree only): No

11.2.  URN Registration

   This specification uses a URN to describe the XML namespace, in
   accordance with [RFC3688].

   URI: urn:ietf:params:xml:ns:lgr-1.0

   Registrant Contact: See the Authors of this document.

   XML: None.

11.3.  Disposition Registry

   This document establishes a vocabulary of "Label Generation Ruleset
   Dispositions", which has been reflected as a new IANA registry.  This
   registry is divided into two subregistries:

   o  Standard Dispositions - This registry lists dispositions that have
      been defined in published specifications, i.e., the eligibility
      for such registrations is "Specification Required" [RFC5226].  The
      initial set of registrations are the five dispositions in this
      document described in Section 7.3.

   o  Private Dispositions - This registry lists dispositions that have
      been registered "First Come First Served" [RFC5226] by third
      parties with the IANA.  Such dispositions must take the form
      "entity:disposition" where the entity is a domain name that
      uniquely identifies the private user of the namespace.  For
      example, "example.org:reserved" could be a private extension used
      by the example organization to denote a disposition relating to
      reserved labels.  These extensions are not intended to be
      interoperable, but registration is designed to minimize potential
      conflicts.  It is strongly recommended that any new dispositions
      that require interoperability and have applicability beyond a
      single organization be defined as Standard Dispositions.

   In order to distinguish them from Private Dispositions, Standard
   Dispositions MUST NOT contain the ":" character.  All disposition
   names shall be in lowercase ASCII.

   The IANA registry provides data on the name of the disposition, the
   intended purposes, and the registrant or defining specification for
   the disposition.

12.  Security Considerations

12.1.  LGRs Are Only a Partial Remedy for Problem Space

   Substantially unrestricted use of non-ASCII characters in security-
   relevant identifiers such as domain name labels may cause user
   confusion and invite various types of attacks.  In many languages, in
   particular those using complex or large scripts, an attacker has an
   opportunity to divert or confuse users as a result of different code
   points with identical appearance or similar semantics.

   The use of an LGR provides a partial remedy for these risks by
   supplying a framework for prohibiting inappropriate code points or
   sequences from being registered at all and for permitting "variant"
   code points to be grouped together so that labels containing them may
   be mutually exclusive or registered only to the same owner.

   In addition, by being fully machine processable the format may enable
   automated checks for known weaknesses in label generation rules.
   However, the use of this format, or compliance with this
   specification, by itself does not ensure that the LGRs expressed in
   this format are free of risk.  Additional approaches may be
   considered, depending on the acceptable trade-off between flexibility
   and risk for a given application.  One method of managing risk may
   involve a case-by-case evaluation of a proposed label in context with
   already-registered labels -- for example, when reviewing labels for
   their degree of visual confusability.

12.2.  Computational Expense of Complex Tables

   A naive implementation attempting to generate all variant labels for
   a given label could lead to the possibility of exhausting the
   resources on the machine running the LGR processor, potentially
   causing denial-of-service consequences.  For many operations,
   brute-force generation can be avoided by optimization, and if needed,
   the number of permuted labels can be estimated more cheaply ahead of
   time.

   The implementation of WLE rules, using certain backtracking
   algorithms, can take exponential time for pathological rules or
   labels and exhaust stack resources.  This can be mitigated by proper
   implementation and enforcing the restrictions on permissible label
   length.

13.  References

13.1.  Normative References

   [RFC2045]  Freed, N. and N. Borenstein, "Multipurpose Internet Mail
              Extensions (MIME) Part One: Format of Internet Message
              Bodies", RFC 2045, DOI 10.17487/RFC2045, November 1996,
              <http://www.rfc-editor.org/info/rfc2045>.

   [RFC2119]  Bradner, S., "Key words for use in RFCs to Indicate
              Requirement Levels", BCP 14, RFC 2119,
              DOI 10.17487/RFC2119, March 1997,
              <http://www.rfc-editor.org/info/rfc2119>.

   [RFC3339]  Klyne, G. and C. Newman, "Date and Time on the Internet:
              Timestamps", RFC 3339, DOI 10.17487/RFC3339, July 2002,
              <http://www.rfc-editor.org/info/rfc3339>.

   [RFC5646]  Phillips, A., Ed. and M. Davis, Ed., "Tags for Identifying
              Languages", BCP 47, RFC 5646, DOI 10.17487/RFC5646,
              September 2009, <http://www.rfc-editor.org/info/rfc5646>.

   [UAX42]    The Unicode Consortium, "Unicode Character Database in
              XML", May 2016, <http://unicode.org/reports/tr42/>.

   [Unicode-Stability]
              The Unicode Consortium, "Unicode Encoding Stability
              Policy, Property Value Stability", April 2015,
              <http://www.unicode.org/policies/
              stability_policy.html#Property_Value>.

   [Unicode-Versions]
              The Unicode Consortium, "Unicode Version Numbering", June
              2016, <http://unicode.org/versions/#Version_Numbering>.

   [XML]      Bray, T., Paoli, J., Sperberg-McQueen, M., Maler, E., and
              F. Yergeau, "Extensible Markup Language (XML) 1.0 (Fifth
              Edition)", World Wide Web Consortium, November 2008,
              <http://www.w3.org/TR/REC-xml/>.

13.2.  Informative References

   [ASIA-TABLE]
              DotAsia Organisation, ".ASIA ZH IDN Language Table",
              February 2012,
              <http://www.dot.asia/policies/ASIA-ZH-1.2.pdf>.

   [LGR-PROCEDURE]
              Internet Corporation for Assigned Names and Numbers,
              "Procedure to Develop and Maintain the Label Generation
              Rules for the Root Zone in Respect of IDNA Labels",
              December 2012, <http://www.icann.org/en/resources/idn/
              draft-lgr-procedure-07dec12-en.pdf>.

   [RELAX-NG]
              The Organization for the Advancement of Structured
              Information Standards (OASIS), "RELAX NG Compact Syntax",
              November 2002, <https://www.oasis-open.org/committees/
              relax-ng/compact-20021121.html>.

   [RFC3688]  Mealling, M., "The IETF XML Registry", BCP 81, RFC 3688,
              DOI 10.17487/RFC3688, January 2004,
              <http://www.rfc-editor.org/info/rfc3688>.

   [RFC3743]  Konishi, K., Huang, K., Qian, H., and Y. Ko, "Joint
              Engineering Team (JET) Guidelines for Internationalized
              Domain Names (IDN) Registration and Administration for
              Chinese, Japanese, and Korean", RFC 3743,
              DOI 10.17487/RFC3743, April 2004,
              <http://www.rfc-editor.org/info/rfc3743>.

   [RFC4290]  Klensin, J., "Suggested Practices for Registration of
              Internationalized Domain Names (IDN)", RFC 4290,
              DOI 10.17487/RFC4290, December 2005,
              <http://www.rfc-editor.org/info/rfc4290>.

   [RFC5226]  Narten, T. and H. Alvestrand, "Guidelines for Writing an
              IANA Considerations Section in RFCs", BCP 26, RFC 5226,
              DOI 10.17487/RFC5226, May 2008,
              <http://www.rfc-editor.org/info/rfc5226>.

   [RFC5564]  El-Sherbiny, A., Farah, M., Oueichek, I., and A. Al-Zoman,
              "Linguistic Guidelines for the Use of the Arabic Language
              in Internet Domains", RFC 5564, DOI 10.17487/RFC5564,
              February 2010, <http://www.rfc-editor.org/info/rfc5564>.

   [RFC5891]  Klensin, J., "Internationalized Domain Names in
              Applications (IDNA): Protocol", RFC 5891,
              DOI 10.17487/RFC5891, August 2010,
              <http://www.rfc-editor.org/info/rfc5891>.

   [RFC5892]  Faltstrom, P., Ed., "The Unicode Code Points and
              Internationalized Domain Names for Applications (IDNA)",
              RFC 5892, DOI 10.17487/RFC5892, August 2010,
              <http://www.rfc-editor.org/info/rfc5892>.

   [RFC6838]  Freed, N., Klensin, J., and T. Hansen, "Media Type
              Specifications and Registration Procedures", BCP 13,
              RFC 6838, DOI 10.17487/RFC6838, January 2013,
              <http://www.rfc-editor.org/info/rfc6838>.

   [RFC7303]  Thompson, H. and C. Lilley, "XML Media Types", RFC 7303,
              DOI 10.17487/RFC7303, July 2014,
              <http://www.rfc-editor.org/info/rfc7303>.

   [TDIL-HINDI]
              Technology Development for Indian Languages (TDIL)
              Programme, "Devanagari Script Behaviour for Hindi Ver2.0",
              <http://tdil-dc.in/index.php?option=com_download&task=show
              resourceDetails&toolid=1625&lang=en>.

   [UAX44]    The Unicode Consortium, "Unicode Character Database", June
              2016, <http://unicode.org/reports/tr44/>.

   [WLE-RULES]
              Internet Corporation for Assigned Names and Numbers,
              "Whole Label Evaluation (WLE) Rules", August 2015, 2016,
              <https://community.icann.org/download/
              attachments/43989034/WLE-Rules.pdf>.

Appendix A.  Example Tables

   The following presents a minimal LGR table defining the lowercase LDH
   (letters, digits, hyphen) repertoire and containing no rules or
   metadata elements.  Many simple LGR tables will look quite similar,
   except that they would contain some metadata.

   <?xml version="1.0" encoding="utf-8"?>
   <lgr xmlns="urn:ietf:params:xml:ns:lgr-1.0">
   <data>
       <char cp="002D" comment="HYPHEN (-)" />
       <range first-cp="0030" last-cp="0039"
         comment="DIGIT ZERO - DIGIT NINE" />
       <range first-cp="0061" last-cp="007A"
         comment="LATIN SMALL LETTER A - LATIN SMALL LETTER Z" />
   </data>
   </lgr>

   In practice, any LGR that includes the hyphen might also contain
   rules invalidating any labels beginning with a hyphen, ending with a
   hyphen, and containing consecutive hyphens in the third and fourth
   positions as required by [RFC5891].

   <?xml version="1.0" encoding="utf-8"?>
   <lgr xmlns="urn:ietf:params:xml:ns:lgr-1.0">
   <data>
       <char cp="002D"
             not-when="hyphen-minus-disallowed" />
       <range first-cp="0030" last-cp="0039" />
       <range first-cp="0061" last-cp="007A" />
   </data>
   <rules>
       <rule name="hyphen-minus-disallowed"
             comment="RFC5891 restrictions on U+002D">
         <choice>
           <rule comment="no leading hyphen">
             <look-behind>
               <start />
             </look-behind>
             <anchor />
           </rule>
           <rule comment="no trailing hyphen">
             <anchor />
             <look-ahead>
               <end />
             </look-ahead>
           </rule>
           <rule comment="no consecutive hyphens
                   in third and fourth positions">
             <look-behind>
               <start />
               <any />
               <any />
               <char cp="002D" comment="hyphen-minus" />
             </look-behind>
             <anchor />
           </rule>
         </choice>
       </rule>
   </rules>
   </lgr>

   The following sample LGR shows a more complete collection of the
   elements and attributes defined in this specification in a somewhat
   typical context.

   <?xml version="1.0" encoding="utf-8"?>

   <!-- This example uses a large subset of the features of this
        specification.  It does not include every set operator,
        match operator element, or action trigger attribute, their
        use being largely parallel to the ones demonstrated. -->

   <lgr xmlns="urn:ietf:params:xml:ns:lgr-1.0">
   <!-- meta element with all optional elements -->
   <meta>
       <version comment="initial version">1</version>
       <date>2010-01-01</date>
       <language>sv</language>
       <scope type="domain">example.com</scope>
       <validity-start>2010-01-01</validity-start>
       <validity-end>2013-12-31</validity-end>
       <description type="text/html">
           <![CDATA[
           This language table was developed with the
           <a href="http://swedish.example/">Swedish
           examples institute</a>.
           ]]>
       </description>
       <unicode-version>6.3.0</unicode-version>
       <references>
         <reference id="0" comment="the most recent" >The
               Unicode Standard 9.0</reference>
         <reference id="1" >RFC 5892</reference>
         <reference id="2" >Big-5: Computer Chinese Glyph
            and Character Code Mapping Table, Technical Report
            C-26, 1984</reference>
       </references>
    </meta>

    <!-- the "data" section describing the repertoire -->
    <data>
       <!-- single code point "char" element -->
       <char cp="002D" ref="1" comment="HYPHEN" />

       <!-- "range" elements for contiguous code points, with tags -->
       <range first-cp="0030" last-cp="0039" ref="1" tag="digit" />
       <range first-cp="0061" last-cp="007A" ref ="1" tag="letter" />

       <!-- code point sequence -->
       <char cp="006C 00B7 006C" comment="Catalan middle dot" />

       <!-- alternatively, use a When Rule -->
       <char cp="00B7" when="catalan-middle-dot" />

        <!-- code point with context rule -->
       <char cp="200D" when="joiner" ref="2" />

       <!-- code points with variants -->
       <char cp="4E16" tag="preferred" ref="0">
         <var cp="4E17" type="blocked" ref="2" />
         <var cp="534B" type="allocatable" ref="2" />
       </char>
       <char cp="4E17" ref="0">
         <var cp="4E16" type="allocatable" ref="2" />
         <var cp="534B" type="allocatable" ref="2" />
       </char>
       <char cp="534B" ref="0">
         <var cp="4E16" type="allocatable" ref="2" />
         <var cp="4E17" type="blocked" ref="2" />
       </char>
     </data>

     <!-- Context and whole label rules -->
     <rules>
       <!-- Require the given code point to be between two 006C
            code points -->
       <rule name="catalan-middle-dot" ref="0">
           <look-behind>
               <char cp="006C" />
           </look-behind>
           <anchor />
           <look-ahead>
               <char cp="006C" />
           </look-ahead>
       </rule>

       <!-- example of a context rule based on property -->
       <class name="virama" property="ccc:9" />
       <rule name="joiner"  ref="1" >
           <look-behind>
               <class by-ref="virama" />
           </look-behind>
           <anchor />
       </rule>

       <!-- example of using set operators -->

       <!-- Subtract vowels from letters to get
            consonant, demonstrating the different
            set notations and the difference operator -->
        <difference name="consonants">
            <class comment="all letters">0061-007A</class>
            <class comment="all vowels">
                    0061 0065 0069 006F 0075
            </class>
        </difference>
        <!-- by using the start and end, rule matches whole label -->
        <rule name="three-or-more-consonants">
            <start />
            <!-- reference the class defined by the difference,
                 and require three or more matches -->
            <class by-ref="consonants" count="3+" />
            <end />
       </rule>

       <!-- rule for negative matching -->
       <rule name="non-preferred"
             comment="matches any non-preferred code point">
           <complement comment="non-preferred" >
               <class from-tag="preferred" />
           </complement>
       </rule>

      <!-- actions triggered by matching rules and/or
           variant types -->
       <action disp="invalid"
               match="three-or-more-consonants" />
       <action disp="blocked" any-variant="blocked" />
       <action disp="allocatable" all-variants="allocatable"
               not-match="non-preferred" />
     </rules>
   </lgr>

Appendix B.  How to Translate Tables Based on RFC 3743 into the XML
             Format

   As background, the rules specified in [RFC3743] work as follows:

   1.  The original (requested) label is checked to make sure that all
       the code points are a subset of the repertoire.

   2.  If it passes the check, the original label is allocatable.

   3.  Generate the all-simplified and all-traditional variant labels
       (union of all the labels generated using all the simplified
       variants of the code points) for allocation.

   To illustrate by example, here is one of the more complicated set of
   variants:

       U+4E7E
       U+4E81
       U+5E72
       U+5E79
       U+69A6
       U+6F27

   The following shows the relevant section of the Chinese language
   table published by the .ASIA registry [ASIA-TABLE].  Its entries
   read:

    <codepoint>;<simpl-variant(s)>;<trad-variant(s)>;<other-variant(s)>

   These are the lines corresponding to the set of variants listed
   above:

   U+4E7E;U+4E7E,U+5E72;U+4E7E;U+4E81,U+5E72,U+6F27,U+5E79,U+69A6
   U+4E81;U+5E72;U+4E7E;U+5E72,U+6F27,U+5E79,U+69A6
   U+5E72;U+5E72;U+5E72,U+4E7E,U+5E79;U+4E7E,U+4E81,U+69A6,U+6F27
   U+5E79;U+5E72;U+5E79;U+69A6,U+4E7E,U+4E81,U+6F27
   U+69A6;U+5E72;U+69A6;U+5E79,U+4E7E,U+4E81,U+6F27
   U+6F27;U+4E7E;U+6F27;U+4E81,U+5E72,U+5E79,U+69A6

   The corresponding "data" section XML format would look like this:

     <data>
       <char cp="4E7E">
       <var cp="4E7E" type="both" comment="identity" />
       <var cp="4E81" type="blocked" />
       <var cp="5E72" type="simp" />
       <var cp="5E79" type="blocked" />
       <var cp="69A6" type="blocked" />
       <var cp="6F27" type="blocked" />
       </char>
       <char cp="4E81">
       <var cp="4E7E" type="trad" />
       <var cp="5E72" type="simp" />
       <var cp="5E79" type="blocked" />
       <var cp="69A6" type="blocked" />
       <var cp="6F27" type="blocked" />
       </char>
       <char cp="5E72">
       <var cp="4E7E" type="trad"/>
       <var cp="4E81" type="blocked"/>
       <var cp="5E72" type="both" comment="identity"/>
       <var cp="5E79" type="trad"/>
       <var cp="69A6" type="blocked"/>
       <var cp="6F27" type="blocked"/>
       </char>
       <char cp="5E79">
       <var cp="4E7E" type="blocked"/>
       <var cp="4E81" type="blocked"/>
       <var cp="5E72" type="simp"/>
       <var cp="5E79" type="trad" comment="identity"/>
       <var cp="69A6" type="blocked"/>
       <var cp="6F27" type="blocked"/>
       </char>
       <char cp="69A6">
       <var cp="4E7E" type="blocked"/>
       <var cp="4E81" type="blocked"/>
       <var cp="5E72" type="simp"/>
       <var cp="5E79" type="blocked"/>
       <var cp="69A6" type="trad" comment="identity"/>
       <var cp="6F27" type="blocked"/>
       </char>
       <char cp="6F27">
       <var cp="4E7E" type="simp"/>
       <var cp="4E81" type="blocked"/>
       <var cp="5E72" type="blocked"/>
       <var cp="5E79" type="blocked"/>
       <var cp="69A6" type="blocked"/>
       <var cp="6F27" type="trad" comment="identity"/>
       </char>
     </data>

   Here, the simplified variants have been given a type of "simp" and
   the traditional variants one of "trad", and all other ones are given
   "blocked".

   Because some variant mappings show in more than one column, while the
   XML format allows only a single type value, they have been given the
   type of "both".

   Note that some variant mappings map to themselves (identity); that
   is, the mapping is reflexive (see Section 5.3.4).  In creating the
   permutation of all variant labels, these mappings have no effect,
   other than adding a value to the variant type list for the variant
   label containing them.

   In the example so far, all of the entries with type="both" are also
   mappings where source and target are identical.  That is, they are
   reflexive mappings as defined in Section 5.3.4.

   Given a label "U+4E7E U+4E81", the following labels would be ruled
   allocatable per [RFC3743], based on how that standard is commonly
   implemented in domain registries:

       Original label:     U+4E7E U+4E81
       Simplified label 1: U+4E7E U+5E72
       Simplified label 2: U+5E72 U+5E72
       Traditional label:  U+4E7E U+4E7E

   However, if allocatable labels were generated simply by a straight
   permutation of all variants with type other than type="blocked" and
   without regard to the simplified and traditional variants, we would
   end up with an extra allocatable label of "U+5E72 U+4E7E".  This
   label is composed of both a Simplified Chinese character and a
   Traditional Chinese code point and therefore shouldn't be
   allocatable.

   To more fully resolve the dispositions requires several actions to be
   defined, as described in Section 7.2.2, that will override the
   default actions from Section 7.6.  After blocking all labels that
   contain a variant with type "blocked", these actions will set to
   "allocatable" labels based on the following variant types: "simp",
   "trad", and "both".  Note that these variant types do not directly
   relate to dispositions for the variant label, but that the actions
   will resolve them to the Standard Dispositions on labels, i.e.,
   "blocked" and "allocatable".

   To resolve label dispositions requires five actions to be defined (in
   the "rules" section of the XML document in question); these actions
   apply in order, and the first one triggered defines the disposition
   for the label.  The actions are as follows:

   1.  Block all variant labels containing at least one blocked variant.

   2.  Allocate all labels that consist entirely of variants that are
       "simp" or "both".

   3.  Also allocate all labels that are entirely "trad" or "both".

   4.  Block all surviving labels containing any one of the dispositions
       "simp" or "trad" or "both", because they are now known to be part
       of an undesirable mixed simplified/traditional label.

   5.  Allocate any remaining label; the original label would be such a
       label.

   The rules declarations would be represented as:

     <rules>
       <!--"action" elements - order defines precedence-->
       <action disp="blocked" any-variant="blocked" />
       <action disp="allocatable" only-variants="simp both" />
       <action disp="allocatable" only-variants="trad both" />
       <action disp="blocked" any-variant="simp trad" />
       <action disp="allocatable" comment="catch-all" />
     </rules>

   Up to now, variants with type "both" have occurred only associated
   with reflexive variant mappings.  The "action" elements defined above
   rely on the assumption that this is always the case.  However,
   consider the following set of variants:

       U+62E0;U+636E;U+636E;U+64DA
       U+636E;U+636E;U+64DA;U+62E0
       U+64DA;U+636E;U+64DA;U+62E0

   The corresponding XML would be:

       <char cp="62E0">
       <var cp="636E" type="both" comment="both, but not reflexive" />
       <var cp="64DA" type="blocked" />
       </char>
       <char cp="636E">
       <var cp="636E" type="simp" comment="reflexive, but not both" />
       <var cp="64DA" type="trad" />
       <var cp="62E0" type="blocked" />
       </char>
       <char cp="64DA">
       <var cp="636E" type="simp" />
       <var cp="64DA" type="trad" comment="reflexive" />
       <var cp="62E0" type="blocked" />
       </char>

   To make such variant sets work requires a way to selectively trigger
   an action based on whether a variant type is associated with an
   identity or reflexive mapping, or is associated with an ordinary
   variant mapping.  This can be done by adding a prefix "r-" to the
   "type" attribute on reflexive variant mappings.  For example, the
   "trad" for code point U+64DA in the preceding figure would become
   "r-trad".

   With the dispositions prepared in this way, only a slight
   modification to the actions is needed to yield the correct set of
   allocatable labels:

   <action disp="blocked" any-variant="blocked" />
   <action disp="allocatable" only-variants="simp r-simp both r-both" />
   <action disp="allocatable" only-variants="trad r-trad both r-both" />
   <action disp="blocked" all-variants="simp trad both" />
   <action disp="allocatable" />

   The first three actions get triggered by the same labels as before.

   The fourth action blocks any label that combines an original code
   point with any mix of ordinary variant mappings; however, no labels
   that are a combination of only original code points (code points
   having either no variant mappings or a reflexive mapping) would be
   affected.  These are the original labels, and they are allocated in
   the last action.

   Using this scheme of assigning types to ordinary and reflexive
   variants, all tables in the style of RFC 3743 can be converted to
   XML.  By defining a set of actions as outlined above, the LGR will
   yield the correct set of allocatable variants: all variants
   consisting completely of variant code points preferred for simplified
   or traditional, respectively, will be allocated, as will be the
   original label.  All other variant labels will be blocked.

Appendix C.  Indic Syllable Structure Example

   In LGRs for Indic scripts, it may be desirable to restrict valid
   labels to sequences of valid Indic syllables, or aksharas.  This
   appendix gives a sample set of rules designed to enforce this
   restriction.

   Below is an example of BNF for an akshara, which has been published
   in "Devanagari Script Behaviour for Hindi" [TDIL-HINDI].  The rules
   for other languages and scripts used in India are expected to be
   generally similar.

   For Hindi, the BNF has the form:

       V[m]|{C[N]H}C[N](H|[v][m])

   Where:

   V    (uppercase) is any independent vowel

   m    is any vowel modifier (Devanagari Anusvara, Visarga, and
        Candrabindu)

   C    is any consonant (with inherent vowel)
   N    is Nukta

   H    is a halant (or virama)

   v    (lowercase) is any dependent vowel sign (matra)

   {}   encloses items that may be repeated one or more times

   [ ]  encloses items that may or may not be present

   |    separates items, out of which only one can be present

   By using the Unicode character property "InSC" or
   "Indic_Syllabic_Category", which corresponds rather directly to the
   classification of characters in the BNF above, we can translate the
   BNF into a set of WLE rules matching the definition of an akshara.

     <rules>
       <!--Character class definitions go here-->
       <class name="halant" property="InSC:Virama" />
       <union name="vowel-modifier">
         <class property="InSC:Visarga" />
         <class property="InSC:Bindu" comment="includes anusvara" />
       </union>
       <!--Whole label evaluation and context rules go here-->
       <rule name="consonant-with-optional-nukta">
           <class by-ref="InSC:Consonant" />
           <class by-ref="InSC:Nukta" count="0:1"/>
       </rule>
       <rule name="independent-vowel-with-optional-modifier">
           <class by-ref="InSC:Vowel_Independent" />
           <class by-ref="vowel-modifier" count="0:1" />
       </rule>
       <rule name="optional-dependent-vowel-with-opt-modifier" >
         <class by-ref="InSC:Vowel_Dependent" count="0:1" />
         <class by-ref="vowel-modifier" count="0:1" />
       </rule>
       <rule name="consonant-cluster">
         <rule count="0+">
           <rule by-ref="consonant-with-optional-nukta" />
           <class by-ref="halant" />
         </rule>
         <rule by-ref="consonant-with-optional-nukta" />
         <choice>
           <class by-ref="halant" />
           <rule by-ref="optional-dependent-vowel-with-opt-modifier" />
         </choice>
       </rule>
       <rule name="akshara">
         <choice>
           <rule by-ref="independent-vowel-with-optional-modifier" />
           <rule by-ref="consonant-cluster" />
         </choice>
       </rule>
       <rule name="WLE-akshara-or-other" comment="series of one or
           more aksharas, possibly alternating with other types of
           code points such as digits">
         <start />
         <choice count="1+">
           <class property="InSC:other" />
           <rule by-ref="akshara" />
         </choice>
         <end />
       </rule>
       <!--"action" elements go here - order defines precedence-->
       <action disp="invalid" not-match="WLE-akshara-or-other" />
     </rules>

   With the rules and classes as defined above, the final action assigns
   a disposition of "invalid" to all labels that are not composed of a
   sequence of well-formed aksharas, optionally interspersed with other
   characters, perhaps digits, for example.

   The relevant Unicode character property could be replicated by
   tagging repertoire values directly in the LGR; this would remove the
   dependency on any specific version of the Unicode Standard.

   Generally, dependent vowels may only follow consonant expressions;
   however, for some scripts, like Bengali, the Unicode Standard
   supports sequences of dependent vowels or their application on
   independent vowels.  This makes the definition of akshara less
   restrictive.

C.1.  Reducing Complexity

   As presented in this example, the rules are rather complex --
   although useful in demonstrating the features of the XML format, such
   complexity would be an undesirable feature in an actual LGR.

   It is possible to reduce the complexity of the rules in this example
   by defining alternate rules that simply define the permissible
   pair-wise context of adjacent code points by character class, such as
   a rule that a halant can only follow a (nuktated) consonant.  Such
   pair-wise contexts are easier to understand, implement, and verify,
   and have the additional benefit of allowing tools to better pinpoint
   why a label failed to validate.  They also tend to correspond more
   directly to the kind of well-formedness requirements that are most
   relevant to DNS security, like the requirement to limit the
   application of a combining mark (such as a vowel modifier) to only
   selected base characters (in this case, vowels).  (See the example
   and discussion in [WLE-RULES].)

Appendix D.  RELAX NG Compact Schema

   This schema is provided in RELAX NG Compact format [RELAX-NG].

   <CODE BEGINS>
   #
   # LGR XML Schema 1.0
   #

   default namespace = "urn:ietf:params:xml:ns:lgr-1.0"

   #
   # SIMPLE TYPES
   #

   # RFC 5646 language tag (e.g., "de", "und-Latn")
   language-tag = xsd:token

   # The scope to which the LGR applies.  For the "domain" scope type,
   # it should be a fully qualified domain name.
   scope-value = xsd:token {
       minLength = "1"
   }

   ## a single code point
   code-point = xsd:token {
       pattern = "[0-9A-F]{4,6}"
   }

   ## a space-separated sequence of code points
   code-point-sequence = xsd:token {
       pattern = "[0-9A-F]{4,6}( [0-9A-F]{4,6})+"
   }

   ## single code point, or a sequence of code points, or empty string
   code-point-literal = code-point | code-point-sequence | ""

   ## code point or sequence only
   non-empty-code-point-literal = code-point | code-point-sequence

   ## code point sent represented in short form
   code-point-set-shorthand = xsd:token {
       pattern = "([0-9A-F]{4,6}|[0-9A-F]{4,6}-[0-9A-F]{4,6})"
                 ~ "( ([0-9A-F]{4,6}|[0-9A-F]{4,6}-[0-9A-F]{4,6}))*"
   }

   ## dates are used in information fields in the meta
   ## section ("YYYY-MM-DD")
   date-pattern = xsd:token {
       pattern = "\d{4}-\d\d-\d\d"
   }

   ## variant type
   ## the variant type MUST be non-empty and MUST NOT
   ## start with a "_"; using xsd:NMTOKEN here because
   ## we need space-separated lists of them
   variant-type = xsd:NMTOKEN

   ## variant type list for action triggers
   ## the list MUST NOT be empty, and entries MUST NOT
   ## start with a "_"
   variant-type-list = xsd:NMTOKENS

   ## reference to a rule name (used in "when" and "not-when"
   ## attributes, as well as the "by-ref" attribute of the "rule"
   ## element).
   rule-ref = xsd:IDREF

   ## a space-separated list of tags.  Tags should generally follow
   ## xsd:Name syntax.  However, we are using the xsd:NMTOKENS here
   ## because there is no native XSD datatype for space-separated
   ## xsd:Name
   tags = xsd:NMTOKENS

   ## The value space of a "from-tag" attribute.  Although it is closer
   ## to xsd:IDREF lexically and semantically, tags are not unique in
   ## the document.  As such, we are unable to take advantage of
   ## facilities provided by a validator.  xsd:NMTOKEN is used instead
   ## of the stricter xsd:Names here so as to be consistent with
   ## the above.
   tag-ref = xsd:NMTOKEN

   ## an identifier type (used by "name" attributes).
   identifier = xsd:ID

   ## used in the class "by-ref" attribute to reference another class of
   ## the same "name" attribute value.
   class-ref = xsd:IDREF

   ## "count" attribute pattern ("n", "n+", or "n:m")
   count-pattern = xsd:token {
       pattern = "\d+(\+|:\d+)?"
   }

   ## "ref" attribute pattern
   ## space-separated list of "id" attribute values for
   ## "reference" elements.  These reference ids
   ## must be declared in a "reference" element
   ## before they can be used in a "ref" attribute
   ref-pattern = xsd:token {
       pattern = "[\-_.:0-9A-Z]+( [\-_.:0-9A-Z]+)*"
   }

   #
   # STRUCTURES
   #

   ## Representation of a single code point or a sequence of code
   ## points
   char = element char {
       attribute cp { code-point-literal },
       attribute comment { text }?,
       attribute when { rule-ref }?,
       attribute not-when { rule-ref }?,
       attribute tag { tags }?,
       attribute ref { ref-pattern }?,
         variant*
   }

   ## Representation of a range of code points
   range = element range {
       attribute first-cp { code-point },
       attribute last-cp { code-point },
       attribute comment { text }?,
       attribute when { rule-ref }?,
       attribute not-when { rule-ref }?,
       attribute tag { tags }?,
       attribute ref { ref-pattern }?
   }

   ## Representation of a variant code point or sequence
   variant = element var {
       attribute cp { code-point-literal },
       attribute type { xsd:NMTOKEN }?,
       attribute when { rule-ref }?,
       attribute not-when { rule-ref }?,
       attribute comment { text }?,
       attribute ref { ref-pattern }?
   }

   #
   # Classes
   #

   ## a "class" element that references the name of another "class"
   ## (or set-operator like "union") defined elsewhere.
   ## If used as a matcher (appearing under a "rule" element),
   ## the "count" attribute may be present.
   class-invocation = element class { class-invocation-content }

   class-invocation-content =
       attribute by-ref { class-ref },
       attribute count { count-pattern }?,
       attribute comment { text }?

   ## defines a new class (set of code points) using Unicode property
   ## or code points of the same tag value or code point literals
   class-declaration = element class { class-declaration-content }

   class-declaration-content =
       # "name" attribute MUST be present if this is a "top-level"
       # class declaration, i.e., appearing directly under the "rules"
       # element.  Otherwise, it MUST be absent.
       attribute name { identifier }?,
       # If used as a matcher (appearing in a "rule" element, but not
       # when nested inside a set-operator or class), the "count"
       # attribute may be present.  Otherwise, it MUST be absent.
       attribute count { count-pattern }?,
       attribute comment { text }?,
       attribute ref { ref-pattern }?,
       (
         # define the class by property (e.g., property="sc:Latn"), OR
         attribute property { xsd:NMTOKEN }
         # define the class by tagged code points, OR
         | attribute from-tag { tag-ref }
         # text node to allow for shorthand notation
         # e.g., "0061 0062-0063"
         | code-point-set-shorthand
       )

   class-invocation-or-declaration = element class {
     class-invocation-content | class-declaration-content
   }
   class-or-set-operator-nested =
     class-invocation-or-declaration | set-operator

   class-or-set-operator-declaration =
     # a "class" element or set-operator (effectively defining a class)
     # directly in the "rules" element.
     class-declaration | set-operator

   #
   # set-operators
   #

   complement-operator = element complement {
       attribute name { identifier }?,
       attribute comment { text }?,
       attribute ref { ref-pattern }?,
       # "count" attribute MUST only be used when this set-operator is
       # used as a matcher (i.e., nested in a "rule" element but not
       # inside a set-operator or class)
       attribute count { count-pattern }?,
       class-or-set-operator-nested
   }

   union-operator = element union {
       attribute name { identifier }?,
       attribute comment { text }?,
       attribute ref { ref-pattern }?,
       # "count" attribute MUST only be used when this set-operator is
       # used as a matcher (i.e., nested in a "rule" element but not
       # inside a set-operator or class)
       attribute count { count-pattern }?,
       class-or-set-operator-nested,
       # needs two or more child elements
       class-or-set-operator-nested+
   }

   intersection-operator = element intersection {
       attribute name { identifier }?,
       attribute comment { text }?,
       attribute ref { ref-pattern }?,
       # "count" attribute MUST only be used when this set-operator is
       # used as a matcher (i.e., nested in a "rule" element but not
       # inside a set-operator or class)
       attribute count { count-pattern }?,
       class-or-set-operator-nested,
       class-or-set-operator-nested
   }
   difference-operator = element difference {
       attribute name { identifier }?,
       attribute comment { text }?,
       attribute ref { ref-pattern }?,
       # "count" attribute MUST only be used when this set-operator is
       # used as a matcher (i.e., nested in a "rule" element but not
       # inside a set-operator or class)
       attribute count { count-pattern }?,
       class-or-set-operator-nested,
       class-or-set-operator-nested
   }

   symmetric-difference-operator = element symmetric-difference {
       attribute name { identifier }?,
       attribute comment { text }?,
       attribute ref { ref-pattern }?,
       # "count" attribute MUST only be used when this set-operator is
       # used as a matcher (i.e., nested in a "rule" element but not
       # inside a set-operator or class)
       attribute count { count-pattern }?,
       class-or-set-operator-nested,
       class-or-set-operator-nested
   }

   ## operators that transform class(es) into a new class.
   set-operator = complement-operator
                  | union-operator
                  | intersection-operator
                  | difference-operator
                  | symmetric-difference-operator

   #
   # Match operators (matchers)
   #

   any-matcher = element any {
       attribute count { count-pattern }?,
       attribute comment { text }?
   }

   choice-matcher = element choice {
       ## "count" attribute MUST only be used when the choice-matcher
       ## contains no nested "start", "end", "anchor", "look-behind",
       ## or "look-ahead" operators and no nested rule-matchers
       ## containing any of these elements
       attribute count { count-pattern }?,
       attribute comment { text }?,
       # two or more match operators
       match-operator-choice,
       match-operator-choice+
   }

   char-matcher =
     # for use as a matcher - like "char" but without a "tag" attribute
     element char {
       attribute cp { non-empty-code-point-literal },
       # If used as a matcher (appearing in a "rule" element), the
       # "count" attribute may be present.  Otherwise, it MUST be
       # absent.
       attribute count { count-pattern }?,
       attribute comment { text }?,
       attribute ref { ref-pattern }?
   }

   start-matcher = element start {
       attribute comment { text }?
   }

   end-matcher = element end {
       attribute comment { text }?
   }

   anchor-matcher = element anchor {
       attribute comment { text }?
   }

   look-ahead-matcher = element look-ahead {
       attribute comment { text }?,
       match-operators-non-pos
   }
   look-behind-matcher = element look-behind {
       attribute comment { text }?,
       match-operators-non-pos
   }

   ## non-positional match operator that can be used as a direct child
   ## element of the choice-matcher.
   match-operator-choice = (
     any-matcher | choice-matcher | start-matcher | end-matcher
     | char-matcher | class-or-set-operator-nested | rule-matcher
   )

   ## non-positional match operators do not contain any "anchor",
   ## "look-behind", or "look-ahead" elements.
   match-operators-non-pos = (
     start-matcher?,
     (any-matcher | choice-matcher | char-matcher
      | class-or-set-operator-nested | rule-matcher)*,
     end-matcher?
   )

   ## positional match operators have an "anchor" element, which may be
   ## preceded by a "look-behind" element, or followed by a "look-ahead"
   ## element, or both.
   match-operators-pos =
     look-behind-matcher?, anchor-matcher, look-ahead-matcher?

   match-operators = match-operators-non-pos | match-operators-pos

   #
   # Rules
   #

   # top-level rule must have "name" attribute
   rule-declaration-top = element rule {
       attribute name { identifier },
       attribute comment { text }?,
       attribute ref { ref-pattern }?,
       match-operators
   }

   ## "rule" element used as a matcher (either "by-ref" or contains
   ## other match operators itself)
   rule-matcher =
     element rule {
       ## "count" attribute MUST only be used when the rule-matcher
       ## contains no nested "start", "end", "anchor", "look-behind",
       ## or "look-ahead" operators and no nested rule-matchers
       ## containing any of these elements
       attribute count { count-pattern }?,
       attribute comment { text }?,
       attribute ref { ref-pattern }?,
       (attribute by-ref { rule-ref } | match-operators)
     }

   #
   # Actions
   #

   action-declaration = element action {
       attribute comment { text }?,
       attribute ref { ref-pattern }?,
       # dispositions are often named after variant types or vice versa
       attribute disp { variant-type },
       ( attribute match { rule-ref }
         | attribute not-match { rule-ref } )?,
       ( attribute any-variant { variant-type-list }
         | attribute all-variants { variant-type-list }
         | attribute only-variants { variant-type-list } )?
   }

   # DOCUMENT STRUCTURE

   start = lgr
   lgr = element lgr {
       meta-section?,
       data-section,
       rules-section?
   }

   ## Meta section - information recorded with an LGR that generally
   ## does not affect machine processing (except for "unicode-version").
   ## However, if any "class-declaration" uses the "property" attribute,
   ## a "unicode-version" element MUST be present.
   meta-section = element meta {
       element version {
           attribute comment { text }?,
           text
       }?
       & element date { date-pattern }?
       & element language { language-tag }*
       & element scope {
           # type may by "domain" or an application-defined value
           attribute type { xsd:NCName },
           scope-value
       }*
       & element validity-start { date-pattern }?
       & element validity-end { date-pattern }?
       & element unicode-version {
           xsd:token {
               pattern = "\d+\.\d+\.\d+"
           }
       }?
       & element description {
           # this SHOULD be a valid MIME type
           attribute type { text }?,
           text
       }?
       & element references {
           element reference {
               attribute id {
                   xsd:token {
                       # limit "id" attribute to uppercase letters,
                       # digits, and a few punctuation marks; use of
                       # integers is RECOMMENDED
                       pattern = "[\-_.:0-9A-Z]*"
                       minLength = "1"
                   }
                },
                attribute comment { text }?,
                text
           }*
       }?
   }

   data-section = element data { (char | range)+ }

   ## Note that action declarations are strictly order dependent.
   ## class-or-set-operator-declaration and rule-declaration-top
   ## are weakly order dependent; they must precede first use of the
   ## identifier via "by-ref".
   rules-section = element rules {
     ( class-or-set-operator-declaration
       | rule-declaration-top
       | action-declaration)*
   }

   <CODE ENDS>

Acknowledgements

   This format builds upon the work on documenting IDN tables by many
   different registry operators.  Notably, a comprehensive language
   table for Chinese, Japanese, and Korean was developed by the "Joint
   Engineering Team" [RFC3743]; this table is the basis of many registry
   policies.  Also, a set of guidelines for Arabic script registrations
   [RFC5564] was published by the Arabic-language community.

   Contributions that have shaped this document have been provided by
   Francisco Arias, Julien Bernard, Mark Davis, Martin Duerst, Paul
   Hoffman, Sarmad Hussain, Barry Leiba, Alexander Mayrhofer, Alexey
   Melnikov, Nicholas Ostler, Thomas Roessler, Audric Schiltknecht,
   Steve Sheng, Michel Suignard, Andrew Sullivan, Wil Tan, and John
   Yunker.

Authors' Addresses

   Kim Davies
   Internet Corporation for Assigned Names and Numbers
   12025 Waterfront Drive
   Los Angeles, CA  90094
   United States of America

   Phone: +1 310 301 5800
   Email: kim.davies@icann.org
   URI:   http://www.icann.org/

   Asmus Freytag
   ASMUS, Inc.

   Email: asmus@unicode.org