TSON 2026 Revision 37 — Change Log

This document records the changes accepted into Revision 37 of the TSON specification series, adjudicated from the implementation’s spec-feedback register (21 entries, renumbered from #1 against Revision 36: the four Revision 36 left open, and seventeen raised since). SPEC-FEEDBACK.md remains a record against Revision 36 and is not modified by this revision; each of its entries receives a disposition here.

Provenance markers used in the disposition table:

• [settled] — the resolution was settled by the spec author inside the feedback register itself, is already implemented in the revised companion artifacts shipped with this revision, or was decided by the spec author at adjudication.
• [open] — a genuine design decision deliberately left open. Collected in §5.

This revision changes the lexer. [TSON-DATA] §1.3 said the lexer was frozen within the 2026 series from Revision 35 and that a later revision touching it would say so. This one does: a bare + becomes the fifteenth special token (§7.2.4, §7.2.5), reserved by the schema grammar for the at-least-one field group (#18). Every token a Revision 36 data document lexes to is unchanged — +5 and +0.5 begin unquoted tokens as before — and a bare + in a data value moves from a lexer error to a parse error.

It also changes the schema grammar (§12.1) in two places — the parameter list, which admits a written type, and the field-group productions — and the vocabulary of both documents: the sentinel _ is the void sentinel and a slot holding it has a void value, where Revision 36 said “absent” (#21).

1. Baseline: the revised artifacts

Fifteen of the register’s proposals arrived already implemented in the companion artifacts, and Revision 37 adopts them as its baseline. Summarising the normative effect:

• #7 — identifier is a text family, and an enum states its type. The kernel declares identifier_type => text_type & atom_specification & { … }, a UAX #31 identifier profile as data (start, continue, start_add, continue_add, medial, exclude, over identifier_base => !enum [XID ID NONE]), and identifier => !identifier_type { continue_add: "-" }, which is [TSON-DATA] §7.7’s profile exactly. An identifier is string-class, inherits the text facets, and every value typed by an identifier family is a name: [TSON-DATA] §8.2’s per-name mechanisms reach it, and the keys of a map keyed by one and the elements of a set of one are look-alike scopes. enum.profile and enum_profile are gone: enum_type => atom & { type: type_name members: enum_set } is the constructor, with enum => enum_type ^ { type?: = identifier } and text_enum => enum_type ^ { type?: = text } as its two pins. (Part 1 §2.6, §7.1, §7.7, §8.2; Part 2 §4.2, §5.4, §5.7, §7.4, §9, §11.4; kernel, fixtures.)
• #8 — value and void each have a constructor. value_type => atom & {} with value => !value_type {}, and void_type => atom & {} with void => !void_type {}; core’s sibling is void => !void_type {}. unit is gone, and with it the one place the series identified a type by its name. (Part 2 §4.1, §4.2, §5.4, §7.3, §9; kernel, core.)
• #9 — a template parameter carries its type. template.parameters is [template_param] over template_param => { name: param_name type: type_ref bound?: type_ref }. The type is derived from the positions the parameter stands in, a declaration may narrow it by writing <T: text, N: int8>, and an application is checked against the parameter list at the call site. (Part 2 §1.3, §5.2, §5.3, §5.10, §8.1, §12.1; kernel, all three fixtures.)
• #10 — order is a facet every container states. array.unordered ~ false becomes ordered ~ true, set_type pins ordered: = false, and map gains ordered ~ false. (Part 2 §5.1, §5.3, §5.6, §7.5, §9; kernel.)
• #11 — a set’s bounds are an array’s. set_type loses its min_items default of 1; the kernel’s member sets say min_items: 1, and meta gains scope_set and decimal_member_set as named entries. (Part 2 §5.3, §7.4, §7.5, §7.8, §9; kernel, meta.)
• #12 — tuple1<T> and voidable_tuple1<T> in core, the one-position tuple the bracket sugar cannot spell. (Part 2 §5.3, §9; core.)
• #13 — uri is RFC 3986’s URI beside a uri_reference. uri_type gains allow_relative ~ true and allow_fragment ~ true, and scheme: text becomes schemes: scheme_set; the kernel’s uri is !uri_type { allow_relative: false }, and core declares uri_reference => !uri_type {} with uri => !uri_reference ^ { allow_relative: false }. (Part 1 §3.3, §5.2, §5.5, §8.1; Part 2 §5.4, §5.5, §5.7, §5.8, §9; kernel, core.)
• #14 — iri_type for RFC 3987. Meta declares iri_type with uri_type’s facets, core declares iri_reference and iri, a URI is US-ASCII, and a directive argument is an IRI-reference. (Part 1 §2.2.1, §3.3, §5.5, §10; Part 2 §5.4, §5.5, §9, §13; meta, core.)
• #15 — core holds only what a schema cannot do without. positive_integer, non_negative_integer, negative_integer, non_positive_integer, non_empty_text, annotation and documentation leave core (48 entries remain), the kernel’s documentation goes with them so that doc is @annotation text in both, and the four sign bounds leave [TSON-DATA] §5.6’s built-in vocabulary. (Part 1 §5.6; Part 2 §1.6, §3.3.3, §7.4, §8.3, §9; kernel, core.)
• #16 — meta’s annotation vocabulary. todo, since and lang leave meta; deprecated becomes @annotation void; comment => @annotation text joins; examples becomes @annotation [text]. (Part 2 §3.3.3, §6, §9, §13; meta.)
• #17 — @doc’s text is CommonMark 0.31.2, with no extensions and raw HTML never executed. (Part 2 §6, §13; kernel, core.)
• #18 — a field group’s option may hold several fields. field_group becomes { members: [[field_name; 1..]; 1..] optional_members?: [field_name; 1..] optional?: boolean ~ false }; | separates options, a ? on a member’s name marks it optional within its option, and ( a: A | b: B )+ is sugar for “at least one of”. (Part 1 §1.3, §4.4, §7.1, §7.2, §7.2.4–§7.2.6, §7.3; Part 2 §5.9, §5.10.1, §5.11, §8.1, §12; kernel, fixtures.)
• #19 — normalization is text_type’s. It is the form a value is put into, over normalization => !enum [NONE NFC NFKC NFKC_CASEFOLD ASCII_CASEFOLD], defaulting to NONE on text_type and NFC on identifier_type, and fixed to NONE on regex_type, uri_type, iri_type and email_type. uri_type.schemes is a set of the kernel’s scheme_name, an ASCII_CASEFOLD identifier. (Part 1 §2.6, §7.2.1; Part 2 §5.2, §5.5, §5.7, §7.4, §7.5, §7.7, §9; kernel, meta.)
• #20 — a leap second, and the precision sentence. Core’s time and datetime docs state that second 60 is refused, and meta’s time_type doc separates reading from writing. (Part 1 §5.4; Part 2 §5.5; meta, core.)
• #21 — two kinds of nothing, one name each. optional is a slot that may be missing (a record field, a field group); voidable is a slot whose value may be void (a record field, an array’s element, a map’s value, a tuple position). element_state is retired and each fact is a boolean. (Part 1 and Part 2 throughout; kernel, core, fixtures.)

The artifacts also carry housekeeping the register did not raise, adopted here on the same footing:

• @bounded has a stated meaning, and seven core types change their value. Meta’s bounded doc now reads “true when the value space has a finite least and greatest value”. Under it date, datetime and duration become @bounded:true, and uuid, mac, cidr4 and cidr6 become @bounded:false — the first two having no order to have a least value in, the networks being partially ordered with no least element. No rule in either part reads the annotation, so no text changes.
• bytes => !bytes_type {} in core, relying on the constructor’s BASE64 default rather than restating it. Part 2 §5.5’s example follows.
• record.discriminators and template.discriminators are [field_name; 1..], so the list is non-empty by type and “whether it is present” replaces “whether it is empty” as what says how a family is dispatched. (Part 2 §5.2, §5.10, §8.1.)
• The hash input is stated in the kernel’s header: sha256 over every byte of a document past its !!id line, which is [TSON-DATA] §2.2.1’s rule.
• The artifact docs are shortened. Rationale the specifications or the guide already carry is removed from entry docs, and the resolved fixtures omit entry docs by a stated convention. Every base kind, role and internal enum in the kernel gains a one-line doc.
• Meta’s set template no longer types a meta field: scoped.scope is scope_set and decimal_type.members is decimal_member_set, so the template serves meta-layer extensions only.

2. Disposition summary

Counts: 17 accepted (4, 5, 7–21 — of which 9’s constructor bound is open), 0 declined, and 4 open questions carried (1, 2, 3, 6), plus the open remainder of Revision 35’s #25 (§5). Revision 36’s open remainder of #15 is declined under #4, and Revision 35’s #4 — typed template parameters — is taken as #9. Open questions are recorded here and are NOT reflected as open text in the specification.

3. Accepted changes by target document

3.1 Part 1 — Text Data Format

1. Header, §1.1, §2.1, §2.2 — “void sentinel” for “absent sentinel” (#21).
2. §1.3 — the lexer-freeze sentence states this revision’s change: a bare + is a special token (#18).
3. §2.2.1 — a directive argument is read as an IRI-reference, so an identity may carry characters beyond US-ASCII as themselves, compared as written (#14).
4. §2.3, §7.3, §7.4 — the absent and absent-token productions are renamed void and void-token (#21).
5. §2.4 — the trailing-comma paragraph speaks of a void value occupying a slot (#21).
6. §2.6 — a key typed by an identifier family is a name, by its type’s statement; a text key type’s normalization can make more keys equal (#7, #19).
7. §2.9 — retitled The Void Sentinel and restated in the new vocabulary: present with a void value, distinct from a missing key (#21).
8. §3.3 — a directive argument is an IRI-reference (RFC 3987); the table’s Argument column follows (#13, #14).
9. §4.4 — bare - and + are special tokens and a bare . a lexer error (#18).
10. §5.2 — the !uri and !iri scheme requirement joins the range constraints as a value rule (#13).
11. §5.4 — second 60 is refused, as a resolver error (#20).
12. §5.5 — the table gains !uri_reference, !iri and !iri_reference; !uri’s row says the scheme is required and the text US-ASCII; a paragraph states the four atoms’ categories (#13, #14).
13. §5.6 — the sign-bound row leaves the table, and the paragraph says where a sign bound is written now (#15). !integer joins the table: the arbitrary-precision integer was never in the vocabulary, though core declares it and every width refines it, and with the sign bounds gone no built-in annotation named one. An oversight, as !boolean was in Revision 36. [settled]
14. §6 — JSON null maps to void (#21).
15. §7.1 — the identifier profile is the kernel’s identifier instance; an enum’s members are names where its type is an identifier family; bare + has a role (#7, #18).
16. §7.2 — rule 5 is the void sentinel; rule 7 lists fifteen special tokens (#18, #21).
17. §7.2.1 — under a schema a text type’s normalization generalises the identifier-position rule (#19).
18. §7.2.4, §7.2.5, §7.2.6 — + takes -’s boundary rule; fifteen special tokens, thirteen reserved by the schema grammar; a bare + in a data value is a parse error (#18).
19. §7.7 — the grammar is the identifier instance’s profile; a map key typed by an identifier family is judged as a value of it, under a schema (#7).
20. §8.1 — the resolver and validation lists follow §5.5 and the renamed sentinel (#13, #21).
21. §8.2 — the per-name mechanisms reach every identifier position and every value typed by an identifier family, under the family’s own profile; a profile’s additions are exempt from Identifier_Status; a per-segment unit divides at the profile’s own separators; two data scopes are named, by reference to [TSON-SCHEMA] §11.4 (#7).
22. §9.4 — a value is a name where a schema types it by an identifier family (#7).
23. §10 — RFC 3987 is a normative reference (#14).

3.2 Part 2 — Type System and Schema

1. §1.3 — a resolved-output consumer checks an application against the template entry’s typed parameter list and never reads the held body (#9).
2. §1.6 — the example declares title_text => !text ^ { min_length: 1 } (#15).
3. §3.3.1 — template_param joins the supporting records that are not constructors (#9).
4. §3.3.3 — meta’s annotation list is deprecated, title, comment, examples, read_only, write_only, ordering, bounded, exact, numeric, disjoint; the kernel carries doc and annotation, and core declares doc alone (#15, #16).
5. §3.3.4 — a family is open across schemas and judged over the closure that holds it (#4). The paragraph also gives §5.2’s existing references to this section a target.
6. §4.1, §4.2 — value_type and void_type replace unit; the name-dispatch rule goes; core’s void is !void_type {}; identifier is a text family (#7, #8).
7. §5.1 — the constructor-refinement example is set_type => array ^ { ordered?: = false } (#10).
8. §5.2 — void vocabulary throughout (#21); a selector’s text pins compare in the field type’s normalization form (#19); a family member is declared (#4); the discriminator list is present or not (housekeeping).
9. §5.3 — voidable for state: OPTIONAL at an element, a map value and a tuple position; ordered on array and map; set_type without a default bound, and a bounded set at a field as a named entry; tuple1<T> and voidable_tuple1<T>; an argument read as its parameter’s type (#9, #10, #11, #12, #21).
10. §5.4 — an enum’s class is read from its type; identifier families and the URI and IRI atoms are string-class; value is the atom with no class (#7, #8, #13, #14).
11. §5.5 — bytes => !bytes_type {}; new clauses for text and normalization, for the URI and IRI families, and for the leap second; the precision sentence reworded (#13, #14, #19, #20).
12. §5.6 — the end-state example reads set_type’s ordered: = false (#10).
13. §5.7 — the permission and member-set kinds name uri_type’s facets; fixed at construction is a facet kind (normalization, the identifier profile facets, an enum’s type); profile leaves settable-once (#7, #13, #19).
14. §5.8 — the uri_type example is the kernel’s current declaration; a restated group member may drop its ? (#13, #18, #19).
15. §5.9 — rule 7 is the new removal rule for groups (#18).
16. §5.10 — a parameter’s type is derived and recorded; <T: X>; agreement of several uses; the earlier-parameter rule; bounds and their inheritance; the same-name same-body sentence; the call-site check; “Parameters carry no bounds” is replaced by what remains deferred (#9). bounded_set’s head is corrected to !set_type (housekeeping).
17. §5.10.1 — guards read voidable and chosen options (#18, #21).
18. §5.11 — rewritten: options, the in-option ?, +, the three validity rules, the declaration rules, resolution to members/optional_members/optional, refinement, composition and removal (#18, #21).
19. §6 — the advisory list follows meta; @doc is CommonMark 0.31.2; @examples are text; the form example is @title:"Order" (#16, #17).
20. §7.1, §7.8 — the cell rule and the typed-position restriction no longer overlap (#5); scoped.scope is scope_set; a key’s void value (#11, #21).
21. §7.3 — void is a void_type instance and admits the void sentinel (#8, #21).
22. §7.4 — enum member semantics rewritten around enum_type.type; identifier families replace “the identifier primitive”; text members are values in the type’s form; identifier-profile coherence joins the coherence list (#7, #19).
23. §7.5 — a set may be empty; ordered and the output rule for arrays, sets and maps; uri_type.schemes among the set-typed fields (#10, #11, #13).
24. §7.6 — retitled The Void Sentinel Under a Schema; the table reads voidable (#21).
25. §7.7 — typed key equality reads the key type’s normalization form (#19).
26. §8.1 — template_param; field_group’s shape; voidable on tuple_element, array and map; ingest verifies a recorded parameter type and re-runs the family checks over the closure; the body-patterns table (#4, #9, #18, #21).
27. §8.2 — a family member is declared; the merge key is declined (#4).
28. §8.3 — the chain example is user_id → id → uuid (#15).
29. §9 — both table rows; core’s contents and the principle behind them; the type_name exception for enum_type.type in the extension guidance (#7–#19).
30. §11.4 — the enum scope is conditional on an identifier family; two data scopes; values, defaults, pins, data bodies and annotation values typed by an identifier family; profile additions and the per-segment unit (#7).
31. §12.1, §12.2, §12.3 — type-param with an optional written type; group-def, group-option and group-member; the + suffix; notes and the adjacency table (#9, #18).
32. §13 — RFC 3987, UAX #15, UAX #31 and CommonMark 0.31.2 added; RFC 5646 removed with lang; pins to the artifacts as received.
33. §5.7, §1.6, §5.10 — a parametric modifier takes the name mark its literal spelling takes. w?: int32 ~ N is a default, w?: int32 = N an injected pin, and w: int32 = N a marker; w: int32 ~ N is refused as w: T ~ v is. Revision 36 had the modifier written on an unmarked name with the mark supplied at closing; the ? is the field’s presence statement and says nothing about the value, so it is written where it is meant. §1.6’s flagged is re-spelled priority?: priority ~ N. Not in the register; settled by the spec author on review, and what the register’s own examples and the kernel’s template_param doc already write. [settled]

3.3 Companion artifacts

The revised artifacts shipped with this revision carry the baseline of §1 and every kernel-, meta- and core-level change in this log. Their entries are not edited by this adjudication; §6 lists two edits to prose: one unpinned fixture’s conventions note, and four doc lines of meta-kernel.tn, edited after its pin was stamped and re-stamped since.

3.4 Part 3 — JSON Encoding

[TSON-JSON] is edited directly as findings arise, and arrived with the Part 3 halves of five entries already in place: #5 (§3.3, §8.5, §9.4), #7 (§1.6 item 7, §5.2, §9.4), #13 and #14 (§5, §5.6) and #18 (§6.1.4). The second pass completed it for the rest, and §7.1 lists the edits. Part 3 defines no type-system rule, so each is a spelling of a rule §3.1 or §3.2 owns; the one place it adds processor requirements of its own is the ordered map (§7.1).

4. Notable normative changes (reader’s digest)

Changes a Revision 36 implementer must act on:

1. A bare + is a special token. It takes -’s boundary rule: followed by a continuation character it begins an unquoted token, otherwise it is emitted alone. In a data value it is a parse error, where it was a lexer error.
2. A field group has options. ( host: H port: P | socket: S ) chooses one option; a ? on a member’s name makes it optional within its option; ( email: E | phone: P )+ is at least one. field_group is { members: [[…]] optional_members? optional? }, so every schema with a group resolves differently, with the same text and value set.
3. optional and voidable everywhere. state: OPTIONAL on an array, map or tuple position is voidable: true; on a field group it is optional: true. element_state is gone. The sentinel is the void sentinel.
4. identifier is a text family. It is string-class, so ( identifier | int32 ) is disjoint; a value typed by an identifier family meets name hygiene, and an identifier-keyed map’s keys and a set of identifiers are look-alike scopes. A schema that wants one declares it: identifier => !identifier_type { continue_add: "-" }.
5. An enum states its type. !enum [OPEN DONE] is unchanged in source and records source: enum; profile: TEXT is written !text_enum [...]; !enum_type { type: kebab members: [...] } enumerates labels of a schema’s own vocabulary.
6. unit is gone. value is !value_type {} and void is !void_type {}, and a processor recognises both by constructor.
7. Template parameters are typed. Output carries parameters: [{ name: T type: type_ref } { name: N type: non_negative_integer }]; <T: text> bounds a type parameter and <N: int8> narrows a value parameter; an argument is checked at the application, against the parameter list.
8. ordered replaces unordered, and map has it too, defaulting to false.
9. A set may be empty. set<text> admits []; a non-empty set says min_items: 1.
10. uri requires a scheme and is US-ASCII. uri_reference, iri and iri_reference are new atoms, in core and in the built-in vocabulary. scheme is schemes, a set.
11. normalization is a text facet. A text value is its token’s text put into the type’s form, and every facet and comparison judges that value. A round trip writes the value.
12. Core is smaller. The four sign-bound integers, non_empty_text, annotation and documentation are gone from core, and the sign bounds from [TSON-DATA] §5.6.
13. Meta’s annotations change. @deprecated is bare; @comment is new; @examples takes text; @since, @todo and @lang are gone. @doc is CommonMark 0.31.2.
14. A family member is declared. A use-site template application that composes onto a record is a resolver error, and a family is judged over the whole closure.
15. A !!schema at a declared position is a validation error, by the cell rule; at a position that is not scoped it remains a resolver error.
16. A leap second is refused, as a resolver error, under time and datetime.
17. tuple1<T> and voidable_tuple1<T> are in core.
18. @bounded changes on seven core types (§1).
19. A parametric default is written a?: T ~ N, with the ?, as a literal one is; the unmarked Revision 36 spelling a: T ~ N is refused.
20. !integer is in the built-in vocabulary.

5. Open questions carried by this change log

Adjudicated 2026-10-07. The following remain deliberately open. They live in this change log only — the specification text carries no open questions.

Decisions taken 2026-10-07: #4 both halves, and the merge key declined; #5 the cell rule decides at every scoped position; #9 the narrow same-type sentence, ingest verifying a recorded type, and the constructor bound left open; #15 the inline prohibition stands; #20 a leap second is a resolver error; #21 the grammar productions are renamed with the prose. Two further decisions by the spec author on review of the first pass: a parametric field modifier takes the name mark its literal spelling takes (§3.2 item 33), and !integer joins the built-in vocabulary (§3.1 item 13).

6. Artifact work

1. Received at Revision 37 — all six artifacts carry “2026 Revision 37 draft” and /2026/37/ identities, and each pin was checked against sha256 over the bytes past the !!id line of the copy received: kernel cfbc5c47…, meta 001ce80d…, core 0315cb4d…. Item 3’s edit moved them, and Part 2 §13.2 carries the re-stamped values. Hash values remain non-normative; only the pin’s shape is.

2. One fixture edited. meta-kernel-resolved.tn is unpinned, and its conventions note is corrected in three places: set-typed fields are compared as written, in source declaration order (Revision 36 §7.5), where the note still had comparison tools canonicalising them; a parametric modifier takes the name mark as written (§3.2 item 33), where the note had = P staying a required field until closing; and “REQUIRED field” is “required field”. No entry changes.

3. meta-kernel.tn edited, and re-stamped. Four doc edits, by the spec author’s direction; none changes an entry or resolved output:

   • The type_argument doc said “The group is REQUIRED”. Under #21 a group is optional or it is not, and it now reads “The group is not optional”.
   • The record_field doc said a parametric = P “is a required FREE field … and becomes optional and FIXED when substitution makes the value concrete”. It now states §3.2 item 33: a parametric modifier takes the name mark its literal spelling takes, the parameter stands in value until substitution, and closing changes neither optional nor role.
   • The type_ref and array_type docs said “REQUIRED field”, a Revision 35 state name; both read “required field”, as the record_field doc beside them does.

   The hash chain is re-stamped over the shipped bytes. Because meta pins the kernel and core pins meta, the edit moved all three, in that order, with the pinned references in meta.tn and core.tn and the three rows of Part 2 §13.2. §8.4’s edits moved them again: kernel f2c2b278…, meta 568b589b…, core 4845f0c1….

4. Not verified, carried as a note: #9’s ingest check of a recorded parameter type is stated in §8.1 and is not running in the implementation, which re-resolves from source; and #10’s ordered facet is consulted by bind mode, which binds an ordered map to an order-keeping host map, while tree mode keeps every map’s order, which honours both values. The first is implementation work rather than a question for the specification.

7. Second pass: Part 3 and the Developer Guide

Both ship with Revision 37. Neither carries a decision that §2 does not record; where a sentence had to be written that no register entry dictates, it is marked [editorial] and is the spec author’s to confirm.

7.1 Part 3

tson-part3-json.md is the copy received, completed for the entries it had not yet consumed. Header and status read Revision 37; references move to /2026/37/ and gain RFC 3987.

1. §1.6 gains items 8 to 14 — the identifier family, normalization, the URI and IRI atoms, field-group options, optional/voidable and the void vocabulary, the order facet and the empty set, and the value/void constructors with #5’s cell rule — each with where it lives and which sections here consume it.
2. Void vocabulary (#21) throughout. §6.1.2 is retitled “Missing and void”, §7 “Void and Null”, and §7.2 is rewritten over the marks; §1.1 and §1.5 no longer speak of field states. JSON null is the void sentinel’s spelling, a member not written is missing, and “absent” is used for neither.
3. §2’s example declared a pack size as an atom refinement written inline at a field, which #15’s standing prohibition refuses. It now declares pack => !integer ^ { min: 1 } and writes pack_size: pack. [editorial]
4. §3.1, §6.4 — key identity. Duplicate member names at a map position are judged after the key type’s normalization has been applied (#19), so two spellings one form apart are one key in both encodings.
5. §5, §5.2, §5.6 — atoms. Identifier families are string-class and take the family’s profile and normalization (#7); an enum member is matched in the member type’s form; text values are put into their declared form before facets are judged (#19); uri and iri refuse a relative reference and uri anything outside US-ASCII (#13, #14); a leap second is refused as in the notation (#20).
6. §5.7 reads value and void as constructors (#8) and a scoped position under #5’s cell rule.
7. §6.1.4 states group counting in #18 and #21’s words: a group that is not optional admits exactly one option, an optional group at most one, and the + form any non-empty subset of its members.
8. §6.2, §6.3 — a set may be empty unless its type says otherwise (#11); an array’s ordered facet is read from the type (#10); tuple1<T> is a one-element JSON array (#12).
9. §6.4 — an ordered map (#10). [editorial] The register gives a map an ordered facet and says nothing of JSON, whose object members are unordered by RFC 8259. The paragraph added here states the consequence: where the type states ordered: true an encoder MUST write the entries in the value’s order and a decoder MUST deliver them in the order read, in object form and pairs form alike, with a note that an ordered map in object form keeps its value only through intermediaries that keep member order. This is the only place the second pass adds requirements that no register entry spells, and it is listed in §5 for confirmation.
10. §8.3, §8.5, §9.1, §9.3, §9.4 — identifier families join the class-stable list; the scoped-position errors follow #5 (a cell-rule failure is a validation error, a non-scoped position a resolver error); the round-trip latitude names the order of an unordered map’s entries and the spelling of a text value its type normalises; the error list gains a relative reference, scheme or fragment a URI or IRI type does not admit.
11. Cross-references. Seven bare section numbers that meant Part 1 or Part 2 and resolved, read literally, to a section of Part 3 are qualified — among them the MUST in §9.4 that a look-alike member name is refused under [TSON-DATA] §8.2. Present in the copy received; no change of meaning.

7.2 Developer Guide

tson-guide.md is realigned on the same two principles as before: it describes the design as it stands, and history lives in the change logs.

1. §2.6 — field-group options, the + form, the one-spelling rules, and the census that motivated them (#18).
2. §2.7 — why a parameter carries a type, and what the call-site check buys (#9).
3. §2.9 — “two kinds of nothing” (missing against void, #21), and the rule that a parametric modifier takes the name mark its literal spelling takes (§3.2 item 33).
4. §2.10 — a family is judged over its closure and a member is declared (#4).
5. §3.1, §3.5 — bare + as a special token; the enum type as data (enum_type { type members }, enum and text_enum), and identifier as a type with a profile rather than a lexer category (#7).
6. §4.4, new §4.5 — the meta annotation vocabulary and @doc as CommonMark (#16, #17); “Why core is small”, the restraint rule and the names that left (#15).
7. §6.1 is retitled “value, void, and identifier” (#8); §6.2, §6.4 take the named member sets, scope_set (#11) and #5’s errors.
8. New §6.6, “Text: a value is put into its form” — normalization, URIs against IRIs and references against absolute forms, the leap second (#13, #14, #19, #20).
9. New §6.7, “Containers: order, emptiness, and the one-position tuple” (#10, #11, #12).
10. §7’s worked example follows Part 2 §1.6 as revised (title_text, typed use of priority ~ N), and its resolved output shows template_param entries with their types and the held body. The held field is written optional: true role: DEFAULT value: N, which follows from §3.2 item 33 and is not copied from a fixture — no shipped fixture holds a parametric default. [editorial] The section now says it extends §1.6’s schema with two fields, which it has done since Revision 35.
11. §8.1, §9.3 in the void vocabulary; references gain RFC 3987 and CommonMark; identities are /2026/37/. Three small corrections ride along: §2.6’s example names datetime, which core declares, where it named timestamp; §4.3 no longer lists identifier among the kernel types meta’s declarations use; §6.2 no longer says integer_type’s step is typed integer (it is non_negative_integer).

The Revision 33 guide candidates remain outstanding.

7.3 An example re-spelled

{text => value} stood at a type position in Part 3 §6.1.1 and Part 2 §7.2, in schemas that import core. value is a kernel entry and core declares no sibling of that name, so under Part 2 §3.3.2 the reference did not resolve as written. Both fragments predate this revision. Both now read {text => text}, which is what Part 3’s JSON example beside it already carried; core gains nothing, in keeping with #15.

8. Third pass: verification against the implementation

Each accepted entry was checked against the text of all three parts and the artifacts, and the implementation run where an entry states what it does. Where the text and the running code disagreed on wording alone, the text now says what runs; what remains open is SPEC-FEEDBACK.md’s, renumbered against this revision.

8.1 Part 1

1. §2.2.1 — the !!id argument is an IRI-reference (§3.3), where the opening said a URI (#14).

2. §7.2.1 — two string values differing only in composition stay distinct as written and compare in NFC wherever they are compared ([TSON-SCHEMA] §5.5); the example’s “decomposed” spelling is written "cafe\u0301", where both spellings were precomposed.

3. §2.2.1 — an identity without a host has an absolute path, so that path-only and hosted identities are disjoint: a relative tson.io/2026/37/m/core.tn would be the identity https://tson.io/2026/37/m/core.tn reduces to. [settled] by the spec author.

4. §8.2 — a policy has a vocabulary, and no document selects its own. The policy’s home was “an artifact of a kind this series does not yet define”. It is now a data document of the policy type the companion artifact policy.tn declares — the identifier and token policies and the resource limits — which is also the shape a processor states its policy in, and §8.2 adds the constraint the paragraph’s own reasons imply: no document may name, import or otherwise select the policy it is judged under. Fetch allow-lists and host mappings stay deployment-internal. [settled] by the spec author.

8.2 Part 2

1. §2.2, §2.2.3, §7.1 — a directive value is an IRI-reference ([TSON-DATA] §3.3), where three sentences said a URL string (#14).

2. §5.2 — the family-member error is located at the declaration whose closing minted the member where the application sits in a template’s held body (kt for kt => kennel_of<text>) (#4).

3. §5.4 — “a unit placeholder” is “a slot that only ever holds the void value”, unit being gone (#8).

4. §5.7, §7.4 — a facet fixed at construction may be pinned by a refinement of the constructor, which is how enum and text_enum state type, and an instance refinement restates it verbatim; §7.4’s “never restates the type” is “never changes” (#7, #19).

5. §5.10 — the value-parameter example is <N: int8> { w?: integer ~ N }; the one it replaced, <N: int8> at max_items, was refused by its own rule, int8 not being IS-A non_negative_integer (#9).

6. §5.10 — two same-named entries with the same resolved body are one type wherever §5.10 compares parameter types, the agreement of several uses included, where the text named two checks of the four (#9).

7. §7.8 — a container is a position whose own type is not scoped, its elements and values being positions of their own, where “or a container of one” left a !!schema on [declared] itself in neither branch (#5).

8. §5.5 — no comparison goes below NFC (#19). Two text values are one when, each in its type’s form, they are NFC-equal, under every normalization member: NONE and NFC share one equality and differ in the value, which is what the other facets judge and a round trip writes. The text had said a composed and a decomposed É stay two values under NONE and ASCII_CASEFOLD, which [TSON-DATA] §2.6 contradicted for map keys — one pair, one key in a map and two members of a set. NFC can carry identity where name hygiene cannot: the normalization stability policy keeps the verdict fixed across Unicode versions, and no identifier or token policy relaxes it. [settled] by the spec author.

9. §5.5, §5.8, §7.8, §9 — iri_type is the kernel’s and identities are schema_identity. scoped.schemas was typed by the kernel’s uri, which requires a scheme and US-ASCII and admits a fragment, so a schema whose identity is an IRI or path-only could not be named at a scoped position, and a fragment could. Meta’s schema_identity => !iri_type { allow_fragment: false } now types the keys; since meta cannot construct an instance of a constructor it declares (§3.3.1), the IRI family moves to the kernel, where its iri types atom_specification.spec, and uri_type moves to meta. extern_of and extern_type record S: schema_identity. [settled] by the spec author.

10. §5.2, §5.10, §8.1, §8.2 — a family’s members are the applications a declaration names. §5.2’s “a family member is declared” exempted applications of a family’s base, but a record-bodied template’s applications were its members, so k: box<text> at a use site minted a member of box’s family under a content-derived name, and a value at a position typed box was then told to name it. A use-site application is now a type read where it is written and no member: an instantiation of a family base carries the template in its supertypes only where a declaration names it. Nothing is refused that loaded before. [settled] by the spec author.

11. §5.10 — where a parameter’s type is read is decided by the template’s form. A recorded type was a bare name read in the schema’s namespace first, so a schema declaring its own non_negative_integer with another body changed what vec => <N> !array { … min_items: N } admitted: vec<0> was refused under a min: 1 shadow, and vec<"abc"> passed the call-site check under a text one. A template applying a meta constructor other than record now reads its value parameters’ types, written ones included, in the structure namespace; a record template reads them in the schema’s; a bound is the schema’s in every form. The entry’s source says which form it is, so resolved output needs no new field. [settled] by the spec author.

12. §5.10 — the constructor bound is declined, and the rule is stated per slot. A value conforms to where it is defined: a value parameter’s type is read where the slot it binds into is defined, the meta’s for a constructor’s own slot and the schema’s for a field the schema types; a type parameter names a local type, so its bound is local. A constructor bound (<T: text_type> set<T>) would judge a schema’s type by meta vocabulary, the crossing item 11 removed. “Any text-valued atom” is not a type — a construction founds a family of its own, and an author joins text’s by refining it — and “any scalar” is a style a schema states with an element or key type. [settled] by the spec author.

13. The bounded type slot at a field is withdrawn, not deferred. The field half of Revision 36’s #6 — type: <T: text> members: set<T>, a field binding a type that a later field’s type depends on — would introduce a dependent record: a field order a streaming reader must rely on, a resolved form for the binding, and a kernel change to record_field and enum_type. Now is the wrong time to introduce such a feature: the template half has only just landed, no bundled schema writes a typed parameter to prove its shape, and what the field half would make structural — an enum’s members against its type (§7.4), a field’s value against its type (§5.2) — is already stated and enforced as prose rules. It is not carried as an open question; a later revision that wants it raises it afresh. [settled] by the spec author.

8.3 Part 3

1. §1.6 item 14 cites §8.5 and §9.4 for a $schema at a declared position, where it cited §3.3, which does not state it.
2. §6.1.4 — the + form lowers to one option holding every member, each marked optional, where the sentence said one option per member, which is how it is written.
3. §3.1 — a text map key is compared in NFC after its type’s form (§5.5’s floor).

8.4 Companion artifacts

1. @ordered is @ordering. Meta’s ordering classification over [NONE PARTIAL TOTAL] shared its name with the container facet #10 introduced, which says something unrelated: a path type is @ordering:NONE (no order relation on its values) and ordered: true (element order is part of a value). The annotation is renamed, its doc says which is which, and core’s forty markers, both resolved fixtures and Part 2 §3.3.3, §6 and §9 follow. No rule in either part reads it.
2. Doc lines. meta-kernel.tn: template_param’s type is “never required of the author”, where it said “never written” beside the paragraph on writing one; template’s parameters lists template_param records, where it said names; void binds no host value, where it said “the host value is absent”. meta.tn: the float bounds are “optional”, where a Revision 35 state name stood. meta-kernel-resolved.tn: “plain required type_ref-typed fields”, the one “REQUIRED” §6 item 2 missed.
3. iri_type and uri_type change places (§8.2 item 9): the kernel declares iri_type and iri, meta uri_type and schema_identity; core is unchanged but for extern_of’s doc.
4. Re-stamped, as §6 item 3 records, and Part 2 §13.2 carries the pins.
5. policy.tn joins the companion artifacts (§8.1 item 4): the processor policy’s vocabulary at /2026/37/m/policy.tn, pinned to meta and core, with restriction_level, script_policy, identifier_policy, limits and policy. A script is named by its Unicode Script property value alias; limits states nesting depth, §9.1’s other limits joining as processors expose them; a policy’s unicode_data_version is optional, a report stating it and a deployment’s document leaving it out. Part 2 §1.5 and §13.2 list it.