0032: Formal model admission

FullBaseline checks the formal numeric and literal domains imported by RFC 9880 Appendix A. These grammar obligations are separate from the exact information-level meaning of numbers in Appendix C.1, following ADR 0012: Exact numbers and JSON input policies. Initial, expanded assessments, source inspection and selected-sample admission retain their existing scope and exact-decimal behavior. The formal admission contract records the boundaries.

Imported number admits every JSON number, read as an exact decimal without a precision or range limit. RFC 8610 Appendix E maps CDDL onto JSON and describes JSON numbers as decimal, without limits to their precision or range. It also calls a floating-point precision restriction unlikely to be useful for JSON. RFC 9880 itself uses a decimal that binary64 cannot express: Figure 5 in Section 6.2.1 overrides a length with "minimum": 5e-2. This is Esdieff’s RFC interpretation. Later guidance from the RFC authors would be evidence for revisiting it, not a precondition for applying it.

Imported uint admits integral values from 0 through 2^64-1, following the prelude. Fractional or exponential notation does not change an integral value’s membership. No supplied decimal is rounded to establish admission. For example, 0.1, 5e-2 and 1e400 are numbers, 10.0 is a uint, and 2^64 is a number but not a uint.

Full and Direct numeric parameters, lengths and Full array counts use colon entries in Appendix A. An established value outside their imported domain is a formal grammar error: a colon entry cuts off fallback to map extension entries, as specified in RFC 8610 Section 3.5.4. Every JSON number satisfies a numeric parameter, so only lengths and counts can be excluded. Grouping repetition counts instead use =>. Their non-uint values, including negative, fractional and nonnumeric values, can fall through to the framework’s extension entries. Full records known grammar admission and unavailable base repetition and extension meaning. It does not retain the narrower grouping-count error for those fallback values. No independent normative prohibition has been adopted for them.

The allowed-types union for Full const and default includes the feature alternative any. Known content outside the base alternatives is therefore framework-admitted, with extension meaning and applicable final const meaning left undetermined. Homogeneous number, text or Boolean arrays are base alternatives; nested, mixed or null-element arrays use the extension alternative. Text-key maps have unrestricted values, so their nested numbers and arrays do not inherit the scalar number or array restrictions. The feature annotation describes an extension alternative without itself requiring a named critical feature, following RFC 9165 Section 4.

number membership needs no arithmetic once a number is decoded. One private exact predicate compares a decoded count with virtual decimal digits for uint, which avoids a large-integer dependency, rounding and exponent-sized expansion. The role-aware literal classifier stays beside the existing effective traversal, without a second JSON model or public interface.

Use the original request’s remaining semantic and numeric counters, and charge scans and comparisons before performing them. Decoding limits, unknown critical context, unavailable donor meaning and exhausted work leave the affected obligation undetermined; they never become domain exclusions. Source composition patches do not impose deleted operands on the effective model. Selected and catalogue scopes retain their existing boundaries, and inherited findings retain both their effective locations and original source operands. Optional detail omission preserves aggregate errors and completed work, following ADR 0009: Bounded processing and reporting.

Considered Options

Restricting number to integers from -2^64 through 2^64-1 and values exactly expressible in binary64 follows RFC 8610’s float types literally and was the earlier decision. It made RFC 9880’s own Figure 5 a grammar error, and it rejected ordinary decimal steps in real models, such as multipleOf: 0.1 with unit s and multipleOf: 0.005 for a 5 ms delay step. Sample checking already treated those values exactly. Rounding through binary64 would change the supplied value and accept rounded large integers. Rejecting every extended literal or grouping fallback would replace the normative framework grammar with stricter validation syntax. Leaving known extension admission unfinished would confuse grammar membership with unavailable semantics. Extending these checks to every workflow would change the narrower Initial, expanded and sample contracts. A general CDDL interpreter or arbitrary-precision dependency would add unnecessary surface for these fixed domains and shallow literal alternatives.