Against everything else

Each row is a property a reader can verify in an afternoon. No row is a matter of taste, and no opponent is chosen for being easy — CBOR and Ion are the closest existing things and are the ones worth losing to.

The matrix

property	json	yaml	xml	toml	cbor	ion	protobuf	rdf	unimsg
human-writable	yes	yes	poor	yes	no	yes	no	poor	yes
exact integers beyond 2^53	no	no	n-a	yes	yes	yes	yes	yes	yes
exact decimals	no	no	n-a	no	tag	yes	no	yes	yes
native binary	no	no	no	no	yes	yes	yes	no	yes
timestamps with precision kept	no	partial	yes	yes	tag	yes	no	yes	yes
comments	no	yes	yes	yes	no	yes	n-a	yes	yes
self-describing	yes	yes	yes	yes	yes	yes	no	yes	yes
canonical form in the base spec	no	no	no	no	yes	no	no	no	yes
lossless text/binary duality	n-a	n-a	n-a	n-a	no	yes	no	no	yes
extensions that degrade legibly	no	no	partial	no	partial	partial	no	yes	yes
no evaluation at parse time	yes	no	no	yes	yes	yes	yes	yes	yes
content-addressed definitions	no	no	no	no	no	no	no	partial	yes

What each column was checked against

column	version	checked-on	against
json	RFC 8259	unchecked	docs/01-survey.md
yaml	YAML 1.2.2	unchecked	docs/01-survey.md
xml	XML 1.0, 5th ed.	unchecked	docs/01-survey.md
toml	TOML 1.0.0	unchecked	docs/01-survey.md
cbor	RFC 8949	unchecked	spec/unimsg-v0.umsg, which profiles it
ion	Ion 1.0	unchecked	docs/01-survey.md
protobuf	proto3	unchecked	docs/01-survey.md
rdf	RDF 1.1	unchecked	docs/01-survey.md
unimsg	v0, provisional	this-build	spec/unimsg-v0.umsg, parsed by this build

Lossless duality and degradable extensions. Everything else in that table is available somewhere already, and a format offering only the rest would not be worth writing.

The reasoning under each column

From the survey, unedited. A column in that table is a summary of a paragraph someone wrote after reading the format's specification; the paragraph is the part worth arguing with.

json

The reigning champion by accident.

No integer type. Numbers are IEEE-754 doubles in practice; 2^53 + 1 silently

corrupts. Fatal for identifiers, ledgers, and scientific data.

No binary. Base64 costs 33% and obscures intent.
No date/time type. Every application invents one.
No NaN or Infinity. Round-tripping numeric computation is lossy.
No comments. The single biggest social failure; it is why every JSON-based

config format eventually gets replaced.

Duplicate keys undefined. Real interoperability hazard.
No cycles or shared references. Graphs must be hand-encoded.
No streaming framing, no chunking, no random access.
No extension mechanism at all. Every extension is a naming convention

({"$type": …}) and every such convention collides with someone else's.

The last point is the decisive one. JSON is not extensible; it is merely permissive.

yaml

JSON plus ergonomics plus footguns.

Anchors and aliases give genuine DAGs and cycles — a real advantage nobody credits.
Indentation encodes shape, which is why it reads well.
The specification is enormous and implementations diverge; YAML 1.1 vs 1.2 is still

unsettled in the wild.

Value-dependent coercion: NO parses as false (the "Norway problem"). Any

coercion rule that depends on the *value* rather than the *lexical form* is a defect.

yaml.load was a remote-code-execution primitive in Python for a decade.

Better than JSON for humans, worse for machines, not a candidate for a wire format.

xml

Genuinely good at *documents*, genuinely bad at *data*.

Mixed content, attributes, and namespaces are real strengths for prose-like data

and are underappreciated.

Everything is a string until XSD bolts types on.
Attribute-vs-element is a permanent, unresolvable modelling coin-flip.
Whitespace handling is ambiguous.
Closing tags roughly double glyph count without adding information — the

signal-to-noise ratio is visible from across the room, and this is why it lost

socially.

Security legacy: XXE, billion-laughs entity expansion. Both are consequences of

resolving and expanding things at parse time, a mistake this project treats as

categorically prohibited (see 04).

toml

Config-scale formats. TOML's nesting story degrades badly beyond two levels. None

target universality and none should be judged against it. KDL is worth studying purely

for syntax ergonomics.

cbor

The two closest existing things to unimsg. Treated in detail in

07-foundation-cbor-ion.

Briefly: CBOR (RFC 8949) has a small typed kernel, native binary, an IANA tag registry,

a deterministic encoding profile, a schema language (CDDL, RFC 8610), a signing layer

(COSE), a content-addressed dialect (DAG-CBOR), and very broad deployment — WebAuthn /

FIDO2, the EU digital COVID certificates, and Bluesky's repository format. Ion has

arbitrary-precision decimals, proper timestamps, stackable annotations, symbol tables,

s-expressions, and a lossless text/binary duality.

Between them, roughly 80% of unimsg already exists.

ion

The two closest existing things to unimsg. Treated in detail in

07-foundation-cbor-ion.

Briefly: CBOR (RFC 8949) has a small typed kernel, native binary, an IANA tag registry,

a deterministic encoding profile, a schema language (CDDL, RFC 8610), a signing layer

(COSE), a content-addressed dialect (DAG-CBOR), and very broad deployment — WebAuthn /

FIDO2, the EU digital COVID certificates, and Bluesky's repository format. Ion has

arbitrary-precision decimals, proper timestamps, stackable annotations, symbol tables,

s-expressions, and a lossless text/binary duality.

Between them, roughly 80% of unimsg already exists.

protobuf

Fast, compact, and not self-describing. Bytes are meaningless without the schema,

which makes them transports rather than storage formats.

Protobuf field numbers are a durable but fragile coordination mechanism.
Protobuf 3 spent years discarding unknown fields, silently breaking forward

compatibility for anything in the middle of a pipeline. This is the strongest

available evidence that **unknown-preserving round-trip must be a conformance

requirement**, not a nicety.

Avro's schema-resolution rules are the best thought-out evolution story in this

family and are worth studying.

rdf

RDF provides global identity via IRIs and a genuine graph model. The ecosystem is

heavy and RDF/XML is a canonical example of expressive-but-unreadable.

IPLD / DAG-CBOR is the most relevant: content-addressed Merkle DAGs, meaning a

document can reference data larger than itself, verifiably. This capability is

adopted directly in these designs.

unimsgnot in the survey

The survey does not discuss this format on its own. The column stands on the rows alone.

Three of them are not rivals in the same way

on cbor

unimsg is CBOR. The kernel is RFC 8949's, the canonical encoding is a deterministic profile of it, and a unimsg document decodes in any CBOR library on earth. What is added is a text syntax that projects it losslessly, and a discipline for extensions. Presenting CBOR as a rival misstates the relationship: it is the foundation, and the site says so on the comparison page rather than in a footnote.

on ion

Amazon Ion had the text/binary duality first, and had it properly. The differences are the degradable-extension rule and the content-addressed definitions, plus a much smaller surface. Ion's decimals and timestamps are better than most of what follows them.

on json

JSON is not extensible; it is permissive. Every extension it has is a naming convention, and every naming convention collides with someone else's. That is the sentence the comparison page is built around.

Where it loses

In the same table, same size, not in a collapsed section. A comparison page that only wins is read as advertising and discarded whole, including the true parts.

against	loses-on
JSON	ubiquity, and it is not close. Every language parses JSON in its standard library and none parses unimsg.
Protobuf	size and speed on a known schema, permanently. Self-description costs bytes and always will.
Parquet	anything at scale. unimsg has no columnar story and no random access into a large value.
YAML	familiarity in configuration, where its footguns rarely fire and its ergonomics are known.
XML	mixed content. Prose with markup inside it remains XML's, and this document does not pretend otherwise.

What this site is not doing

a package registry, a *hosted* validator, or accounts of any kind. The emphasis matters and was missing until a reader asked why a format's site would refuse to tell anyone whether their document is valid. It does not refuse: the playground parses, encodes and round-trips in the reader's own browser, which is validation, and /doc reports the same. What is ruled out is a validator *we run* — an endpoint, a queue, an abuse surface, and a service to keep alive, none of which a static site has and all of which it would need. A checker that runs on the reader's machine costs a download; one that runs on ours costs forever.
a comparison that flatters unimsg by choosing weak opponents — CBOR and Ion are the closest things to it and are treated as the serious rivals they are
claiming adoption. There is none, and a site that implies otherwise is lying about the only fact a reader can check in ten seconds