aweft identity, v0
Status: draft. Normative for the parts marked so; the encoding is open.
Every observable carries an id. Ids are how two replicas agree on which observable a change refers to without agreeing on where it sits in the tree.
1. Requirements
An id MUST be unique within a document, and MUST remain valid when the observable it names is detached and reattached.
Ids MUST be generated from a cryptographically secure source. An implementation MUST NOT provide a fallback to a non-cryptographic generator, and MUST NOT expose a way to replace the generator at runtime.
Rationale: a global generator override exists to make tests deterministic, and the failure it produces is that tests run on one source of randomness while production runs on another, so no test can observe the weakness. Determinism in tests is obtained by injecting a generator at construction, not by replacing a global.
2. Ids are not credentials
An id MUST NOT be used as an authentication token, session identifier, or capability.
Rationale: ids are published. Every id in a synchronized document is visible to every participant, by design. A value that is broadcast cannot also be a secret. Credentials are minted by a separate mechanism with a separate call site, so that reaching for the wrong one is not expressible.
3. Encoding
An id is 12 bytes, 96 bits from the source in section 1. Its textual form, for a URL, a log line, or a key in a document, is base64url with no padding: exactly sixteen characters drawn from A-Z, a-z, 0-9, - and _.
Twelve bytes is a multiple of three, so the textual form carries no padding to strip and no character needing an escape anywhere an id is likely to be written.
Rationale, measured: ids are 35.5% of the raw bytes of a representative editing stream, so the width is the largest single lever on the size of the format. Measured on that stream against 16 bytes, 12 bytes is 6.1% smaller compressed and 8 bytes is 11.8% smaller.
Twelve was chosen over eight because a collision is silent, permanent, and corrupts data with no recovery path, and 64 bits reaches a probability at scales an application can actually get to: a document of a million observables sits at roughly one in 37 million, and a hundred thousand such documents at roughly one in 370. At 96 bits that same fleet is around one in a trillion. Twelve was chosen over sixteen because sixteen buys nothing that 96 bits has not already bought, and costs 6.1% of every byte sent, forever.
Width defends against accidental collision only. A participant that wants a collision can simply choose one, so no id width is a defence against a hostile writer. That belongs to the authority layer, and it is why 96 bits is enough here.
3.1 Ordering
An id MUST NOT carry a timestamp or a counter, and MUST NOT be ordered by anything but its bytes.
Rationale: every id in a synchronized document is visible to every participant by design, so a timestamp component publishes the creation time of every object to everyone who can read the document, including anyone it is shared with later. What a timestamp buys is sortable storage keys, and storage keys are local, unshared, and free to carry one.
Array ordering does not use ids at all. See spec/format.md section 6.6.
4. Open
Whether an id ever needs re-minting, and what a replica holding the old one should do.