How ops reference other ops, and why the two natural alternatives were rejected. The op shape itself is FRAMEWORK.md "The op shape"; the byte layout is GRAMMAR_SPEC.md Part A.
- An op carries one flat reference set
refs(u)of full 32-byte ids. Its role fields (anchor, targets, overwrites) name members of that set; the rest are frontier pins. Naming is committing: a named id is a ref by construction. - The preimage stores
refs(u)once, as a sorted, unique table; role fields address it by index. Indices are encoding, never identity. - The named / pin split is positional and unflagged: nothing consumes it, so the artifact does not record it.
- In memory a node is stored normalized,
HashNode { pins, op }withpins = refs ∖ named(src/hash_node.rs);pins ∩ named = ∅is an invariant every constructor keeps and the decoders enforce (DecodeError::RedundantPin). - Values (payloads, keys, map values,
Place . placed_at) are ids too, but they are not refs: never in the table, never buffered on (HETEROGENEITY.md "Payloads are ids").
Proposal: the op commits only to its observed tips, and anchors, targets
and overwrites are positions in a canonical enumeration of the committed
closure S(u) = refs*(tips). Attractive because the wire already encodes
refs by index, it would shrink large multi-target preimages, and it would
make "refs cannot escape the commitment" hold by type.
Rejected because no enumeration is both deterministic and local:
- Sorted by id: resolving "the i-th id within
S(u)" means materializing ancestry, and positions carry no recency. - Frontier walk with dedup: honest refs become small integers, but at a merge a position depends on what the branches shared, so the enumeration cannot be maintained incrementally and repeated merges force repeated O(|S|) walks.
- Frontier walk without dedup (path addressing): incremental, but position widths grow exponentially with merge depth.
- The cold anchor: the first edit after clicking into an old document anchors thousands of ops deep. Every scheme needs a full-id "far pointer" for it, which is the id-based model again.
- Positional ops are not self-describing: a remove's targets cannot be read, filtered or relayed without the DAG.
What the proposal wanted is delivered elsewhere: canonical bytes (ENCODING_SPEC.md), one copy of each id per op (the refs table), and range compression of removes (the wire encoding).
Proposal: commit only to the tips; resolve named ids against the closure without committing them individually.
Rejected: a named id outside the commitment becomes representable, so
soundness needs a containment check named ⊆ refs*(tips). Checking presence
at apply is delivery-order dependent (two replicas can reach opposite
permanent verdicts), and exact containment is an ancestry oracle, expensive
exactly in the honest cold-anchor case. With naming as committing the check
does not exist: a cold anchor costs one id, resolved in O(1), and buffering
waits on refs(u).
An explicit tips field separate from the role fields would record which
refs the author called frontier. Nothing reads that partition, so it would
be committed bytes with no semantics; the flat set with roles as indices is
the shape.