A Grant is a signed IPLD blob:

subject: what is covered. A document (follows updates), a specific version (frozen DAG), or a path prefix (a subtree of a space).

audience: who may read. everyone | space members | a specific account key | a bearer secret hash.

epoch: the space's auth-chain version this grant was issued under (see Ordering).

scope flags: history: true|false. The default is false, forward-only from this epoch (see the redaction rabbit hole).

issuer + signature: an account with authority over the subject, or a delegate holding a grant-issuing capability, chained exactly like today's WRITER capability delegation.

Everything the system does today and everything the proposal wanted is an instance:

No envelope blob, no separate capability-for-reads type, no parallel policy channel. One kind, one verifier, one index.

A grant on a document covers the blobs reachable from it through owner-signed structure: change deps, ref heads, embedded file blobs. The existing blob_visibility_rules table evaluates this, unchanged. A grant never follows a bare CID mention. It never crosses an ownership boundary: an embed of someone else's doc renders only if the viewer holds their own path to it. And it is evaluated at indexing time, never at query time.

That last point is the whole performance story. blob_visibility (blob, space) generalizes to an access table (blob, audience). Every query surface (search, citations, feeds, listings, comments, sync, the blockstore itself) filters by joining that table against the caller's resolved audiences, exactly the shape of today's publicOnly filters. One evaluator runs at indexing time, and everything else is a dumb join. Public content takes the same fast path it takes today, with the same CDN-cacheable headers. The whole cost of the system lands on private reads only, which preserves the asymmetry that prior art says to protect.

Each space maintains an auth chain: a linear, hash-linked sequence of signed auth events (grants, revocations, membership changes, delegations). The head of the chain defines the current epoch. Ref already carries a generation counter, which is client-supplied and unvalidated today. This design promotes it into the epoch and makes the indexer enforce it.

Ordering is by chain position. Timestamps appear in blobs but carry zero security meaning. This resolves the time rabbit hole and fixes two live bugs: today the alive-vs-deleted decision and the visibility register are both last-writer-wins on unvalidated Ref timestamps.

Revoking one grant means appending a revocation. Removing a member means bumping the epoch. Standing grants to the remaining members are re-materialized automatically, and the removed member's access ends at the bump.

A replica serving private content must hold the chain up to the epoch it enforces. How stale it may be is an explicit serving policy, decided on purpose.

What revocation means is stated plainly, in the spec and in UI copy: honest servers stop serving, and whoever already has replicated bytes keeps them.

This is trusted-server enforcement. A grant constrains which principals your home server and its authorized replicas will serve. It does not give cryptographic secrecy. These design consequences keep it honest:

Sync respects grants. Private blobs replicate only to principals with an audience that covers them: your other devices, space members' servers, and gateways the space designates. The filtered-sync machinery exists (authenticated peers, filtered RBSR fingerprints, push allowlists). But today it answers the access question with a different rule than HTTP does, and Bitswap fails open on blobs with no visibility rows while the blockstore fails closed. Unification means one grant evaluator that both layers query, and both fail closed. See Network.

Uniform denial. Unauthorized and nonexistent are byte-identical responses on every surface, which closes the existence oracle. The .dagjson endpoint's "blob is not public" error message is today's counterexample.

Enforcement is not opt-in. Today every privacy gate is inert unless the operator sets -public-only, and a default desktop daemon serves all private content to any local caller. Grant evaluation becomes the only read path, and the flag disappears.

The encryption slot. A Grant may carry an optional wrapped content key. Nothing uses it in v1. Because the schema has it, Tahoe-style cryptographic privacy for high-sensitivity spaces can come later as an added layer, without a redesign.

Comments were the proposal's loose thread, and the grant model ties it off cleanly. Today a comment copies the target document's visibility into its own signed blob at creation, and nothing ever reconciles it. A doc that flips from public to private leaves its old comments publicly listed, and a flip from private to public strands the discussion. The comment blob even resolves the target's visibility with a fail-open default to public.

The fix starts from the fact that a comment is its own tiny document with its own author. The doc owner cannot grant read on someone else's comment, and the commenter shouldn't need per-reader grants. So a comment declares its audience at creation as a reference to another audience: target's-audience, meaning "whoever can read the target, evaluated at read time." Comment visibility then tracks document visibility automatically. When the doc goes public, the discussion goes public. When an outsider gets a doc grant, they see the discussion. A comment on a version-scoped share stays inside that share. A commenter who wants a narrower audience, such as a private note to the author, declares a narrower one, with the same statement kind and no special case.

Scope discipline from the critique: this is one feature, sharing. It is not a permissions platform.

Grant blob kind and verification, with audience everyone / space / account-key / bearer. This subsumes the READER role that the Capability enum never got.

An access table generalizing blob_visibility. publicOnly filters become audience joins, and both HTTP and sync query the same evaluator and fail closed.

An auth chain per space with epochs (promoted from Ref's generation). Revocation, member removal, and unpublish are defined against it, and timestamps are demoted to advisory everywhere.

Share-a-doc and share-link flows in the clients, riding the existing bearer-token and signed-assertion auth transports.

Ownership claims required on raw uploads, and index-time validation of visibility transitions (closing VULN-5's class for good).

Explicitly deferred: nested groups, the encryption layer, path-prefix grants beyond spaces, and cross-space federation of audiences.

The test of the design is that the table above has no third column. Nothing in the current system, and nothing in the sharing feature, needed a second concept. Public and private were always the same kind of thing: public is a grant with the widest possible audience.