Connected compilation now uses a dedicated server-owned task contract while local model modes retain local compilation.
Compiled output is schema-checked and every model-supplied source ID is validated against Phasoric's bounded source manifest before local persistence.
Recompilation creates retained, linked versions. Unchanged source fingerprints reuse the current version without another model request.
Newly compiled domains stay inactive until reviewed and enabled.
Compilation now allocates one 58K source budget adaptively: short notes are included whole and longer notes use representative excerpts across the document.
Full source fingerprints detect changes outside an excerpt, while source deltas record additions, changes, and removals for future reconciliation.
Connected compilation validates and preserves explicit Phasoric source IDs, including adversarial delimiter-injection defenses.

Intelligence Chat and Agent Studio now share one active-domain selector.
Matched concepts, decision rules, terminology, and bounded raw evidence can reach the model context.
Disabled and superseded domains are consistently excluded.
Aggregate retrieval diagnostics expose the mix of compiled and raw context without sending note content or titles as analytics parameters.

Compiled domains show their version, compiled-source coverage, partial status, and provenance-warning count.
The recompile action reports when the current fingerprint is unchanged.
Knowledge Compiler documentation no longer makes unsupported numerical efficiency or perfect-provenance claims.
Research Studio reports compilation character coverage, representative excerpts, and source deltas.
Paid users can retain the Phasoric Skill JSON format or export a real Portable Agent Skill ZIP with safe paths and collision-resistant filenames.

Paid plans can create, edit, duplicate, export, import, disable, and delete reusable Prompt Recipes stored in the active local workspace.
Recipes are available from Intelligence Chat, Agent Studio's shared execution path, the editor, Capture Inbox, Meeting Memory, and installed Feature Packs.
Connected recipe execution uses a dedicated server-owned prompt policy and a server-side plan gate; user recipe text remains untrusted input to that policy.
Personal AI Memory is visible, editable, scoped, disableable, and removable. Inferred entries remain disabled until explicitly confirmed.
Connected chats that include Personal AI Memory use a separate server-gated task contract instead of relying only on a client-side plan check.
Recipe results never mutate notes directly. They remain read-only, use explicit editor actions, or enter Forge as reviewable proposals.