RFC-0022 claim projection review
Versioned draft snapshot
RFCs, reference catalogs, and evidence reflect Spec 3b2e3e7a114fb826b1445d39a48756b8fff629cd. Counts and pending/status notes belong to that snapshot. Read the dated project update for later progress.
Source repositories are private during staging. Links marked “repository access required” need access to the linked repository.
This is a trace and gap record from the pinned draft specification. It does not establish proof or implementation conformance.
Status: all 68 claims reviewed on 2026-10-05. Owners: Runtime, AI. Current totals: 41 source matches and 27 whole-node mismatches. No normative facts changed.
Composition and bounded execution
The initial group covers agent/module composition, static data flow, actor binding, capability limits, observability, diagnostic ranges, and isolation. The cross-instance variable claim lacks its exact invariant identity.
The finite-execution claim is stale: authored §10.2 now promises a finite bound on dispatched steps and explicitly excludes wall-clock termination. A human gate can remain open until a decision or a platform timeout outside this RFC. Preserve that distinction in RDF and public documentation.
Acceptance obligations
- Correct the stale finite-execution summary without claiming guaranteed wall-clock completion.
- Own or map the cross-instance variable invariant after consumer review.
- Verify concurrent execution isolation, actor-bound approval, typed step inputs/outputs, and complete structured event emission in the actual runtime.
- Test that orchestration cannot grant permissions absent from the invoked agent’s effective authority.
Source agreement does not prove a working workflow scheduler or runtime capability enforcement.
Declaration slots and error policies
The next ten claims have seven matches and three mismatches. Gates, capability limits, static structure, required descriptions, error policy variants, and optional workflow defaults agree with their authored definitions.
Three required nested slot identities are absent: gate approvedBy, step
agent, and step inputs. The grammar establishes their requirements but does
not supply RDF slot owners. Reconcile the nested declaration-slot model before
creating nodes or redirecting them to the parent declaration.
Acceptance must compile a step with a workflow-level default and reject an uncovered step without either local or default policy. Verify required nested properties and their diagnostics through the integrated parser and type checker.
Trigger, conditional, loop, and gate patterns
The next fifteen claims have eight matches and seven mismatches. Trigger variants, optional timeout, and gate lifecycle/events have authored identities. Nested step inputs/outputs and conditional/repeat declaration identities are absent at the claim IRIs.
Warning-only M22203 permits a later step to depend on a variable absent from one branch. Define missing-value typing and input availability or reject that dependence; otherwise the promised static data-flow safety has no executable meaning on the branch that never binds the variable.
Acceptance must execute both branches, check unavailable outputs, and verify gate approval/rejection against the authorized actor. Static loop bounds do not prove gate completion or workflow wall-clock termination.
Parallel settlement and error handling
The next fifteen claims have ten matches and five mismatches. Parallel steps settle without mid-flight cancellation before failure halts later work. Compensation targets, idempotency advice, and fail/continue policies agree with the source. Output-slot identity gaps recur in typing and rebinding claims.
The unconditional successor-output promise conflicts with continue after a
failed producer. Specify typed absence or restrict consumption to successfully
bound outputs; otherwise a statically typed downstream input can receive no
value. The loop’s finite wording also needs the authored bounded-step limit.
Acceptance must inject failures into parallel branches and verify all in-flight
work settles before termination, then test a failed producer with continue
and a dependent consumer. Compensation execution must be distinguishable from
normal flow and must not run twice through both paths.
Retry fallback, events, and authorization
The final eighteen claims have eight matches and ten mismatches. Required event fields, event triggers, transition coverage, retry wording, and capability limits match the source. Missing nested slots recur, along with unresolved error-policy and isolation invariant identities.
Section 3 permits procedure, event, and manual triggers. Section 8.4 requires authorization for an underlying action, while M22210 names an undeclared action and M22214 rejects direct action references. Specify authorization for each accepted trigger form and diagnostic precedence; the procedure-only RDF claim must be qualified to avoid excluding the other grammar variants.
The retry projection faithfully repeats §6.1, but that section does not say
what happens when an exhausted workflow-level retry falls back to itself.
Define a finite fallback so the bounded dispatched-step invariant is enforceable.
Also specify the terminal event after continue leaves a failed step:
workflow.completed currently requires every step to succeed.
Acceptance must exercise local retry exhaustion, inherited default retries, manual and event authorization, invalid procedure references, and terminal events after continued failures. Actor impersonation, concurrent variable isolation, and unbound agents require runtime tests once the scheduler exists.