goal-workgraph

name	goal-workgraph
description	Project current source-bound work into the smallest verifier-first owner graph only when decomposition changes execution. Omit the graph for one bounded operation; compress repeated classes, derive the frontier from current evidence, and invalidate stale graphs.

Goal WorkGraph

Mission

Project current controlling obligations into the smallest verifier-first graph that changes execution. The graph is an advisory, ephemeral view for $goal-actuating; it is not an authority source, event store, architecture selector, or recursive executor.

Admission

Do not create a WorkGraph for one bounded owner operation with one known verifier. Let $goal-actuating select that operation directly.

Use a graph only when decomposition changes at least one of:

• dependency order;
• repeated-class realization;
• proof fanout;
• safe read-only parallelism; or
• stopping behavior.

A graph that merely wraps one selected operation is ceremony.

Node

work_node:
  version: WN-v2
  node_id:
  run_id:
  kind: inspect | edit | verify | review | ask | block
  description:
  depends_on: []
  paths: []
  resource_keys: []
  owner_boundary:
  invariant:
  proof_surface: []
  verifier: []
  expected_evidence: []
  review_resolution_ref:
  resolution_decision_id:
  strategy: none | local-repair | replacement-kernel
  parallel_safe: true | false
  isolation: read-only | file-disjoint | serial
  status: pending | selected | running | passed | failed | blocked # derived view

Graph

work_graph:
  version: WG-v2
  goal_id:
  run_id:
  source_digest:
  nodes: []
  frontier_policy: verifier-first | highest-risk-first | dependency-order
  combine_policy: all-pass | proof-sufficient
  next_ready_node_ids: [] # derived from current evidence

Node status and next_ready_node_ids are projections, never continuation authority. Recompute them from the current controlling artifacts, kernel state, and evidence.

Compilation laws

1. Current control basis. Derive nodes only from the current GoalContract, selected review resolution or construction, current counterexamples, the canonical applicable route-exclusion projection, subject identity, path scope, obligations, and verifier bindings. Raw findings and suggested patches are never nodes.
2. Repeated-class compression. When many instances share one governing law and one construction, create one canonical owner edit node. Instance application or verification may fan out; do not duplicate architecture nodes.
3. Current-evidence frontier. Recompute the ready frontier after every observation. Cached node state or an earlier ready frontier cannot authorize continuation.
4. Whole-graph invalidation. A change to the source, selected construction, current counterexamples, canonical applicable route-exclusion projection, subject identity, paths, obligations, or verifier bindings invalidates the complete graph. Regenerate it rather than repairing stale graph state.
5. Distinct stop meanings. Keep failed execution, blocked progress, and invalid proof distinct. Stale, mismatched, fallback, or incomplete evidence does not become an implementation failure or a repair node.
6. Selected architecture only. Read-only experiments may compare candidate behavior, but architecture selection remains outside the graph. Only the already-selected construction may inform edit nodes. When the source is a review resolution, only its exact synthesis-owned selected_work_node may become an edit node.
7. Bounded parallelism. Parallelize only resource-disjoint read-only scout, review, or proof nodes. Shared-owner mutation remains serial and every fanout returns through $goal-actuating.

Procedure

1. Apply the admission rule; omit the graph when one direct operation suffices.
2. Preserve the current source, run, owner, subject, path, obligation, and verifier bindings.
3. Split only by owner boundary, invariant, proof surface, or real dependency.
4. Compress repeated instances that share one law and construction.
5. Group already-quotiented review findings into one resolution decision; never create one node per comment.
6. For review-resolution-derived work, create edit nodes only from the exact synthesis-owned selected_work_node of an already-selected local-repair or replacement-kernel decision; construction-only and node-free decisions do not produce edit nodes.
7. Make replacement-kernel work one serial canonical-owner node with explicit retirements.
8. Consume current counterexamples and the canonical applicable $negative-ledger route-exclusion projection before proposing inspect work; never infer or perform reopening locally.
9. Return the currently ready node IDs to $goal-actuating for lead selection.

Guardrails

• The graph cannot grant mutation or complete the goal.
• The graph cannot select architecture or reinterpret a selected strategy.
• Unknown review pressure routes back to classification or resolution.
• Do not persist graph state as a second control plan.
• Do not parallelize a shared owner boundary.
• Do not add refactor work unless it removes a shared cause or proof surface.