Phase graph — the working model creativity inherits [Concept]

The Phase Reasoning Graph (PRG; Phase Graph in compact UI copy) is the descriptive layer of graph thinking: a mined working model of the world's meta-machinery — its patterns, conventions, attractors, agents, rules, pressures, and landmarks. It answers "how does this world WORK?" where the coordination plan answers "how do we navigate it across arcs?" and the per-arc CRG answers "how does this arc reason inside it?" (reasoning-graph-engine). Opt-in at exactly one lever: the operator generates a graph and sets it current; every creative surface then inherits it uniformly, and with no active graph every surface renders nothing.

flowchart TB
    Gen["**generatePhaseGraph**<br/>mined from domain context<br/>(optional seed: a prior graph · optional guidance)"]
    Gen -->|"immutable once stored"| Store[("domain.phaseGraphs<br/>+ currentPhaseGraphId (the ONE opt-in lever)")]

    Store -->|"getActivePhaseGraph<br/>(none active → '' everywhere)"| Embed

    subgraph Embed["buildActivePhaseSection — SELF-EMBEDDED by each engine fn, per scope"]
        direction TB
        D["descriptive data block<br/>&lt;mode&gt; — 7 node types with temporal stance"]
        A["scope-tailored application directive<br/>&lt;mode-application&gt; — same semantics everywhere"]
        D --- A
    end

    Embed --> S1["expand · arc-direction"]
    Embed --> S2["reasoning-arc · reasoning-plan"]
    Embed --> S3["scene-structure · scene-plan · scene-prose"]
    Embed --> S4["perspective · variables (Present + Reading cohort)"]

    Store -.->|"arc.phaseGraphId pins the model<br/>the arc was generated under"| Arcs[("arcs — reference-counted GC")]

    X["Position monitor · Opinion — evidence surfaces,<br/>DECLINED by decision (readout purity)"]
    Embed x--x X

Invariants

• Descriptive, never prescriptive. The PRG states what the machinery is; the coordination plan and CRG decide what to do with it. Node types carry temporal stance (a pressure accumulates, an attractor pulls, a landmark anchors) and downstream prompts read the stance, not just the label.
• One opt-in lever, uniform reach. Every embedding surface calls buildActivePhaseSection itself (the engine self-embeds — never a caller-threaded parameter), so UI and headless callers can never diverge. phase-coverage-guard.test.ts pins the exact surface map and the declared scopes.
• Evidence surfaces are excluded by decision. The Position monitor and Opinion selection trace evidence; pre-embedding a creativity substrate there would bias the read toward the model's own priors. The exclusion is pinned in the guard — changing it is a product decision, not an edit.
• Immutable + pinned, refreshed by hand. A graph is never edited; regeneration mints a new one (optionally seeded by the old). Arcs pin the graph they were generated under (arc.phaseGraphId), and a coordination plan pins the model its backward induction reasoned under (plan.phaseGraphId); unreferenced graphs are pruned. There is no automatic refresh — but the decay is no longer silent: the plan view runs a pure staleness read (coordinationPlanPhaseDrift) and tells the author when the active model has moved past the plan. The author decides; nothing regenerates itself.
• Rupture is named, never silent. Higher-priority inputs may break a phase rule; every scope's directive requires marking the rupture (a chaos node, a superseding delta, named subtext) so a deliberate break never reads as drift.

Where it lives: src/lib/engine/ai/phase-graph.ts (generation + the embed helper), src/lib/engine/prompts/phase/ (data block · scope directives · hierarchy entries), src/lib/core/graph/phase-graph.ts (active pointer · reference counting · GC), src/__tests__/phase-coverage-guard.test.ts (the pinned surface map). Siblings: reasoning-graph-engine · projection-anatomy.