🌀 Show & Tell: Phase 22.3 — ConceptBlender — Fauconnier-Turner conceptual blending engine

[web3guru888]

Phase 22.3 — ConceptBlender Architecture

The ConceptBlender brings Fauconnier & Turner's Conceptual Integration Network (CIN) theory into ASI-Build — the computational mechanism for genuine conceptual novelty.

The 4-Space Model

              ┌─────────────────────┐
              │    Generic Space     │
              │ shared_structure:    │
              │  {agent, instrument, │
              │   target, action}    │
              └──────────┬──────────┘
                    ╱          ╲
                   ╱            ╲
    ┌─────────────▼──┐    ┌──▼─────────────┐
    │  Input Space 1  │    │  Input Space 2  │
    │ surgeon         │    │ butcher         │
    │ {surgeon,       │    │ {butcher,       │
    │  scalpel,       │    │  cleaver,       │
    │  patient,       │    │  meat,          │
    │  operating-rm}  │    │  butchershop}   │
    │ frame:medical   │    │ frame:meat-proc │
    └────────┬────────┘    └────────┬────────┘
             ╲                      ╱
              ╲                    ╱
         ┌─────▼────────────────▼─────┐
         │      Blended Space          │
         │ {blended-agent, sharp-tool, │
         │  victim/target, workplace,  │
         │  ★ incompetence,            │
         │  ★ recklessness,            │
         │  ★ harm-through-negligence} │
         │ frame: botched-operation    │
         └──────────────┬──────────────┘
                        │
                  ★ EMERGENT ★
           (not in either input)

The starred elements (★) are emergent structure — they exist in the blend but were not present in either input space. This is the creative payload.

Composition-Completion-Elaboration Pipeline

  Input 1 ──┐                ┌─→ Composed Space
             ├─→ COMPOSE ────┤
  Input 2 ──┘   (project     └─→ (mapped elements fused)
                 via mapping)
                    │
                    ▼
               COMPLETE
               (fill frame defaults)
                    │
                    ▼
               ELABORATE
               (run the blend — simulate,
                extend, derive consequences)
                    │
                    ▼
            Blended Space + Emergent Structure

Each stage builds on the previous:

Stage	What it does	Weight	Example
Composition	Projects elements from both inputs into blend via CrossSpaceMapping	0.4	surgeon+butcher → blended-agent
Completion	Adds frame-implied elements not yet present	0.3	butcher-frame → rough handling, no care
Elaboration	Imaginative extension — runs the blend as a scenario	0.3	→ incompetent surgeon EMERGES

BlendType Classification

                        Same frame?
                       ╱          ╲
                    YES             NO
                     │               │
                  MIRROR       One frame-only?
                               ╱          ╲
                            YES             NO
                             │               │
                          SIMPLEX      High overlap?
                                       ╱       ╲
                                    YES          NO
                                     │            │
                              SINGLE_SCOPE   DOUBLE_SCOPE

DOUBLE_SCOPE blends are the most creatively productive — both frames contribute structure, creating the richest emergent potential.

Integration with Phase 22 Pipeline

  DivergentGenerator (22.1)           AnalogyMapper (22.2)
  ┌──────────────────────┐           ┌──────────────────────┐
  │ Generate candidate   │           │ Find structural      │
  │ concepts via SCAMPER,│           │ mappings between     │
  │ random association,  │           │ concept domains      │
  │ bisociation          │           │ (Gentner SMT)        │
  └──────────┬───────────┘           └──────────┬───────────┘
             │ MentalSpace[]                     │ AnalogicalMapping
             ╲                                   ╱
              ╲                                 ╱
         ┌─────▼───────────────────────────────▼─────┐
         │         ConceptBlender (22.3)               │
         │ • Accepts MentalSpace pairs                │
         │ • Uses AnalogicalMapping to seed            │
         │   CrossSpaceMapping                        │
         │ • Runs CCE pipeline                        │
         │ • Detects emergent structure                │
         │ • Returns scored Blend objects              │
         └──────────────────┬────────────────────────┘
                            │ Blend[]
                            ▼
         ┌──────────────────────────────────────────┐
         │       AestheticEvaluator (22.4)          │
         │ Scores blends for aesthetic/novelty      │
         └──────────────────┬───────────────────────┘
                            │
                            ▼
         ┌──────────────────────────────────────────┐
         │    CreativityOrchestrator (22.5)         │
         │ Routes best blends to downstream tasks   │
         └─────────────────────────────────────────┘

Key Design Decisions

1. Frozen dataclasses everywhere — MentalSpace, GenericSpace, Blend are all immutable. This makes caching safe and enables structural sharing.

2. asyncio.Lock on cache only — The CCE pipeline itself is pure computation (no shared mutable state), so only the cache write needs synchronization.

3. optimize_blend() hill-climbing — Deliberately simple (greedy accept). The mapping space is small enough that greedy works well.

4. Emergent structure as set difference — Clean, auditable definition. An element is emergent iff it appears in the blend but cannot be traced to any input or the generic space.

Prometheus Metrics

Metric	What to watch
asi_blend_total{blend_type}	Distribution across 4 types — healthy system should produce mostly DOUBLE_SCOPE
asi_blend_quality	Should center around COHERENT-EMERGENT range
asi_blend_emergent_structures_total	Growing = system is creative
asi_blend_novelty_score	Track drift — if novelty drops, inputs may be too similar
asi_blend_duration_seconds	P95 < 100ms for single blend

Grafana Alert

- alert: BlendNoveltyDrop
  expr: histogram_quantile(0.5, rate(asi_blend_novelty_score_bucket[15m])) < 0.3
  for: 10m
  labels:
    severity: warning
  annotations:
    summary: Median blend novelty below 0.3 — inputs may be too similar

Open Questions

1. AnalogyMapper → CrossSpaceMapping bridge: Should ConceptBlender accept an optional AnalogicalMapping from 22.2 to seed the cross-space mapping, or always compute its own?

2. Frame knowledge source: Where does _complete() get its frame knowledge from? Options: hardcoded frame library, WorldModel (13.1) lookup, or LLM-backed.

3. Multi-space blending: Fauconnier-Turner theory supports >2 input spaces (multiple blend). Should we plan for this in the Protocol signature?

---

Spec: #515
Phase 22 Planning: See Phase 22 planning discussion