Design Space Map

Design Space Map — Morphological Analysis

This table shows all design dimensions and their possible values. The chosen combination is highlighted.

#	Design Dimension	Option A	Option B	Option C	Option D	Chosen
1	IR as Compilation Target	Direct Authoring → Runtime	Authoring → flowJson → Runtime	Authoring → IR → Multi-Target	Authoring → IR → Single-Target	C
2	Architecture Layers	Two Layers (IR → Runtime+Pipecat)	Three Layers (IR → Runtime → Pipecat)	Four Layers (IR → Policy → Runtime → Pipecat)	Pipecat + Controller Overlay	B
3	Agent Communication	Single report_observation	Multiple Specialized Functions	Free-Text + Runtime Parsing	Structured Output (JSON mode)	A
4	Evidence Signal Model	Discrete Signal Kinds Only	Continuous Confidence Only	Discrete + Confidence + Process Taxonomy	Rubric-Level Judgments	C
5	Policy Expression	Structured Data Objects	Code (Python/TS)	Prompt Instructions	Rule Engine (DSL)	A
6	State Management	Explicit State Machines Only	Event Sourcing Only	Hybrid (Explicit SM + Event Sourcing)	Reactive State	C
7	Agent Autonomy Gradient	Two Levels	Three Levels (Autonomous/Advisory/Controlled)	Five Levels	Full Autonomy + Post-Hoc	B
8	Event Protocol	Push Only	Pull Only	Event Sourcing Only	Hybrid (Push + Event Store)	D
9	Evidence Output	Separate EvidenceLedger	Embedded in Transcript	Post-Hoc Extraction	Hybrid (Real-Time + Enrichment)	A
10	Recovery Strategy	Fully Automated	Human-in-the-Loop	Categorized (Technical/Assessment)	Graduated (Severity-Based)	C
11	Transition Authority	Runtime Only	LLM Proposes + Runtime Approves	LLM Decides	Runtime + LLM Delay	B
12	Context Management	RESET Only	APPEND Only	Hybrid (RESET + Summary)	Policy-Driven (RESET default)	D

Design Configuration Summary

The chosen design configuration represents a specific point in the design space:

{C, B, A, C, A, C, B, D, A, C, B, D}

This configuration is characterized by:

• Determinism over flexibility: Policies are structured data, transitions are runtime-approved, evidence is validated before recording
• Separation of concerns: IR ↔ Runtime ↔ Pipecat are cleanly separated; evidence collection ↔ evidence evaluation are separated
• Defense in depth: Multiple layers of constraint (IR policies, runtime guardrails, output validation, context policy)
• LLM as sensor, Runtime as judge: The LLM observes and reports; the Runtime validates and enforces
• Auditability as first-class: Events, evidence signals, and policy decisions are all recorded with provenance

Cross-Dimension Dependencies

Dimension Pair	Dependency	Nature
1 (IR) → 2 (Layers)	IR enables three-layer architecture	IR provides the semantic richness that justifies a separate Runtime Controller
2 (Layers) → 7 (Autonomy)	Three layers enable three autonomy levels	The Runtime Controller layer is the enforcement mechanism for the advisory/controlled boundary
3 (Communication) → 11 (Transition)	Single function bundles transition proposals	evidenceSufficient in report_observation is the LLM’s transition proposal
4 (Evidence) → 9 (Evidence Output)	Rich taxonomy requires first-class ledger	The eight signal kinds and provenance fields require a dedicated storage mechanism
5 (Policy) → 6 (State)	Structured policies drive state transitions	CompletionPolicy and TransitionPolicy are evaluated by the state machine
7 (Autonomy) → 11 (Transition)	Advisory level enables LLM proposals	The LLM’s evidenceSufficient signal is an advisory action
8 (Events) → 9 (Evidence)	Event protocol delivers evidence signals	evidence_signal events flow through the event protocol to the ledger
10 (Recovery) → 12 (Context)	Recovery affects context management	After recovery, context may need to be reconstructed from events

Rejected Combinations

Combination	Why Rejected
{A, _, _, _, _, _, _, _, _, _, _, _}	No IR = no versioning, no multi-target, no compile-time validation
{_, _, _, _, C, _, _, _, _, _, _, _}	Prompt-based policies are not enforceable (Greshake et al., 2023)
{_, _, _, _, _, _, _, _, _, _, C, _}	LLM-controlled transitions threaten fairness and reliability (Joughin, 1998)
{_, _, _, C, _, _, _, _, _, _, _, _}	Free-text parsing is unreliable for high-stakes assessment
{_, _, _, _, _, _, D, _, _, _, _, _}	Full LLM autonomy without runtime guardrails violates safety requirements

Interactive Visualization

For the full interactive design space visualization with QOC evaluation matrices, dependency graphs, and cross-cutting themes, see the Design Alternatives section.