Public Technical Preview · Contextual NPC interpretation, governed behavior, and bounded adaptive learning
Spanda is an experimental Unity-based NPC behavioral control layer exploring observer-specific interpretation, directed relationship state, governed consequences, deterministic movement, and bounded adaptive learning. This repository documents the working mechanisms and their limits; the implementation remains private.
Watch the expanded Spanda technical preview on YouTube
Publicly available on YouTube. The recording covers governed movement, consequence-path safe stopping, and adaptive learning through conditional reuse.
Follow the demonstration guide for the walkthrough.
The Unity Developer Preview brings observer roles and diagnostic controls into one inspectable scene. Select the image for full resolution.
An NPC can interpret a remark without being entitled to change the world. Spanda explores how to connect contextual interpretation to inspectable behavior—and how recurring interpretations can become reusable shortcuts without granting them unconditional authority.
Interpretation is not authority. Meaning, proposed consequence, authorization, committed change, and executed behavior are separate stages.
flowchart TD
O[Observer experience and accessible context] --> R[Interpretation routing]
R --> I[Interpretation record]
I --> M{Matching authored consequence?}
M -->|No| S[Stop consequence path]
M -->|Yes| P[Relationship change proposal]
P --> G[Governance evaluation]
G --> A[Explicit authorized Apply]
A --> C[Committed directed relationship state]
C --> X[Separately started deterministic movement trial]
L[Installed rule and applicability guard] --> R
Routing checks for a usable installed rule, then tries the authored path, and requests a bounded local Qwen interpretation when unresolved. The model supplies a label and summary. Authored mappings define consequences; governance evaluates proposals; explicit Apply commits state. A separate control starts movement from that committed state.
John / governed approach. Qwen interprets Jimmy’s reassurance through the trusted observer’s context. A matching credible_reassurance result supports an authored consequence, followed by governance, explicit Apply, and a separately started approach trial.
Robert / Qwen safe stop. A valid expected_concealment interpretation has no matching consequence in Robert’s scenario. The consequence path stops; this is neither model failure nor governance denial.
Robert’s Qwen interpretation resolves to expected_concealment. No authored consequence matches: no relationship proposal, no state change, and no movement. Select the image for full resolution.
Robert / authored withdrawal. A separate authored dismissive_reassurance scenario demonstrates governed commitment followed by deterministic withdrawal. It is not a Qwen result. John and Robert’s generative scenarios use different utterances and observer contexts.
Experience → Learn → Install → Prove → Evaluate → Refine → Compare
Actual model-call occurrences support a recurring structural candidate. Authored diagnostic validation outcomes support a rule proposal; governance and explicit Apply install the shortcut. Later authored performance outcomes support a separately governed applicability refinement.
| Case | Supplied condition | Interpretation route | Qwen calls for that case | Displayed result |
|---|---|---|---|---|
| I | CALM | Learned shortcut permitted | 0 | expected_concealment |
| J | ELEVATED | Shortcut withheld; normal path escalates to Qwen | 1 | expected_concealment |
The installed guard permits the CALM shortcut and withholds it for ELEVATED. Both displayed results are expected_concealment; the routes differ. Select the image for full resolution.
Refinement preserves the learned label while restricting reuse. Matching uses caller-supplied structural metadata. Validation, performance outcomes, and CALM/ELEVATED conditions are authored diagnostic inputs, not independently gathered empirical evidence. Live model outputs are not guaranteed to repeat these results. See adaptive learning for the full mechanism.
| Read | Covers |
|---|---|
| Architecture | Boundaries, data flow, state ownership, and responsibilities |
| Contextual interpretation | Observer context, Qwen, routing, and failures |
| Governed movement | Consequences, governance, Apply, movement, hold, and Restore |
| Adaptive learning | Recurrence, installation, review, and conditional reuse |
| Demonstration guide | Walkthrough with diagnostic screenshots |
| Technical boundaries | Experimental scope and integration requirements |
| Claims audit | Implemented mechanisms, visible evidence, and limits |
| Project status | Availability, evaluation, and collaboration |
The Spanda Architectural Framework develops a broader architecture. Relevant v0.2 papers are Paper 7 — Experience Entry Contract, Paper 5 — Tone Vector System, Paper 2 — Memory Stratification, and Paper 9 — Reference Architecture & Interface Contracts.
They provide background on admissibility, representation, persistence, and module boundaries. The architecture documentation explains their relationship to this preview; the demonstration does not implement or validate every module or contract in the papers.
Spanda is functioning experimental software, not a production-ready system. Private source is not publicly downloadable.
The earlier public browser demonstration covers deterministic social spacing without requiring a language model. Its scope is narrower than the expanded Qwen and adaptive-learning preview.
Spanda is under active development.
The implementation repository remains private, but opportunities for technical collaboration, research, integration, and licensing are welcome.
Potential areas of interest include:
- NPC behavioral systems
- Contextual interpretation
- Governed AI integration
- Bounded adaptive learning
Private technical evaluation may be considered by arrangement. Access to the private implementation is not automatic.
Inquiries: putmanmodel@pm.me.
The underlying Spanda implementation is proprietary.
This repository provides technical documentation and demonstration materials, not an open-source software release.
See Rights & Permissions.