Architectural framework for constrained agency, runtime governance, symbolic drift analysis, and structured emergence.
Spanda explores how NPCs in Unity interpret the same event differently based on their relationship history. The technical preview shows how an interpretation can lead to a governed response and how an approved learning rule can be reused under specific conditions.
- Watch the published technical preview and its video walkthrough.
- Play the earlier browser demonstration to see how relational conditions change NPC spacing in a fixed scene.
- Inspect the evidence in the claims and limitations, architecture papers and normative harness with paired proof and break fixtures.
CDE detects deviation, Kingpin decides whether an operation is authorized, and the gateway enforces that decision. The local MCP developer preview shows both a permitted fixture read and a refusal when a lease is required.
- Run the local MCP developer preview, or try the browser demonstration to explore how attention and authority change during a scenario.
- Inspect the runtime in the gateway architecture and limited SDK.
- Evaluate the evidence with the Agent Governance Test Harness, a separate evaluator for authority transitions such as revocation and retry. Its documentation distinguishes recorded assessments, synthetic controls and live qualifications for specific integrations.
Spanda and Kingpin licensing, integration and research inquiries — putmanmodel@pm.me.
The PUTMAN Model™ is an evolving architecture for building agents and simulations that remain legible under pressure.
It is centered on a simple idea: interesting behavior should emerge from bounded local functions, not from unconstrained generation or hidden-state wishful thinking. The goal is not to fake emotion or hand-wave “intelligence,” but to define the conditions under which agents can accumulate state, adapt, drift, recover, and act under explicit limits.
This work spans memory stratification, runtime governance, deviation tracking, narrative coupling, semantic reuse, multi-agent field behavior, and weak-signal interpretation. Together, these papers and demos form a stack for building systems that are more inspectable, more bounded, and more capable of meaningful behavior inside constrained worlds.
-
PUTMAN Model — Core Architecture for Contextual Meaning Reconstruction
https://doi.org/10.5281/zenodo.15634339 -
Resonant Field Mapping — A Non-Mimetic Empathy Layer for Language Models
https://doi.org/10.5281/zenodo.15701018 -
Spanda Architectural Framework
Papers, companion notes, bridge papers, and demos for the broader stack:
https://github.com/putmanmodel/spanda-architectural-framework
This is builder-facing work.
I am building toward agents that can operate inside constrained, inspectable worlds rather than simply remixing language or simulating coherence after the fact. The aim is to provide foundations that stronger builders, larger labs, and better-resourced teams can actually use.
My value is in architecture, framing, cross-paper coherence, and identifying the control surfaces that make richer systems possible without collapsing into noise, script, or silent drift.
A compact summary of the design philosophy:
No constraint = noise
Too much constraint = script
Useful emergence = structured freedom
The current stack is centered around the paper series and companion notes collected here:
https://github.com/putmanmodel/spanda-architectural-framework
-
Constraint-grounded inference
Agents should infer only from admissible structure, not from convenient projection. -
Memory stratification
State is tiered across short-term, intermediate, and structural layers so persistence is earned, not silently absorbed. -
Constraint Deviation Engine (CDE)
Deviation is treated as an inspectable, governable signal rather than a vague sense that “something is off.” -
Reflex Attention
Short-horizon salience can redirect attention without automatically mutating durable state. -
Experience Entry Contract (EEC)
Promotion into persistent state remains governed, explicit, and boundary-aware. -
Tone Vector System (TVS)
A typed representation layer for structured state exchange, replay, and interoperability. -
Narrative Coupling / Shared Representation
Systems can accumulate thematic pressure and diverge in interpretation even when representation overlaps. -
Multi-agent field dynamics
Local interactions can propagate, amplify, dissipate, and destabilize across groups without requiring magical global knowledge. -
Companion extensions
Additional papers extend the stack into self-relevance, cross-environment semantic continuity, and trajectory-level drift detection.
A growing part of the project is showing that the same weak-signal / layered-interpretation architecture transfers across serious application domains.
Current bridge papers include:
-
Weak-Signal Interpretation for Low-Altitude Drone Detection
Defensive, simulation-oriented sensing architecture for weak acoustic, visual, thermal, RF, and motion-derived partials. -
Governance-Aware Weak-Signal Architecture for Grid Anomaly Triage
Operator-facing early-warning architecture for energy-system anomaly handling under uncertainty. -
Weak-Signal Interpretation for Industrial Fault Detection and Predictive Maintenance
Stratified early-warning architecture for handling weak equipment precursor signals before they collapse into alarms, maintenance work, or ignored noise.
This is one of the main practical claims of the broader architecture: weak-signal handling should be portable across domains as different as low-altitude drone detection, grid anomaly triage, and industrial fault detection, without collapsing attention, interpretation, memory, and authority into one opaque step.
- Runtime LLM governance
- Agent-state monitoring
- Symbolic drift detection
- Memory-safe agent architectures
- Multi-agent escalation analysis
- NPC and simulation behavior modeling
- Structured emotional field systems
- Constrained generative environments
- Human-in-the-loop governance layers
- Weak-signal detection and anomaly triage across physical and operational domains
The main architecture stack repo containing the paper series, companion notes, bridge papers, and demos.
https://github.com/putmanmodel/spanda-architectural-framework
Core conceptual framework for contextual meaning reconstruction and layered interpretive structure.
A non-mimetic empathy/control layer for modulating language-model behavior under bounded persistence and runtime constraints.
Deviation and drift-oriented monitoring for contradiction, mismatch, escalation, and trajectory change across turns and state updates.
A simulation-facing layer for affective field behavior, propagation, memory interaction, and bounded adaptation.
Small demonstrators focused on leases, deny-by-default control, reconvergence, reversibility, and inspectable runtime authority.
-
PUTMAN Model — Core Architecture for Contextual Meaning Reconstruction
https://doi.org/10.5281/zenodo.15634339 -
Resonant Field Mapping — A Non-Mimetic Empathy Layer for Language Models
https://doi.org/10.5281/zenodo.15701018
- Spanda Architectural Framework
https://github.com/putmanmodel/spanda-architectural-framework
This is a solo effort built with limited resources, limited lab access, and a strong bias toward architecture that can survive contact with reality.
I am interested in:
- constructive technical feedback
- serious implementation critique
- opportunities to help other builders work from a stronger architectural foundation
- shared resources, collaboration, or practical paths to growth
I am not presenting this as a polished corporate lab. I am trying to build something structurally sound enough to be useful beyond me.
See the profile repository LICENSE for its noncommercial sharing and adaptation terms, attribution requirements and commercial permission contact. Each linked project has its own terms.
Stephen A. Putman
Email: putmanmodel@pm.me
GitHub: @putmanmodel
X: @putmanmodel
Reddit: @putmanmodel
BlueSky: @putmanmodel.bsky.social
