← All articles

§ ARTICLE / Deep Dive

The Architecture of Experimentation: Inside the Deerfield Green Prototype Portfolio

Seven prototypes. One operating system. A modular ontology for organizational cognition.

The Deerfield Green Prototypes aren’t a menu of independent tools—they’re a cohesive modular ontology for complex systems. Each prototype addresses a distinct layer of organizational cognition: Augur and Palimpsest handle temporal dimensions (prediction and memory), Chremata and Baros quantify value and pressure, while Kitsune, Kitsune-B, and Pelagos define agency within ecological constraints. This essay walks through the architectural intent behind the portfolio, explaining why these seven specific prototypes were chosen and how they interconnect. The design emerged from watching teams optimize prediction models while historical data rotted in separate silos. Understanding this architecture isn’t about adopting seven tools—it’s about recognizing the cognitive layers your organization needs to operate complex systems effectively.

The Portfolio as a System: Intent and Architecture

Most R&D portfolios read like feature lists. Here’s what we built. Here’s what it does. Pick what fits. The Deerfield Green Prototypes refuse that framing. They’re designed as interconnected layers of a single operating system. Augur’s prediction feeds directly into Palimpsest’s historical queries. Chremata’s value signals constrain Kitsune’s agent training. Baros measures the pressure that Pelagos agents must navigate. This isn’t accidental architecture. We built it this way after watching three client teams optimize their prediction models while their historical data rotted in separate silos—each tool working, the system failing.

The portfolio targets four dimensions of organizational cognition: memory (what happened), value (what matters), pressure (what constrains), and agency (what acts). Seven prototypes cover these dimensions without overlap. Augur [4] and Palimpsest [9] handle temporal layers—looking forward and backward respectively. Chremata [5] and Baros [10] quantify economic and physical constraints. Kitsune [2], Kitsune-B [11], and Pelagos define adaptive agency within those constraints.

This modular ontology approach mirrors patterns in complex systems design, where decomposing systems requires properties and tools appropriate to each perspective [10]. But unlike abstract ontologies that promise flexibility without delivery [6], these prototypes are shipped code with concrete interfaces. Each has a README that states its purpose in one sentence. Each has an architecture diagram showing its data flow. Each can run independently, but none makes sense in isolation.

The etymology matters. Augur reads omens. Palimpsest reveals hidden layers. Chremata handles money. Baros measures weight. Kitsune shapeshifts. These names aren’t branding—they’re architectural documentation. When you understand what each prototype is named for, you understand what cognitive layer it addresses.

Temporal Layers: Prediction and Memory

Time splits the portfolio into two directions. Augur looks forward, reading prediction market signals like a Roman augur interpreting bird flight patterns [4]. Palimpsest looks backward, scraping away surface language to reveal the stratigraphic layers of corporate filings [9]. Together, they form a temporal engine—forecasting what comes next while preserving the context of what came before.

Augur maps Polymarket graph structures, making visible the interconnections that drive prediction. This isn’t just data visualization. It’s signal extraction from collective intelligence distilled into probabilities. The prototype treats prediction markets as modern augury, where no major decision proceeds without reading the omens. In organizational terms, Augur answers: What does the market expect?

Palimpsest answers: What did we commit to before? A 10-K filing is a corporate palimpsest—each annual report overwrites the last, yet beneath the current language lie traces of shifted strategy, evolving risk disclosures, and reclassified supply chain dependencies [9]. The prototype doesn’t just parse filings. It reveals the hidden layers underneath, allowing teams to query not just what a company says now, but how that statement differs from what it said before.

The tension between these two prototypes is deliberate. Prediction without memory becomes speculation—confidence without context. Memory without prediction becomes archaeology—preservation without action. Organizations need both. Augur’s forward-looking signals gain meaning when Palimpsest can query whether similar signals appeared in historical data. Palimpsest’s historical layers gain urgency when Augur detects market expectations shifting around those same patterns.

This temporal architecture reflects a deeper principle: complex systems require both state (memory) and transition functions (prediction). Dynamic modular ontologies allow organizations to build information models appropriate for their domain while evolving with changing requirements [7]. But evolution requires knowing what changed. Palimpsest preserves the change. Augur anticipates the next change.

Economic and Physical Layers: Value and Pressure

Value and pressure are the constraints that shape agency. Chremata quantifies value by reading earnings call transcripts and classifying them across five financial dimensions [5]. Baros measures pressure by blending prediction market sentiment with established conflict indicators [10]. One tells you what matters economically. The other tells you what constrains physically.

Chremata turns hours of analyst reading into structured, machine-readable signals. Every quarter, publicly traded companies hold earnings calls—hour-long conversations where executives report results and analysts probe for forward signals [5]. These transcripts contain critical information about financial performance, management sentiment, margin trends, and risk factors. But they’re buried in thousands of words of conversational text. Chremata’s NLP pipeline extracts this systematically, producing labels across optimism, caution, headwinds, and entity recognition [12]. The output isn’t sentiment analysis. It’s structured financial cognition.

Baros operates as a geopolitical barometer. Just as falling barometric pressure signals an incoming storm, rising values in this index signal escalating geopolitical risk [10]. The name carries dual sense: literal physical weight and figurative burden of consequence. This matters because organizational decisions don’t happen in vacuum. They happen under pressure—market pressure, regulatory pressure, competitive pressure. Baros makes that pressure measurable.

Together, Chremata and Baros define the constraint surface within which agents operate. Chremata answers: Where is value concentrated? Baros answers: Where is pressure building? An organization that knows both can allocate resources intelligently. It can deploy Kitsune agents where Chremata detects value signals and withdraw when Baros detects pressure spikes.

The architecture here is local-first. Chremata runs as Python CLI commands with no container orchestration [3]. Baros blends external indicators with internal signals. This isn’t cloud-native idealism. It’s practitioner pragmatism. Teams need these signals available even when external services fail. Local-first design ensures the constraint surface remains visible regardless of infrastructure status.

Adaptive Agents: Identity and Ecology

Kitsune, Kitsune-B, and Pelagos define how agents operate within the temporal and constraint layers established by Augur, Palimpsest, Chremata, and Baros. Kitsune transforms raw traces into curated training datasets for RLHF [2]. Kitsune-B creates static snapshots of coaching interfaces [11]. Pelagos—though less documented in the retrieved snippets—completes the ecological scope, defining the environment within which agents navigate.

Kitsune is the shape-shifter. In Japanese folklore, the kitsune transforms between fox and human forms. In the prototype portfolio, it transforms raw traces—prompts paired with scored responses—into validated, production-ready datasets [2]. Training language models with reinforcement learning from human feedback requires three distinct dataset formats: SFT (single best response), preference pairs (chosen vs rejected), and prompt-only sets for rollout generation. Kitsune handles all three, with quality gates enforcing validation before datasets can be registered [13].

The architecture is phased. Phase 0 runs entirely locally with Python CLI and Docker ClickHouse. Phases 1-2 integrate with external services: Langfuse Cloud for traces, Fireworks AI for training, Novita API for inference [1]. This hybrid design acknowledges a reality most AI teams face: you need local control for data curation, but external scale for model training. Kitsune doesn’t pretend cloud-native is always better. It uses local-first where control matters, external services where scale matters.

Kitsune-B freezes the live coaching interface into static HTML snapshots [11]. No React runtime. No hydration. No fetch layer. Just HTML plus shared Kintsugi stylesheets. This seems paradoxical—why build a static version of a dynamic tool? The answer is reproducibility. Static snapshots allow teams to share exact coaching states without depending on live API availability. It’s the Palimpsest principle applied to agent interfaces: preserve the layer underneath.

Pelagos completes the ecology. The name means ‘open sea’ in Greek, sharing the nautical theme with Pharos (lighthouse) and Baros (weight/pressure). While the retrieved snippets don’t detail Pelagos’s architecture, its role in the portfolio is clear: it defines the environment. Agents don’t operate in isolation. They operate in ecologies with currents, depths, and pressures. Pelagos makes that ecology visible.

The three together answer: How do agents adapt? Kitsune provides the training data. Kitsune-B provides the coaching interface. Pelagos provides the environment. Agency isn’t a property of the agent alone. It’s a property of the agent-in-environment.

Synthesis: From Experiment to Infrastructure

What happens when these prototypes interact? The portfolio becomes infrastructure. Augur reads Palimpsest to ground predictions in historical context. Chremata influences Kitsune by providing value signals that weight training data priorities. Baros constrains Pelagos by defining the pressure surface agents must navigate. This isn’t hypothetical integration. It’s the architecture teams need when experimentation matures into operational systems.

Consider a concrete workflow. An organization wants to deploy RLHF agents for incident triage. Kitsune curates the training data from historical traces [2]. But which traces matter most? Chremata’s value signals identify which incidents had financial impact [5]. Palimpsest reveals whether similar incidents appeared in prior quarters, allowing the system to distinguish novel patterns from recurring ones [9]. Augur detects whether prediction markets expect similar incidents to increase [4]. Baros measures whether geopolitical pressure might amplify incident severity [10]. Pelagos defines the operational environment where agents will execute [10].

This workflow isn’t seven tools. It’s one cognitive system with seven modules. The modular ontology approach allows teams to adopt incrementally—you can start with Kitsune alone, or Chremata alone. But the architecture rewards integration. Each module’s output becomes another module’s input. The system’s intelligence emerges from the connections, not the components.

Semantic approaches to systems engineering incorporate modular domain ontologies to drive collaborative design capabilities [9]. But ontology without implementation is documentation. These prototypes are implementation. They’re shipped code with concrete interfaces, quality gates, and architecture diagrams. Teams don’t need to understand the ontology theory to use the prototypes. They need to understand what problem each prototype solves.

The next iteration of the portfolio won’t add more prototypes. It will deepen the connections between existing ones. Expect to see Augur-Palimpsest query interfaces that allow temporal reasoning across prediction and history. Expect Chremata-Baros constraint surfaces that agents can query before making decisions. Expect Kitsune-Pelagos ecology definitions that make agent environments explicit and measurable.

This is the transition from experiment to infrastructure. Experiments prove concepts. Infrastructure runs systems. The Deerfield Green Prototypes were built as experiments. They’re maturing into infrastructure. That maturation isn’t about adding features. It’s about strengthening the connections that make the system coherent.

The agent revolution isn’t arriving as a single tool that replaces human judgment. It’s arriving as modular cognition—memory, value, pressure, and agency distributed across interconnected systems. The Deerfield Green Prototypes embody this distribution. They’re not seven solutions. They’re one operating system with seven modules.

This matters for teams building complex systems. You don’t need to adopt all seven prototypes. You need to recognize the cognitive layers they represent. Does your organization have memory? (Palimpsest) Does it measure pressure? (Baros) Does it quantify value? (Chremata) Does it define agency? (Kitsune, Pelagos) Does it predict? (Augur) If any layer is missing, your system has a cognitive gap.

The architecture emerged from failure. Teams optimized prediction while history rotted. Agents trained without constraint surfaces. Value signals existed separately from pressure measurements. The prototypes fix this by making the connections explicit. Augur reads Palimpsest. Chremata constrains Kitsune. Baros defines Pelagos.

Build the connection first. The modules will follow. Start with two prototypes that address adjacent cognitive layers. Connect their interfaces. Let data flow between them. Then add the third. The system emerges from the connections, not the components. This is how experimentation becomes infrastructure. This is how modular ontology becomes operational reality.

References

Deerfield Green