Research Note · Theseus Technologies · July 2026
A typed-graph architecture for coherent, auditable reasoning in discretionary investment agents
Download PDFWe describe an architecture for grounding an autonomous investment agent’s decisions in an explicit, internally consistent set of beliefs. The core idea is to treat retrieval as a physical process: a query injects activation into a typed graph of principles, that activation spreads along weighted edges and decays with distance, and edges marked as contradictory carry negative conductance, so that incoherent branches of reasoning cancel themselves out before they ever reach the model. What survives is a small, ordered, self-explaining chain of beliefs, which is then grounded in market data and passed to a forecasting simulation. We situate this design against its nearest intellectual relative, Thagard’s connectionist theory of explanatory coherence, against which it is a novel application rather than a first invention; we compare it to two alternative architectures (an unconstrained neighborhood-retrieval variant and a self-play reinforcement-learning variant), and we are explicit about which parts of the design rest on solid evidence, which rest on recent and largely unreplicated results, and which rest on no evidence at all. We close with the open research questions that determine whether this approach is viable.
The premise of this work is a claim about agent design, not about markets: an agent whose reasoning is drawn from an internally consistent set of beliefs will outperform one whose reasoning is not, holding the quality of the underlying beliefs constant. A large language model given an unstructured pile of context cannot easily tell which parts of that context agree with each other; if the agent’s knowledge base contains both a value-investing principle and a momentum principle that contradict it in a given situation, dumping both into a prompt does not surface the tension, it buries it. The architecture below is an attempt to make coherence a computed property of retrieval itself, rather than a separate check bolted on afterward.
The organizing metaphor is that the context window is RAM; the belief graph is disk. The graph holds everything the firm believes; each retrieval decides what small, coherent slice gets loaded into working memory for one decision. This is not a novel metaphor in retrieval-augmented generation generally, but it motivates a specific design choice: retrieval should be a traversal with its own dynamics (activation, decay, cancellation), not a similarity lookup followed by concatenation.
Claim: a coherent chain of principles, retrieved by a process that lets contradictory beliefs cancel each other, grounded in real data, and projected forward by a simulation, produces better-justified investment decisions than either (a) an LLM reasoning over an undifferentiated context window, or (b) a model with no explicit belief structure at all.
This claim is plausible and is consistent with a body of retrieval-augmented-generation and cognitive-science literature discussed in §2, but it has not been tested end to end for this domain, and the weakest link, whether a short natural-language query can be reliably decomposed into the graph coordinates the system needs (§9), is unproven. We flag evidentiary strength throughout rather than presenting the design as settled.
The architecture combines ideas from three largely separate literatures that, as far as we are aware, have not previously been put together this way for financial reasoning. Being explicit about lineage matters here: it lets us borrow known failure modes instead of rediscovering them by trial and error.
The mechanism in §3 (inject activation at seed nodes, let it spread along edges, decay with distance, gate on a firing threshold) is a direct descendant of Collins and Loftus’s spreading-activation theory of semantic processing[1], developed to explain human semantic priming. That model was never intended as a retrieval algorithm for machines; its adaptation into knowledge-graph RAG systems is recent (§2.3).
The single most important design feature of this architecture, that a contradiction edge carries negative current, so surviving activation IS a coherence score, is not new. It is, essentially, Paul Thagard’s ECHO model of explanatory coherence[2][3]: a connectionist constraint-satisfaction network in which propositions are units, mutually supporting propositions get excitatory links, contradictory propositions get inhibitory links, and the network is allowed to settle; the resulting stable activation pattern is Thagard’s formal definition of which explanation is most coherent. McClelland’s earlier interactive-activation-and-competition models[4] use the same excitatory/inhibitory settling dynamic for perceptual constraint satisfaction (the “Jets and Sharks” model).
The most direct precedent for using a typed knowledge graph as an LLM’s retrieval substrate, with activation spreading over it, is SA-RAG[5], which reports large correctness gains over naive vector RAG on an automatically-constructed graph (itself evidence that graph construction need not be fully hand-curated), consistent with recent surveys of LLM-driven knowledge-graph construction[8]. Two additional data points are commonly cited in favor of graph-grounded retrieval for financial tasks: a Neo4j benchmark showing knowledge-graph retrieval sharply outperforming vector search on a financial-analysis task, and a GraphRAG evaluation on FinanceBench reporting fewer hallucinations and lower token cost than conventional RAG[6].
0 QUERY (a situation / question arrives)
|
v
1 DIMENSION RESOLUTION [stateless inference pass]
- rank which domains of belief the query touches
- assign an energy/resistance coefficient to each
|
v
2 SEED [1..N entry points]
- inject activation at matched principles + the
querying agent's own worldview anchors
|
v
3 SPREADING ACTIVATION [over the typed principle graph]
- current flows along typed edges, decays per hop
- firing threshold acts as a noise gate
- CONTRADICTS edges carry NEGATIVE current
|
v
4 PRINCIPLE PATHWAY <-- the retrieved context window
- the lit, connected chain of principles
- pre-ordered and self-explaining: the trace IS the justification
|
v
5 DATA OVERLAY [empirical grounding]
- quantitative extraction pulls relevant market data
- layers real numbers onto the principle pathway
|
v
6 FORECASTING SIMULATION [agent-based / Monte Carlo]
- projects the grounded situation over an investment horizon
|
v
7 DECISION
- price entry / exit, time entry / exitStages 1–4 are the reasoning half: what does the firm believe about this situation, and does that belief survive contact with itself? Stages 5–7 are the empirical/predictive half: what do the numbers say, and what happens next? This separation matters because the two halves have very different risk profiles: the reasoning half is a retrieval and consistency problem with reasonable precedent (§2); the predictive half is a forecasting problem in a genuinely hard, arguably unsolved research area (§6).
An important consequence of the design worth stating explicitly: because CONTRADICTS edges remove current rather than merely being flagged, coherence is not a post-hoc filter applied to a retrieved set: it is computed during retrieval, as the same operation that decides what gets retrieved. This is the design’s central claim to novelty over standard RAG, and it is inherited directly from the ECHO tradition described in §2.2.
The retrieved pathway in Stage 4 could in principle live on either of two substrates: a typed graph, where every edge has a labeled type and a weight, or a weighted vector space, where traversal cost is a function of which embedding dimensions a query crosses. The vector-space version is simpler to build. We nonetheless argue for the typed graph, for two reasons that are structural rather than a matter of taste:
The dimension-weighted-traversal idea from the vector-space alternative is not wasted, however: it maps cleanly onto edge-type weights in the graph (conduct more freely along causal edges, resist along bookkeeping edges), so the tuning concept survives even though the substrate does not need to change to get it.
Restated precisely, because it is the feature most likely to be quietly lost as the system grows in scope: contradiction edges carry negative conductance; mutually supporting beliefs add current; contradictory branches cancel. The surviving activation of a branch is its coherence score, with no separate coherence checker required. This is, again, Thagard’s constraint-satisfaction formulation of coherence (§2.2), not a new idea, but it has apparently not been applied to belief-grounded financial reasoning before, which is where the actual contribution of this architecture lies.
One parameter remains genuinely open and should be settled empirically rather than by preference: whether contradiction should be soft (dampens activation continuously, producing a rankable worldview-coherence index per thesis) or hard (cancels outright, pass/fail). Thagard’s own ECHO implementations use continuous (soft) activation update rules, which is mild evidence in favor of soft polarity here, but the financial-domain behavior of either choice is untested.
This section corrects a category error that is easy to make when “forecasting” methods are discussed loosely. Two methods are commonly proposed for the forward-looking layer: simulated annealing and iterative inference loops. Neither is a forecasting method.
Recommendation: the predictive layer should be an agent-based or Monte Carlo simulation of the situation, with annealing or inference loops used only for parameter calibration inside that simulation, never as the forecasting mechanism itself. A single, reusable simulation engine, re-parameterized per query, is preferable to a bespoke model per prompt.
The path-tracking pipeline in §3 is one point in a broader design space, not the only viable design. Two alternatives are worth characterizing rigorously rather than dismissing.
Instead of a lit, decayed traversal, this variant resolves a query to a single strong seed vector, fixes a cluster radius and a size cap, and retrieves every node in that neighborhood, collapsing Stages 2–3 of §3 into one step, with no per-edge decay, threshold, or activation tuning.
| Property | Path-tracking (§3) | Cluster retrieval |
|---|---|---|
| Traversal machinery | Central, heavily tuned | Not needed: no traversal to tune |
| Burden on query resolution | Ranks dimensions | Must nail a single start point and radius; the entire result rides on it |
| Dependence on embedding quality | Helpful | Make-or-break: a mediocre vector space returns a bad neighborhood |
| Output shape | Ordered, self-explaining trace | Unordered cloud of principles |
| Coherence arithmetic (§5) | Native | Lost by default; would need to be reconstructed as a separate module |
Cluster retrieval is not simply a worse version of path-tracking; it is a real alternative that wins under two specific conditions worth naming precisely, because they are testable claims rather than matters of preference: (1) the embedding space turns out to be excellent but the traversal algorithms of §3–§5 cannot be made to work reliably; or (2) a sufficiently large context window lets a model use an unordered cloud of principles about as well as a curated trace, at acceptable token cost. Absent evidence for either, the standing objection to cluster retrieval is the same one that motivates staged retrieval generally: loading an undifferentiated block of context up front tends to degrade a model’s reasoning relative to a curated, ordered trace, and this variant forfeits the auditability and coherence-as-arithmetic properties that are this design’s main claims to distinctiveness.
The second alternative discards explicit belief structure entirely: give the system data and market access with minimal constraints, have it trade with simulated capital against real or simulated markets, repeat at scale, and keep what wins.
The approach also produces an interpretability gap: an emergent policy may not yield the auditable trace that is this project’s main product differentiator (§4), and compute cost is unbounded without an explicit cap (§8). It is best framed as a fenced, budget-capped research program run in parallel to the core architecture, not as a competing production path, evaluated on whether it discovers anything that a hand-built ontology would not have.
Cost is not incidental to this comparison; it is a structural property that differs by architecture in ways worth stating precisely, at order-of-magnitude precision only.
| Driver | Path-tracking | Cluster | Self-play |
|---|---|---|---|
| Graph storage / traversal | Low (CPU, cheap) | Low | N/A |
| LLM tokens per query | Low–medium (small curated trace) | Medium–high (large context cloud) | N/A at inference time |
| Simulation compute | Medium (CPU-bound, bursty) | Medium | N/A |
| Training / self-play GPU | N/A | N/A | Very high, continuous |
| Cost predictability | Predictable | Scales with context size | Unbounded without an explicit cap |
The qualitative read-through: path-tracking is the cheapest and most predictable architecture at scale, because small curated traces produce small token bills. Cluster retrieval looks cheaper to build but its token cost grows with context size and can quietly exceed path-tracking’s cost once usage scales. Self-play sits in a different cost regime entirely and should be evaluated as a bet with an explicit stop-loss on spend, not as a line item comparable to the other two, consistent with the critique in §7.2 that its underlying premise (a sufficiently stationary market to learn against) is itself unproven.
The following are unresolved technical questions that determine whether this architecture is viable, stated as research questions rather than as items for a status meeting.
The contribution of this design is not the invention of spreading activation or of coherence-as-constraint- satisfaction (both have decades of precedent in cognitive science, §2), but the specific application of that machinery to grounding an investment agent’s reasoning in an auditable, internally consistent belief graph, with a clean separation between the reasoning half (Stages 1–4) and the empirical/predictive half (Stages 5–7). The reasoning half rests on a reasonably credible, if still thin and recent, evidence base (§2.3) and a well-established cognitive-science precedent (§2.2). The predictive half rests on a genuinely open research problem (§6) that no amount of engineering discipline resolves on its own. The two alternative architectures (§7) are not weaker versions of the same idea; they occupy different, definable points in a real design space, one of which (self-play) carries a structural objection (market nonstationarity) serious enough that it should be treated as a bounded research bet rather than a production candidate. The single highest-leverage next question is not architectural but empirical: whether a query can actually be decomposed into the structure Stage 1 assumes it can (§9, item 1). Everything downstream depends on that being true.