Observe Current State
The reasoning model interprets the initial market state and context.
An Adaptive Compound Financial Reasoning Model trained to retrieve, integrate, and evaluate specialist quantitative intelligence under changing market conditions.
TFL Bloodhound Model 1 is a proprietary financial AI system designed to reason over changing market conditions using quantitative models, specialized financial representations, long context memory, and reinforcement learning from realized market outcomes.
Bloodhound is not a general-purpose language model repurposed to analyze charts. It is a compound reasoning architecture in which language reasoning operates over structured evidence generated by specialist financial models.
This separation prevents the language model from being responsible for calculations and market representations that require specialized numerical processing.
The language model does not independently infer a market regime from raw numbers, calculate volatility relationships, or reconstruct historical distributions through natural-language reasoning. Instead, it requests and receives structured outputs from dedicated TFL models trained for those tasks.
Bloodhound Model 1 reasons from quantitative evidence rather than attempting to replace quantitative modeling with language.
The reasoning layer is built on a fine-tuned variant of Gemma 4 26B A4B — a sparse Mixture-of-Experts foundation model with a context capacity of up to 262,144 tokens, retrained by The Finance Lab for financial market reasoning.
Gemma provides the underlying language, long-context, tool-use, and reasoning capabilities. However, the foundation model alone does not constitute Bloodhound Model 1. The Bloodhound reasoning model is created through additional financial training, tool orchestration, memory integration, and reinforcement-learning processes developed by The Finance Lab.
Its principal transformation is a domain-specific reasoning policy trained using Group Relative Policy Optimization — GRPO. GRPO is used to train the model not only to produce an answer, but to follow a financial evidence-retrieval and hypothesis-testing process.
Interpret the current analytical objective.
Identify which financial evidence is missing.
Request the appropriate specialist TFL model or analytical tool.
Evaluate the retrieved numerical and contextual information.
Construct one or more hypotheses about the current market condition.
Generate a probabilistic expectation for the next movement.
Compare the expectation with the realized market outcome.
Learn which evidence and reasoning decisions contributed most to a calibrated forecast.
The resulting model is not trained merely to describe a market situation. It is trained to determine which information must be retrieved before a market expectation can be formed.
Bloodhound uses an iterative reasoning process in which the model can query its own analytical environment. When the reasoning model encounters an unresolved question, it can request additional evidence from another component of the TFL architecture.
The reasoning model interprets the initial market state and context.
e.g., What is the current volatility environment? Is the market trending or mean-reverting? Which user-defined playbook rules apply?
Dynamically retrieve context from TFL models, tools, and multidimensional memory rather than relying on static prompting.
Reason over quantitative evidence to synthesize competing hypotheses.
Formulate a probabilistic scenario with associated confidence boundaries.
Generate a GRPO reinforcement signal. The model learns which signals were truly relevant over time.
The reasoning model therefore learns when to retrieve information, which source to query, and how much weight that evidence should receive in the final assessment.
Bloodhound Model 1 operates through a multi-model architecture composed of specialist TFL models. Each model is optimized for a particular class of financial computation and returns a structured representation to the reasoning layer.
Classifies the structural state of the market environment.
Price structure, trend persistence, and directional alignment.
Current implied and realized volatility configuration.
Identifies structurally comparable past market conditions.
Correlations and relationships across different asset classes.
Forward-return distributions and market-sequence expectations.
User constraints, portfolio exposure, and risk boundaries.
Multi-horizon momentum configuration and strength.
The architecture behaves functionally like a multi-headed financial intelligence system. Each specialized component provides a different analytical perspective, while the reasoning layer combines the outputs into a coherent interpretation.
This should not be confused with standard multi-head attention inside a transformer. Bloodhound's architecture is a multi-model expert system in which independently specialized quantitative models supply evidence to a central reasoning policy.
Outputs from specialist models are converted into structured context objects before being presented to the reasoning model. This transformation acts as a bridge between quantitative computation and language reasoning.
Instead of asking a language model to derive a complex regime directly from raw time-series values, Bloodhound supplies a machine-generated regime representation produced by a model trained for that purpose.
Instead of asking the model to estimate an outcome distribution through text reasoning, Bloodhound supplies the statistical distribution generated from historical and model-based analysis.
The reasoning model can then focus on higher-order questions: Which evidence is most relevant? Are the analytical components consistent? What information contradicts the leading hypothesis? Which scenario is best supported? What would invalidate the expectation? How should the result be interpreted within the user's playbook and portfolio?
Bloodhound's reasoning policy is trained using realized market behavior as delayed feedback. For each training episode, the system records:
The expected movement is then compared with the realized outcome. The reinforcement process can reward more than simple directional accuracy. It evaluates probability calibration, magnitude error, distributional accuracy, risk-adjusted utility, maximum adverse movement, and relevance of retrieved evidence.
This allows the model to learn not only whether its expectation was confirmed, but also which information was genuinely useful in forming that expectation. A market variable that was highly relevant in one period may become less informative under a different environment.
Bloodhound is therefore designed for a non-stationary domain in which the importance of information changes over time. Historical evidence is treated as context, while recent market feedback helps determine how much relevance that evidence should receive now.
TFL Bloodhound includes a memory architecture called MeMo. MeMo is not limited to storing text documents or retrieving previous conversations. It maintains a multidimensional representation of financial experience.
A memory episode can include the analyzed instrument, time, environment, quantitative model outputs, hypotheses, playbook rules, expected movements, and realized outcomes. These dimensions create a structured memory environment that can be conceptually represented as a financial tensor.
MeMo does not treat all stored experiences as permanently equal. The relevance of a memory can change as the market changes. A historical episode may remain structurally similar while becoming less predictive because of changing liquidity, volatility behavior, or cross-asset correlations.
The defining characteristic of TFL Bloodhound Model 1 is that it does not reason about markets in isolation. It reasons through a network of specialist financial models.
Adaptive compound financial reasoning model trained to retrieve, integrate, and evaluate specialist quantitative intelligence under changing market conditions.