Bloodhound
Pillar 04

Metacognitive Pipeline

Six stages replace the human researcher sitting in front of a computer for hours. The system decomposes the question, allocates resources, generates candidates, evaluates quality, verifies correctness, and orchestrates refinement—autonomously.

1

Decompose

The research question is decomposed into atomic sub-questions, each targeting a specific information need from a specific modality.

Q → {q₁, q₂, ..., qₖ} where each qᵢ maps to a single slice statement
2

Allocate

Computational resources are allocated across sub-questions based on expected information yield—more resources for higher-yield extractions.

GQIC allocation: resource ∝ I(qᵢ; A_Q) / cost(qᵢ)
3

Generate

Multiple candidate extractions are generated for each sub-question using Determinantal Point Processes for diversity.

DPP kernel K balances quality (diagonal) with diversity (off-diagonal)
4

Evaluate

Each candidate is evaluated across five quality dimensions: factual accuracy, logical coherence, domain consistency, completeness, and uncertainty calibration.

Quality vector q ∈ [0,1]⁵ with weighted aggregation
5

Verify

Domain-expert consensus, adversarial evaluation, and formal verification through proof assistants ensure conclusions are correct.

Multi-expert: consensus ∧ adversarial ∧ formal (Lean 4/Coq)
6

Orchestrate

The orchestrator decides: accept, refine, redirect, or escalate. Refinement loops continue until the convergence criterion is met.

Action = argmax P(converge | state, action) subject to budget

Convergence: Gas → Liquid → Crystal

Analysis temperature decreases through refinement iterations. The system transitions from gaseous (high uncertainty, scattered fragments) through liquid (clustering, cross-links forming) to crystal (converged, validated answer).

Phase Transitions

Gas Phase T > 0.5

Understanding fragments are dispersed, uncorrelated. Each source extraction produces an isolated signature with no cross-modal connections. High entropy, low confidence.

Liquid Phase 0.2 < T < 0.5

Fragments begin clustering through composition. Cross-modal links emerge (gene-protein connections, tissue context). Semantic structure forms but has not solidified.

Crystal Phase T < 0.2

The answer has crystallized. All fragments are integrated, cross-validated, and formally verified. The trajectory is complete, reproducible, and provably correct.

CONVERGENCE THEOREM

Under quality-monotonic refinement and sufficient candidate diversity, the analysis temperature T(n) decreases exponentially: T(n) ≤ T₀ · e^(-λn). Convergence to crystal phase is guaranteed in finite iterations.