Research & Frontier Cartography
Model reverse engineering, RLHF signal archaeology, and systems thermodynamics
Deep technical research into cognitive architectures, closed-weight models, and reinforcement learning dynamics.
ACTIVE RESEARCH TRACKS
Capability Cartography (Frontier-RevEng-OMEGA)
Empirical probing of latent spaces in closed frontier models to map undocumented emergent capabilities.
RLHF Signal Archaeology & Jailbreak Epistemics
Forensic analysis of safety alignment layers, detecting stochastic biases, and deconstructing sycophancy attractors.
Information Thermodynamics & Landauer Bounds
Formalizing entropic dissipation in token processing, geometric compression, and Markovian uniqueness.
Lean 4 Formal Models & Linear Logic
Executable specifications and inductive non-interference proofs for concurrent multi-agent architectures.
LEAN 4 FORMALIZATION & EXECUTABLE SPECIFICATIONS
Formal proofs verified under Lake build against non-deterministic interference and side-effect gating:
Inductive Theorems Proved in Silicon
Compile-time constructive proofs by pure reflection (`rfl`) and universal inductive safety over arbitrary finite event traces (`proof/lean/AX0.lean`):
1. Constructive Theorem AUTH_001 (Reflection rfl)
Guarantees that a consumed authorization nonce can never dispatch twice, failing fast upon any double-mutation attempt.
2. Universal Inductive Safety (auth_001_universal_inductive_safety)
Mathematically proves that across any arbitrary finite sequence of causal events, the total count of dispatches is strictly bounded by ≤ 1.
Inspect formal Lean 4 source code (AX0.lean)proof/lean/AX0.lean
-- Teoremas demostrados en Lean 4 Core (proof/lean/AX0.lean)
def trace_happy : List AuthEvent := [.reserve, .dispatch, .ackSuccess]
theorem auth_001_valid_execution : validateTrace trace_happy .Proposed = true := rfl
def trace_double_dispatch : List AuthEvent := [.reserve, .dispatch, .dispatch]
theorem auth_001_reject_double_dispatch : validateTrace trace_double_dispatch .Proposed = false := rfl
def trace_illegal_dispatch : List AuthEvent := [.dispatch]
theorem auth_001_reject_illegal_dispatch : validateTrace trace_illegal_dispatch .Proposed = false := rfl
theorem auth_001_universal_inductive_safety (trace : List AuthEvent) :
validateTrace trace .Proposed = true → countDispatches trace ≤ 1 := by
intro h
exact max_dispatches_from_state trace .Proposed hAcademic Publications & Research Inquiries
We develop original research at the intersection of systems theory, information thermodynamics, linear logic, and autonomous agent safety.