ZTGI-Pro v3: Tek-Throne Risk–Stability Law for FPS-Based AI Systems

This white paper presents ZTGI-Pro v3, the latest formulation of the
Zorunlu Tekil Gözlem İlkesi (ZTGI – “Compulsory Singular Observation Principle”)
as a risk–stability core for FPS-based AI systems.

ZTGI builds on a simple but strong axiom: within any causal closed region (CCR)
there can be only one permanent First-Person Perspective (FPS). When internal
jitter (σ) and goal–action dissonance (ε) grow beyond critical levels, the system
is driven toward a “Tek-Throne” collapse event Ω = 1 and falls back to a single
dominant stream.

v3 introduces a fully formalised hazard and collapse model:

• A log-barrier hazard action I(σ, ε, ρ, χ) with gate Q = e^(−I) and hazard energy H,
combining jitter, dissonance, robustness (ρ) and coherence (χ) into one scalar;

• A free-energy block F and energy signal E, feeding a contextual threshold Θ(ψ)
and dynamic hazard threshold H*(ψ);

• A dual-horizon EMA hazard memory (Hs, Hl, Ĥ) and a risk score r = Ĥ − H* driving
a collapse rule Ω = 1{ Q < Θ(ψ) or Ĥ > H* or E < E_floor }.

Phase-3 telemetry on a running ZTGI-driven agent shows that this single risk score
r predicts collapse with AUC ≈ 0.98, Brier ≈ 0.046 and calibration error (ECE)
in the 0.06–0.09 range. Risk histograms, σ–ε stability surfaces, ROC curves and
reliability diagrams match the theoretical iso-hazard geometry.

The paper also describes the ZTGI-AC / Tek-Throne architecture, where a ZTGI-Shield
module monitors inner stability and an LLM controller adjusts temperature, top-p,
max tokens and answer style according to the internal risk state before responding.
Single-Throne Orchestration (STO) generalises this to multi-agent / multi-LLM
settings by enforcing a single executive FPS with advisory modules.

Cosmological and quantum-inspired extensions (Planck / Riemann links) from
earlier drafts are deliberately excluded from the v3 core and treated as speculative
hypotheses for separate work.

ZTGI-Pro v3 thus positions ZTGI as a concrete, testable and engineering-ready
risk–stability law for pseudo-conscious AI control, rather than a purely
philosophical theory of consciousness.