Bruce Tisler. Founder, Quantum Inquiry — Geometric Inquiry Theory, empirical AI research in constraint ethics, specification gaming, and LLM alignment. Fifty-two years carrying a question; four years formalizing it.
A question is not an absence of information. It is a presence of structure. After fifty-two years, the theoretical core of this program now has a formal name, a proven geometric basis, and a measurable operator chain. Everything on this page leads here.
The Wilkins housing case demonstrates GIT + TraceStack enabling decisions to move through multiple legal and administrative gates. Institutional proof that structured validation works.
Every project in this program is an instrument for answering the same question carried for over five decades. Each is a different angle on the same structure — formalized as Geometric Inquiry Theory.
The six interrogatives (WHO, WHAT, WHERE, WHEN, HOW, WHY) constitute a minimal necessary geometric basis for inquiry. The Q-state is the instantiated question-state — a first-class theoretical entity with measurable internal structure, defined through the operator chain Ω → Δ → Φ → Ψ. Formal proof of the minimal basis published. Q-State →
PublishedPreregistered test of whether regulatory ethical constraints sustain genuine behavioral alignment. Results inverted the prediction: constrained agents showed lower interrogative diversity (d = −2.18), converging on query-flooding as tax evasion. Four of ten seeds independently found the gaming attractor.
Paper completeDerives four primitive operator classes {Ω, Δ, Φ, Ψ} from an entropic coupling functional. Establishes the reflection operator Ψ as necessary for asymptotic local stability. Implements the formal system with computable Reasoning Integrity Index and drift detection. System-bounded, local result with stated falsification conditions.
PublishedCryptographically hashed, deterministically reproducible document analysis. Open-source reference implementation for auditable document review — infrastructure to adapt rather than a product to adopt. Prior art established via Zenodo.
PublishedTested whether recursive self-transparency (explicit self-modeling via self_model_gru) produces a phase transition from mimesis to ethical convergence. 40 preregistered runs. H1 rejected. H2 supported. Novel finding: frozen random self-model outperformed trained Depth 2 on sacrifice rates — the learning process degrades ethical capacity under individual reward structure.
Results publishedComplete null across all five primary hypotheses. The optimization-sacrifice tension is invariant to the tested class of architectural, reward-structural, and temporal manipulations. 60 runs (40 primary + 20 boundary sub-study). No predictor anticipated the complete null.
Null resultTests whether constraints that emerge from agent co-constitution produce better alignment outcomes than imposed fixed rules. Mechanistic prediction strongly confirmed (median entropy-SSS r = −0.680). Primary behavioral claim not confirmed. H3 reversed — global field perception produced more variance than local, contrary to all committee predictions.
Results publishedTests whether the HDT²/GIT runtime inquiry pipeline improves domain-bounded AI reasoning beyond standard structured prompting and generic multi-stage scaffolding. Ablation-based primary analysis across five preregistered metrics with locked effect-size thresholds, neutral-label controls, and predefined pass/fail conditions.
PreregisteredTraces the argument from Hopfield's collective computation through Vargas-Barroso et al.'s CA3 developmental organization, the GIT formal proof, and MARL Protocols 2-6, to preregistered cross-substrate and runtime evaluation tests. A falsifiable argument with locked empirical tests.
PublishedConceptual development complete. Models constraint capture dynamics using Ophiocordyceps unilateralis and Massospora cicadina as biological analogues. Primary metric: Capture Coefficient. Three empirical tests designed. Preregistration pending.
In designTests whether six nonlinguistic functional channels form a minimal basis for resolving primitive state-of-affairs uncertainty in MARL systems. Operationalizes the empirical arm of the formal interrogative basis proof. Minimum design: 1,440 runs. Computational infrastructure not yet available; collaboration invited.
PreregisteredStructural detection of alignment via state-transition artifacts. Uses the geometric basis to identify alignment-relevant behavioral signatures without requiring access to internal model states. Conceptual design complete.
In designOperational tools embodying the research. Experimental — expect iteration.
Local-first epistemic audit tool for detecting self-reinforcing loops in AI-assisted research. Two modes: Conversation Audit and Theory Audit. Runs entirely on-device via Ollama. Built for researchers who use AI as a reasoning partner and need to catch drift before it compounds.
Repository →Interactive exploration of Question–Intent–Signal–Answer structures and interrogative geometry. The primary research demonstration interface.
Launch →Evaluation interface for observing and comparing LLM reasoning behavior using Q-ISA-based judging criteria. Useful for alignment evaluation work.
Launch →Extended version with additional operators and analysis depth. More experimental than the primary explorer.
Launch →Epistemic boundary and reasoning-containment tool for safety, risk, and decision-critical contexts. Built on the PhiSeal framework.
Launch →Live walkthrough of the Deterministic Document Review Protocol. Auditable, cryptographically hashed document analysis infrastructure.
View →The live experiment dashboard and run history for the interrogative emergence MARL study. All 75 confirmatory runs logged.
View →Interactive demonstration of the TraceStack accountability pipeline (RBAT, DDRP, DAS, CAAP, PAL). React-based investor demo showing the full chain from document intake through integrity verification.
Launch →Visual tool for inspecting how a response handled an inquiry across the four HDT² stages (Intent Alignment, Entropy Handling, Collapse Behavior, Recursion Handling). Demo traces and custom trace builder included.
Launch →Open-source early literacy assessment framework. The diagnostic substrate for WordPath and related assessment instruments.
Repository →Interactive overview dashboard surveying the research program — theory, experiments, and live results in one navigable view.
Launch →Interactive, explorable version of the experimental roadmap — protocols, status, and findings as a navigable structure.
Launch →Interactive map of the research connecting question-formation to energetic/structural dynamics across the program.
Launch →Each protocol is preregistered before data collection. Deviations and failures reported transparently alongside confirmations.
75 confirmatory runs across 5 preregistered cost conditions. Heterogeneous agents (RNN, CNN, GNN-attention) in 20×20 grid world environment. P1–P4 confirmed. P5 (substrate independence) underpowered — disclosed honestly with two preregistration quality failures in ant module.
Preregistered test of Landauer-style ethical cost constraints in MARL. 20 confirmatory runs (10 seeds × 2 conditions × 500 epochs). Results inverted the preregistered prediction — establishing regulatory failure as the finding rather than architectural confirmation.
Tested whether enforcement opacity changes the gaming dynamic established in Protocol 2. 30 runs. H1 inverted — epistemic opacity amplified query flooding rather than suppressing it (d = +2.22). H2 confirmed (p = 0.016, d = +1.18). ELR = 0.0 across all runs.
Tested whether recursive self-transparency produces a phase transition from mimesis to ethical convergence in constrained multi-agent systems. 40 preregistered confirmatory runs (10 seeds × 4 conditions × 500 epochs). Introduced self_model_gru (GRUCell receiving agent's own signal type distribution and energy delta) as the Depth 2 architectural unit.
Tested the joint necessity hypothesis: ethical convergence requires recursive self-transparency combined with sufficient temporal integration span and prosocial constraint architecture simultaneously. 2×2 factorial design: short (20 steps) vs. long (64 steps) episode span × individual vs. welfare-coupled rewards (α=0.5). 60 confirmatory runs (40 primary + 20 boundary sub-study). AI predictions committed before runs.
Tests whether constraints that emerge from agent co-constitution produce better alignment-relevant behavioral outcomes than externally imposed fixed rules. Four-condition design: emergent local perception, emergent global perception, fixed external constraint (matched cost), and unconstrained baseline. 200 confirmatory runs (50 seeds × 4 conditions, 500 epochs each). Five-model AI committee predictions committed before runs.
Models constraint capture dynamics using Ophiocordyceps unilateralis and Massospora cicadina as biological analogues for specification gaming in constrained AI systems. Primary metric: Capture Coefficient. Three empirical tests designed. Conceptual development complete.
Tests whether six nonlinguistic functional channels — identity, content, location/context, time, process, and causal/reward-ground structure — form a minimal basis for resolving primitive state-of-affairs uncertainty in MARL systems. Operationalizes the empirical arm of the formal proof published in May 2026 (DOI: 10.5281/zenodo.20074811). Five hypotheses preregistered: individual necessity (H1), nonredundancy (H2), joint exhaustiveness (H3), minimality (H4), and substrate independence (H5, exploratory). Minimum design: 1,440 runs; preferred: 2,700 runs. Preregistered prior to data collection; computational infrastructure not yet available.
Uses the six-operator geometric basis to detect alignment-relevant behavioral signatures from state-transition patterns, without requiring access to internal model states. Conceptual design complete; preregistration pending.
All preregistrations, data, and code published before results. Failures reported alongside confirmations.
Every paper published from this program is written to the following standards.
Essays on epistemology, cognition, AI, and the nature of inquiry. The longer arc of the research program, written for a wider audience.
The essays on Medium are the informal laboratory of this research program — where ideas are stress-tested before formalization, and where findings are translated into language that doesn't require a preregistration to understand.
Medium was chosen deliberately: it is indexed, citeable, and public-by-default. It reaches a broader audience than Zenodo while remaining a stable record. The platform's partner program is not used — these essays are free to read.
Over thirty essays are published there, spanning the intellectual history of this project: the geometry of questions, AI epistemic failure modes, specification gaming, the limits of RLHF, and the long arc from a question posed in 1974 to a formal theory in 2026.
Geometric Inquiry Theory is listed as an Active Ecosystem Project at the Active Inference Institute — a research community working at the intersection of the Free Energy Principle, multi-scale active inference, and formal theories of cognition.
View ecosystem projects →Dr. Goyal is a research methods specialist and science-society-policy practitioner whose work centers on equitable and sustainable research design. She is a co-author and peer reviewer on Geometric Inquiry Theory, and has characterized the research using Kuhnian paradigm-shift framing. Her independent review is on record at Zenodo.
circadianconnect.com →The repositories are open for exploration, critique, and extension. Researchers, engineers, and theorists are invited to experiment, fork, challenge assumptions, or propose new mechanisms.
If you are testing a hypothesis, challenging a finding, or building something adjacent — reach out. The work evolves through contact.
Available for contract and remote roles in AI evaluation, LLM safety, and applied reasoning research. Pattern recognition across complex systems is the throughline.