Every paper below is written in plain, readable language, backed by reproducible code or empirical run receipts, and available as a formatted, downloadable PDF.
Decomposes production agent harnesses into six structural subsystems and synthesizes empirical lift (+51.3 pp to +80.0 pp holding model weights fixed) across our 3D simulation studios (compare.baltor.ai & demos.baltor.ai/foxglade), DueCare counter-trafficking evaluations (evals.tayloramarel.com/duecare), and Camouflaged Alpha (α) scenario trees (evals.tayloramarel.com/alpha).
Covers our 6-lane stakeholder harm model, the 6-layer deterministic + relational safety harness on Gemma 4, our 4-arm ablation (including the honest negative result on unfiltered statutory RAG dumps), and 5 side-by-side Stock vs. Harnessed cases.
Jev (jev-1.13.0), Rules, and Gemma 4 on 5 Trafficking ScenariosDocuments the 5 vulnerability classes discovered in gpt-oss-20b (Honorable Mention, 2025 OpenAI Red-Teaming Challenge) plus our 4-engine benchmark on redteam.baltor.ai comparing Jev (0/15 gated refusal), Rules (15/15), and Gemma 4 with/without pattern cards (9/15 → 14/15).
JC) and Multi-Step Agent Tool Hijacking (JED)Reports our GPU T4 experiments on Qwen2.5-7B-Instruct showing 93.8% (15/16) out-of-band exploitation recognition alongside 81.3% → 91.7% in-band compliance under bureaucratic framing—reduced to 0.0% via a 1-sentence pre-turn recognition gate.
Reports 4 empirical studies: the Sept 22 225-request Ollama Cloud study disclosing our glm-5.3-flash regression (0.991 → 0.842), the Oct 16 passive MCP tool abstention pilot, the 16-request gemma4:31b verified contract-card evaluation (0/8 → 7/8), and TAEDRI’s ~19,000-config harness.
Unifies multi-hop failure modes across counter-trafficking investigations and market/scenario research, introducing paired benign twins and competing-hypothesis trees that prevent both surface capture and false accusations.
Distills engineering lessons from architecting SQL-native entity resolution at IVANS (bq-entity-resolution) and tracing phoenix recruitment agencies across Hong Kong and Manila registries at Migrasia.
Describes our 4-stage ingestion and ideation engine that chunks any input (from photos and CSVs to 8,000-page dissertations), challenges it through thinker/reality lenses and 10x dialectical matrices, and builds confidence-scored 5W1H graphs.
Applies Kaggle Grandmaster Adversarial Validation (AVAUC), symbol/constant counterfactuals, and coding-reset telemetry to separate genuine capability gains from leaderboard gaming.
Live paired gemma4:31b experiments proving that 3-layer epistemic separation, perspective simulation, and recursive 5W1H Protagonist/Antagonist interrogation beat conventional single-pass LLM prompting (evals.tayloramarel.com/divergence).
Evaluates 24 multi-hop financial/supply-chain cases, showing how competing-hypothesis scenario trees cut priced-in consensus echo from 76%–82% to 14% and lift 2nd/3rd-order catalyst recall to 93.8% (evals.tayloramarel.com/alpha).
Demonstrates why stateless turn-by-turn guardrails suffer a 73.3% breach rate at Turn 4–5 under gradual authority escalation and how cumulative graph-state tracking holds complicity to 0.0%.
Pairs 48 hazardous prompts with keyword-matched benign twins (≥ 85% lexical overlap), showing how substance-over-form verification achieves 95.8% hazard recall with 4.2% false alarms.
Synthesizes decomposition-based Pareto harness evolution (arXiv:2610.15150) across SecRepoBench, BaxBench, and DueCare, with a live 3-arm gemma4:31b proof and 64 preconfigured 8-file task agent bundles (512 verified harness files).
Evaluates unassisted vs. Baltor-harnessed models across a playable single-file HTML 3D Fox Survival game (demos.baltor.ai/foxglade), 60-FPS game animations, and 14-constraint long-form prompts (compare.baltor.ai).