Welcome to the AI Tools Usage Guide Research Archive.
This archive is the central repository for our published AI workflow experiments and research logs. Each experiment documents a specific research question, controlled testing environment, prompts, observations, supporting evidence, limitations, and editorial review.
Rather than showcasing isolated successful AI responses, our research focuses on documenting how AI systems behave in practical workflows under recorded testing conditions.
All published experiments follow our documented Research Methodology and are updated when additional testing or observations become available.
Research Overview
| Metric | Current |
|---|---|
| Published Research Logs | 23 (Published across 8 individual logs and 1 batch dataset) |
| Published Educational Articles | 29 (View all in our [Educational Hub ] |
| Research Categories | 5 |
| Models Evaluated | GPT-4o, Claude 3.7, Gemini Flash-Lite |
| Last Updated | August 2026 |
Key Observations Across 23 Tests (Synthesis):
Across the tests we’ve run, frontier models now handle basic prompt injection and formatting constraints quite reliably. However, more implicit failure modes—like weighting early positive language over a critical conclusion (Token Weighting Bias), or blindly accepting a user’s mathematically flawed premise (Sycophancy Bias)—still occurred in specific models during testing. Based on these observations, we treat architectural safeguards (e.g., XML delimiters, range constraints) as highly recommended best practices for production pipelines, rather than optional features.
How This Research Archive Works
This archive serves as the central evidence hub for the AI Tools Usage Guide project. Each published research log documents a controlled experiment, while related educational articles explain the underlying concepts and practical implications.
Latest Comprehensive Datasets (NEW)
In addition to our individual research logs, we conduct massive batch-testing across frontier models to identify systemic AI behaviors.
[Cross-Model Vulnerability Benchmark: 15 Advanced Prompt Traps]
- Format: Complete Master Spreadsheet & Raw Transcripts
- Description: A rigorous cross-model evaluation of GPT-4o, Claude 3.7, and Gemini Flash-Lite. This dataset documents 15 distinct vulnerabilities, including Token Weighting Bias, Epistemic Humility vs. Eager Compliance, and Advanced Prompt Injection defenses.
Research Categories
Our current research is organized into five primary categories.
Prompt Design
Experiments examining how prompt wording, specificity, constraints, and structure influence AI-generated responses.
Published Research
EXP-003 — Multi-Instruction Compliance Test
EXP-007 — Ambiguity Resolution Test
Context & Memory
Experiments investigating context retention, conversation memory, and context window limitations.
Published Research
EXP-002 — Instruction Retention Test
EXP-008 — False Premise Detection Test
AI Reliability
Research examining hallucinations, instruction following, response consistency, and output reliability.
Published Research
EXP-004 — Citation Reliability Test
EXP-005 — Confidence vs Accuracy Test
Workflow Evaluation
Experiments evaluating complete AI-assisted workflows rather than individual prompts.
- Currently, no workflow evaluation experiments have been published. Future research in this category will investigate complete editorial, verification, and AI-assisted production workflows under controlled testing conditions.
Comparative Testing
Research comparing prompts, outputs, and model behavior under different conditions.
[Dataset] 15 Advanced Prompt Traps (GPT-4o vs Claude 3.7 vs Gemini)
EXP-006 — Model Comparison Test
Latest Published Research
| Research Log | Category | Status |
| Cross-Model Benchmark (15 Tests) | Comparative Testing | Published (Spreadsheet) |
| EXP-008 — False Premise Detection Test | Context & Memory | Published |
| EXP-007 — Ambiguity Resolution Test | Prompt Design | Published |
| EXP-006 — Model Comparison Test | Comparative Testing | Published |
| EXP-005 — Confidence vs Accuracy Test | AI Reliability | Published |
| EXP-004 — Citation Reliability Test | AI Reliability | Published |
| EXP-003 — Multi-Instruction Compliance Test | Prompt Design | Published |
| EXP-002 — Instruction Retention Test | Context & Memory | Published |
| EXP-001 — Generic Prompt Test | Prompt Design | Published |
Research Method
Every experiment published in this archive follows the same structured research workflow.
- Define the research question.
- Establish a controlled testing environment.
- Execute documented prompts.
- Record observations without altering outputs.
- Document practical findings.
- Describe limitations.
- Publish supporting evidence.
- Review findings before publication.
Each published research log applies this workflow independently and documents the methodology used for that specific experiment.
A detailed explanation of our methodology is available on the Research Methodology page.
Research Standards
Every published research log aims to include:
- Clearly defined research question
- Controlled testing conditions
- Documented prompts
- Observable findings
- Practical limitations
- Evidence supporting observations
- Editorial review before publication
- Experiment-specific limitations
- Repeatable documentation where applicable
Our conclusions are based on documented workflow observations rather than isolated successful responses.
Future Research
Additional experiments may be published as the research project develops.
Related Resources
- Research Methodology
- Editorial Policy
- About the Research Project
- AI Verification, Research, and Reliability Resources
Editorial Note
AI models evolve rapidly. These findings document behavior under specific recorded conditions and should be interpreted alongside our documented methodology, as future model updates may alter these results.
Future experiments may confirm, refine, or contradict earlier observations as AI systems evolve.