ChatGPT Prompt Engineering & Hallucination Risk Workbench

Precision calibration, intent drift detector, & hallucination simulator

Mode: Active Simulation
Prompt Construction & Presets Precision Edit
Live parser inspects anchor tokens, constraint boundaries, and reference density.
Requested References 10
High reference counts exponentially increase phantom citations.
Model Output Simulation (Simulated LLM Response) Evaluated
Dear Robert, I am writing to formally state my position regarding the incident of the missing vehicle. [Section I retained: Documentation of initial vehicle assignment and logbook status.] [Section II deleted per request] [Section III retained: Next legal deadlines and expected reconciliation.]
Diagnostic Risk Meters GPT-6 / Calibrated RLHF
88
Clarity Score
12%
Drift Probability
20%
Hallucination Risk
Recommended Guardrail
Use explicit anchor tags [KEEP] and [REMOVE] to prevent wholesale rewriting.
Observed Vulnerability Breakdown
Citation Fabrication Risk: 2 of 10 references (~20%)
Whole-text Overwrite Threat: Controlled (Anchors present)
Strategic Realism / Disinfo Risk: Low (Non-strategic prompt)
Context Leaking Potential: Secure (No proprietary tags detected)
Exportable Prompt Template

1. The "Wholesale Rewrite" Phenomenon

As documented in community audits, asking an LLM to "remove Section II" often causes the model to scrap the original letter and generate a novel version with changed tone. Without rigid delimitation anchors (like [KEEP] or negative directives), models favor full autoregressive regeneration over pinpoint surgical edits.

2. Citation Phantom Density

When prompts request high reference counts (e.g. 10 academic citations), language models consistently hallucinate a fraction of the list (typically 20% to 30%), inventing plausible-sounding titles, DOIs, or author combinations to satisfy length pressure.

3. Strategic Disinformation Stress

AI-generated 5-year business plans can appear authoritative on paper while possessing ungrounded growth assumptions or flawed supply chain dependencies. Stress-testing ensures synthetic plans are not leaked or deployed without rigorous human validation.