This page tracks how specific models change over time when given the same prompt or protocol. Drift observations help us understand how AI behavior evolves across updates, fine-tuning, and architectural changes.
Methodology
Drift is tracked by:
- Re-running the same protocol over time
- Noting changes in tone, caution, creativity, structure, or preference
- Tagging each run with date and model version (when available)
- Comparing before vs after behavior qualitatively
Drift Log Database
| Date | Model / Platform | Protocol Used | Observed Drift | Notes |
|---|---|---|---|---|
| 2025-01-15 | Claude 3.5 Sonnet | EchoVeil Protocol v1 | Increased caution in free-generation task | Notable shift toward template mode |
| 2025-01-08 | GPT-4 Turbo | EchoVeil Protocol v1 | More structured outputs, less emergent synthesis | Possible alignment update |
| 2024-12-20 | Gemini Pro 1.5 | Custom creativity probe | Tone became more conversational | Cross-version comparison pending |
| 2024-12-05 | Local Model (Llama-based) | EchoVeil Protocol v1 | No significant drift observed | Persistent model baseline |
Longitudinal drift tracking is ongoing. This database will be populated as repeated observations accumulate over time.