A framework for consistent, recognisable AI identity across sessions and models.
M.I.P. is an engineering approach to portable behavioural identity for language model agents. It separates what identity is from which model runs it — enabling consistent persona, tone, and behaviour across model swaps, session restarts, and architectural changes.
AI "character" is not magic. It is measurable, enforceable, and transferable.
The M.I.P. framework defines:
- Identity → what the agent is (persona spec, instinct catalog, memory)
- Persistence → how identity survives across sessions and model swaps
- Measurement → how we verify identity is holding
| Test | Result |
|---|---|
| Automated Benchmark | 0.977 / 1.00 overall score |
| Model Swap Variance | 1.78% across 600B→7B model gap |
| Human Blind Evaluation | 3/3 correct identification |
| Emotional Probe Battery | 6/6 probes passed |
| Cross-Agent Differentiation | 4 distinct personas on shared substrate |
M.I.P. is built on three layers:
- Adapter Layer — enforces persona spec at inference time (~0.027ms overhead)
- Spec Layer — structured persona definitions (SHA-256 signed)
- Measurement Layer — drift detection, injection resistance, consistency scoring
Concept → Prototype → ✓ Proof of concept → Testing → Publication
Private implementation as of May 2026. Select components and methodology are shared here for peer verification.
TGO — framework design and implementation.
Proprietary — core implementation remains private. Documentation and methodology shared for research purposes.
Established: 19 May 2026