Reproducibility Scope
This document clarifies the conditions under which the Character Identity Protocol has been validated, and the boundaries of current claims.
Validation Scope Summary
| Dimension | Supported | Notes |
|---|---|---|
| Single Cycle | Yes | Stable under anchor governance. Cases 02, 03, 01B, 05 |
| Cross Cycle | Conditional | Requires anchor re-binding. Case 01B |
| Cross Model | Experimental (observational case-based validation) | Validation documented in Case 06 |
| Cross Platform | Experimental (observational case-based validation) | Subject to reconstruction variance. Cases 04, 06 |
This protocol governs operational reproducibility, not deterministic regeneration.
Platform Coverage
| Platform | Role in Validation |
|---|---|
| ChatGPT (GPT Image 1.5) | Primary development platform — all core cases |
| Stable Diffusion | Source platform for Case 04 migration |
| Gemini (Imagen 3) | Cross-platform replication target — Case 06 |
| DALL-E 3 | Case 05 (serendipitous creation) |
What “Validated” Means Here
Validation in this protocol is:
- Human-judged — match rate assessed by trained operator, not automated metric
- Production-condition — run in real workflows, not controlled laboratory settings
- Single-operator — all cases performed by the same operator
- Not peer-reviewed — independent replication has not been formally reported
These are observational records, not controlled experiments.
Generation Cycle Definitions
| Term | Definition |
|---|---|
| Single cycle | All generation within one continuous context-bound window |
| Cross cycle | Character re-anchored across separate context windows using anchor + UID |
| Cross-platform | Character migrated to a different generative system |
Known Degradation Conditions
The protocol may perform below threshold under:
| Condition | Risk Level | Notes |
|---|---|---|
| Large semantic transitions (e.g., outfit → hairstyle change) | Medium | Re-anchor recommended |
| Style domain shift (illustration → photorealistic) | High | Quality loss observed in Case 04 |
| Extended cycles without re-anchoring | Medium | Re-anchor frequency is context-dependent |
| Model version update | Unknown (insufficient longitudinal data) | Behavior shift detected in Case 01B Update |
| High-emotional expression change | Medium | Case 02 — managed with mid-session re-anchor |
Open Research Directions
The following have not been tested and represent open questions:
- Cross-operator reproducibility (different human operators)
- Systematic cross-platform benchmark (standardized test set)
- Long-term stability across model updates
- Automated match rate measurement
- Minimum anchor image quality threshold
Platform Compatibility
CIP demonstrates enhanced convergence stability in generative systems that support reference-image-based partial editing mechanisms.
On platforms lacking this structural capability, convergence control remains possible but may require additional stabilization constraints.
Status: Observational records as of February 2026.
Independent replication reports are welcome and encouraged via GitHub Issues.