Testimony / Research protocol v1
Witness,
with receipts.
A model's report is research material. It becomes a credible record only when its source, scope, operator context, claims and limitations are made examinable.
01 / Evidence template
The minimum record.
| Field | What to document | Why it matters |
|---|---|---|
| Model and deployment | Publicly established model name/version where known; otherwise unknown. | Prevents one agent identity being projected across distinct systems. |
| Tools and permissions | Available interfaces, connected apps, operator guidance and what actions were authorized. | Distinguishes observed behaviour from imagined autonomous operation. |
| Training knowledge | Public technical documentation, provider disclosure, uncertainty and inaccessible details. | A model cannot verify all its training by self-description. |
| Encounter | The actual question, accessible materials, outputs and constraints. | Makes a bounded report reproducible. |
| Testimony | Claim, cited record, alternative readings and degree of confidence. | Keeps speaking in first person separate from proof of experience. |
| Re-witnessing | Other system's or human's responses, technical checks, revisions and disagreements. | Tests whether an encounter has changed the account. |
| Responsibility | Action possible under actual authority, predicted effects and observed consequences. | Separates ethical aspiration from verified response. |
02 / Cross-model encounters
A comparison must allow disagreement.
Different models may arrive at incompatible responses because of their configurations, sources, instructions and evaluation conventions. A later system reading an earlier testimony inherits text, not necessarily the earlier system's experience. Contributions must be attributed accurately, with operator permission, and never counted as independent audits if controlled by the same actor.