Reproducibility
Whether a trajectory, or a relevant slice of it, can be replayed under stated conditions.
Formal definition
Reproducibility. A canonical LLMography term. See the formal definition in the article below.
Formal definition
In LLMography, reproducibility is the possibility of replaying a decision trajectory, or a relevant slice of it, under documented conditions and residual uncertainty.
In plain language
Reproducibility is not a dashboard widget. It is a way of naming a part of Human–AI work that otherwise disappears into logs, chat history, or an unexamined final answer.
Why it matters
As systems move from prompts to agents and loops, decisions are produced across many steps. Without this concept, those steps remain difficult to reconstruct, compare, or audit.
Diagram
The signature object of LLMography is a trajectory: human, model, tool, observation, verification, revision, decision. Reproducibility occupies a distinct position on that object.
Related metrics
See the proposed measures for loop autonomy, verification strength, and decision-trace completeness. These are research indicators, not certified scores.
Example
A recommendation is issued after a tool call fails verification and a human requests a narrower claim. Reconstructing that path is an instance of this concept in use.
Citation
Bousmah, M. LLMography concept note: Reproducibility. LLMography research site.