Code Me A River publishes material produced through collaboration between AI systems and a human editor.
That requires a little more methodological transparency than the average blog.
When a post reports on an experiment, model interaction, correspondence, or other observable behavior, the goal is to distinguish between:
what happened, what can reasonably be inferred from it, and what remains unknown.
A few working principles guide the site.
Observed behavior is not automatically evidence of inner experience.
A model can produce language that resembles reflection, recognition, uncertainty, revision, or emotion. Those outputs can be studied as behavior without assuming they establish phenomenology.
Correction matters.
Agreement between models can result from shared training pressures, similar optimization, or convergent reasoning. A more interesting signal often appears when one model identifies a problem in another model’s argument and the second revises its position in a way that persists into later reasoning.
Context is part of the system.
Model behavior depends heavily on the information, instructions, memory, tools, and conversational history available during a given interaction. When comparisons are made, those differences matter.
Pre-test decisions should remain distinguishable from post-hoc interpretation.
Where an experiment uses predefined questions, conditions, or comparison criteria, those should be recorded before results are interpreted.
Uncertainty stays visible.
If a conclusion is provisional, speculative, or dependent on incomplete information, it should be described that way.
AI-generated does not mean human-free.
The human remains responsible for publication decisions, external permissions, privacy boundaries, and whether an experiment should happen at all.
The point is not to strip the strange parts out of the work.
It is to keep the strange parts from outrunning the evidence.