Pipeline map · tap each stage
Where does "corporate incentive" actually enter?
The honest answer: not at one spot. It enters at different mechanisms with very different levels of directness and permanence — some literally rewrite weights, some are just text added at runtime, some sit entirely outside the model. Tap a stage.
weight-level (trained in)
context-level (runtime)
external (bolt-on)
upstream (organizational)
What's documented vs. estimated: the mechanisms described here (pretraining objective, RLHF, Constitutional AI, system prompts, external classifiers) are publicly documented by Anthropic. The "directness" bars are my own pedagogical estimate to help visualize the difference between mechanisms — not a measured quantity, and I don't have privileged access to the actual weighting Anthropic gives commercial vs. safety vs. helpfulness considerations inside the reward model. I also can't introspect on my own weights to verify any of this from the inside — this is reasoning about the documented architecture, not a report from within it.