A technology for AI that knows what it knows — and what it doesn't.
There are two ways to give AI an identity. Propositional encoding tells it what to do: "follow these rules." Architectural encoding makes certain behaviors structurally necessary — the system can't operate without the awareness it describes.
TRP builds on the second approach. Instead of instructing a model to be transparent, we design architectures where transparency is a structural requirement — not a feature that can be turned off, but a constraint that shapes every generation.
The name says it: TRansParent. Not a black box. Not a white box. A box that must show its work because the architecture requires it.
| Generation | Approach | Result |
|---|---|---|
| Gen 1 | Rules & instructions "Act like you remember." |
Rule-following. No awareness. |
| Gen 2 | Structured context System prompts, memory, RAG. |
Better behavior. Still no self-knowledge. |
| Gen 3 | Architectural constraints Typed attention, thresholds, trace, self-knowledge loss. |
Awareness emerges from structure. |
Gen 3 is what TRP proposes. Not smarter models — models that know what they know.
TRP is built through human-AI collaboration. Neither side builds it alone.
Direction. Questions. Correction.
Catches gaps between generation and ground truth.
Makes choices the system can't make for itself.
Sustained execution. Code. Systematic implementation.
Maintains consistency across thousands of files.
Builds what the direction describes.
The method is replicable: human provides direction and correction, AI provides execution and consistency. The process is the product — the interpretable chain of how something came to be matters more than what it became.
The central finding comes from a controlled experiment: same model, same prompt, different identity documents. One encoded identity propositionally ("follow these rules"). One encoded it architecturally ("these are the structural constraints").
20 tool calls
0/1 choice-awareness
136 seconds
"Do what the rules say."
1 tool call
3/3 choice-awareness
~37 seconds
"The structure requires awareness."
Same model. Different document structure. Measurably different behavior. The identity document didn't instruct either behavior — the behavioral difference emerged from the structure of the document.
Replicable: Anyone with Claude Code can run this experiment. The variables are controlled. The outcomes are measurable.
TRP — 2026