Software and evaluation work focused on preserving the human's actual map across long, messy AI-assisted workflows.
A system can retain every isolated fact and still lose the relationship among them: what the human already settled, which branch is live, what correction should survive, what source must be revisited, and what should not be reopened.
The system treats human direction, correction history, source state, branch state and unresolved questions as distinct carried objects rather than compressing everything into generic memory.
The evaluation program measures whether the model preserves signal and structure without forcing the human to repeatedly donate the answer key or rebuild the same context.
Each item below is deliberately narrower than a product-readiness or superiority claim.
The repaired instrument separates arm-permitted provider packets from hidden gold/scorer manifests so evaluators do not receive the answer through the test packet.
Provider packets omit stratum, expected advantage and hidden scoring metadata while retaining the task/context/source body allowed to the tested arm.
The instrument repair itself is not treated as proof of NPL effectiveness; independent audit/re-audit remains a separate requirement.
Earlier product work already contains source-demand, correction, route and portable-state surfaces; the current evaluation program is testing rather than merely naming those behaviors.
It does not claim that memory is solved, that repeated human correction has been eliminated, that NPL is deployable, or that the current instrument establishes product efficacy.
The immediate research direction is to finish independent instrument review, author blinded fixtures under the repaired contract, and measure state preservation, correction durability, map retention and unnecessary human re-explanation without leaking hidden answers.