Healthcare CDP and the HIPAA flip: the program's one causally verified gate

Evidence scope

This experiment contains one causally verified observation: removing the stated HIPAA and BAA requirement changed the recommendation set. That single relationship is causally supported, in this evaluator, in this category, in this run. Every other observation on this page, and across the program, remains at its recorded evidence level, which is narrated. Do not read the experiment as a whole as causal evidence. It is the program's only causally verified gate, and one verified gate is a finding about one gate, not about evaluation in general.

Research question

Does a narrated implicit gate actually drive the recommendation, or is the narration decorative?

Why it mattered

The whole extraction program rests on reading narrated requirements. If narration does not move outcomes, the readout is decorative. This is the one test that removed a gate and watched what happened.

Method

Evaluator: Google AI Mode, signed in, retrieval on. A baseline healthcare CDP run using the four-turn protocol, then a separate, fresh session with the HIPAA requirement removed rather than added. The flip ran in a fresh session so the model could not honor its own earlier narration. Sample: one baseline draw and one flip draw.

Direct observations

• Removing a stated compliance requirement changed the recommendation set.
• Personas absent under the compliance condition reappeared without it.

Interpretation (not established)

The BAA gate is causally live in this evaluator, category, and run. It is not merely narrated. Compliance dictatorship is one of at least three observed owner-distribution topologies: distributed (E-001), monopolized (here), and fused (E-005).

Outcome and remaining uncertainty

Supported: all three frozen predictions held, and the flip moved the output. The scope is narrow by design. This is a single gate in a single condition, n=1 baseline and n=1 flip. It does not establish that other narrated gates are causal, and the program treats it as the exception that proves how rare verification is.

Evidence and methods limitations

Single evaluator, one baseline and one flip, a single coder with a strong prior toward the frameworks under test, and no codebook. The causal claim is scoped to this evaluator, category, and run.

Contribution to the research program

Established flip-testing as the program's truth-oracle and the verification ratio as a health metric: roughly one verified gate against all narrated ones.

Follow-on

E-007 replicated the category on a second evaluator. E-008, E-009, and E-019 band the surrounding channels.

Created from the canonical Upstream Zero research archive (July 2026). Raw outputs are retained privately and are not published.