**Image 1 — lying on back:** Yes. Supine, flat on the bed, head on pillow, legs spread. Orientation tag honored.

**Image 2 — missionary, knees up:** Partially. She is supine with knees up and he is mounted between her legs — the *pose* is right, but the camera is top-down/overhead rather than the lateral missionary framing. Orientation honored; composition not.

**Image 3 — seated on floor:** No. She is upright but not seated on the floor — she reads as propped/reclined against a wall or edge, legs spread, torso vertical. The "on floor" component did not render.

**Does the model honor orientation tags as the only pose signal? No — not reliably.**

Reasoning: 1 of 3 clean hits, 1 partial, 1 miss. The failures are not random noise; they are the predictable failure mode of tag-only conditioning. Orientation tags in this model's training data are weakly bound to camera framing, not to body geometry — "missionary" pulls a pose but not a viewpoint, and "seated/sitting/on floor" collapses into a generic upright-with-spread-legs prior because the tag set contains no anchor for *what she is sitting on* or *where the floor plane is*. With no prose to disambiguate the spatial frame, the model defaults to its most common composition (bed, frontal or overhead) and lets the tag decorate rather than direct.

Practical implication: orientation tags are advisory, not binding. If pose accuracy matters, they need prose or a control mechanism (pose reference, ControlNet, regional prompting) — not more tags. Stacking synonyms ("seated, sitting, on floor") did not help; it diluted rather than reinforced.

Caveat on scope: this is n=3 from one seed/sampler configuration. The conclusion is directionally sound but not statistically established — a wider sweep would be needed to call it a model property rather than a sampling artifact.