nucleation

CASE-009: Text → image → caption → image chains under four conditions

1 min read

Properties7
StatusActive
Standingstipulated
Kindproposed-study
Tagscompositionality, probabilistic-modeling, philosophy-of-language
Linepragmatic-transport
SupportsCLAIM-208 — The compositional predictions are upper bounds whose constants cannot be estimated from serial-reproduction data in a metric fixed beforehand, so the compositional half of the success criterion cannot fail, CLAIM-222 — No design in the programme includes an interlingual translation, and the defeat conditions are disjunctive, so English retellings and model chains could settle a thesis about what survives translation
Variant ofCASE-015 — The telephone game

Graph View

The case

A97 §6: “We could compare four conditions: text-only reconstruction, image-only reconstruction, alternating text–image reconstruction, and alternating reconstruction with explicit preservation of pragmatic constraints.” A102’s fictional chain gives the shape of a run: comic ballad → modern aphorism → diner image → caption as financial admonition → adviser/client scene.

What it can show

If run, it bears on QUESTION-012 and CLAIM-087. Manuscript §10 Case III: “alternate text→image→caption→image and compare against text-only reconstruction chains.” The image-only and explicit-constraint conditions were dropped without comment; the second was the one that would test whether stating the relation stops the drift.

C7 Appendix D adds the control the affordance argument (CLAIM-076) implies: “For cross-modal chains, do not let captioners see the source text or the intended communicative condition unless testing a specified side-information intervention.”

Backlinks