Observed Imagination Capabilities of Geometric Causal Intelligence (GCI)

Abstract Generative artificial intelligence can produce fluent hypotheses, explanations, and experimental suggestions, but fluent generation alone does not establish that a problem was correctly formulated, that a proposed structure is identifiable, or that a conclusion follows from the supplied evidence. This report examines GCI Imagination, an implemented research system within the Geometric Causal Intelligence programme. For empirical purposes, imagination is defined through observable functions: problem reformulation; construction of candidate structural explanations; derivation under explicit evidence constraints; preservation of competing explanations; formulation of a discriminating experiment when ambiguity is separable; refusal of unique authority when the evidence cannot identify a unique answer; and deterministic traceability.The study evaluates frozen software qualification and staged black-box batteries while separating implementation, qualification, prospective behavior, and scientific-discovery claims. GCI Imagination v4.1.0 passed 201 of 201 inherited qualification tests in a reported Windows run and in a manuscript-preparation rerun. Its predecessor, v4.0.0-r1, failed a frozen internally authored prospective battery despite passing deterministic replay, oracle isolation, and execution-completeness gates: 7 of 22 identifiable cases reached the identifiable terminal class, 5 of 8 non-identifiable cases reached refusal, and 0 of 6 ambiguity cases reached experiment construction. After unsealing, the revised v4.1.0 system obtained terminal-class agreement of 22/22, 8/8, and 6/6 on the same now-spent battery. This is remediation evidence, not prospective generalization evidence. A subsequently frozen v4.1.0-r1 public execution produced identical repeated outputs with no reported oracle access and no capability holds, but its sealed-oracle scientific score was unavailable at manuscript freeze.The results do not demonstrate general superiority over generative AI or independent discovery of a natural law. They demonstrate an implemented, replayable epistemic workflow whose externally visible contract distinguishes discovery, experiment, refusal, and unresolved status. Proprietary internal operators, search procedures, thresholds, routing rules, data structures, and certificate schemas are intentionally excluded. This limits source-level replication but preserves independent falsifiability through preregistered black-box evaluation. Keywords: Geometric Causal Intelligence; causal discovery; identifiability; scientific imagination; experiment design; deterministic computation; generative AI; reproducibility; epistemic refusal.