Does Explicit Episode–Trace Binding Help? An Exploratory Study with Information-Matched, Competence-Audited Controls

We study a narrow association benchmark: whether a model attributes the source of an episode (own, imported, or inferred) better when its record is explicitly bound to a summary of the internal change it produced (D) than when an information-matched control receives the same streams without the binding and must learn the pairing (F). Motivated by provenance in persistent agents, the pilots are small, synthetic, open-loop, and offline, and their contribution is methodological: a decoding ceiling, a control-competence audit, and equivalence tests that make a null result interpretable. In five adaptive, exploratory iterations, generic controls first failed to use the trace, producing a large but uninformative advantage for D (+0.38 BSS). With a shared label-free objective, the control reached 87% of the ceiling and D − F was +0.002 BSS; fresh seeds (+0.001) and a generic masked-token objective (−0.023) gave descriptive pilot equivalence under the provisional ±0.05 BSS margin, with no multiplicity adjustment across the adaptive pilot path, although D began learning sooner when the association was not supervised. Decoder diagnostics indicated that the source was carried by local update statistics and that resetting the state before each episode did not reduce decodability, so the pilots do not test attribution from path-dependent history. With 32 episodes per window, an apparent reversal between objectives was consistent with a shortcut in how one objective was applied to D, and removing it repaired the reversal; D began learning in 9–10 of 10 seeds under every objective and F only under a join-supervised one, while skill once both learned remained undetermined. We close this benchmark and state open problems on functional provenance.