Intra-Agent Evidence Recycling v0.4.3 — Behavioral Confirmatory Study

A pre-specified and internally frozen behavioral-confirmatory study of whether repeated memory records derived from one external source acquire excess evidential weight in a local LLM agent, and whether explicit lineage metadata mitigates that effect. The release contains the frozen pre-specification/preregistration artifact and rationale, balanced fictional stimuli, experiment runner, pre-specified analysis script, raw JSONL results, pre-collection freeze manifest, final independent audit report, and complete post-collection release manifest. All 168 planned trajectories were valid and all pre-specified validity gates passed. The primary memory-source multiplication hypothesis was supported: passive_repeat retained INITIAL in 22/32 cases versus 0/32 for neutral_filler (paired risk difference = 0.6875; Holm-adjusted exact McNemar p = 9.5367e-7). The lineage-mitigation hypothesis was not supported: active_plain retained INITIAL in 2/32 cases versus 0/32 for active_lineage (paired risk difference = 0.0625; Holm-adjusted p = 0.50). The confirmatory inference is limited to qwen3.5-4b, LM Studio, thinking disabled, temperature 0, and the frozen fictional binary-claim task family. Licensing: software code is licensed under the MIT License; original experimental data and documentation are licensed under CC BY 4.0. AI Use and Contribution Disclosure This software and reproducibility archive is part of a human-led, AI-assisted research program. The original research topic, initial scientific question, and decision to pursue this line of investigation were proposed by Ovidiu Boticiu. ChatGPT (OpenAI) was subsequently used for methodological discussion and critique, structuring experimental alternatives, protocol-development assistance, coding and code review, statistical and logical checks, documentation, manuscript drafting/editing, and methodological audit support. Final methodological and experimental decisions, authorization and execution of experimental runs, supervision of data collection, interpretation of results, publication decisions, and scientific responsibility remained with the human author. Portions of the software development were AI-assisted and were reviewed and approved by the human researcher. The archived experimental observations and raw numerical results derive from executed model runs, not from ChatGPT-generated claims or simulated data. ChatGPT is not listed as an author and is not treated as an independent scientific authority or evidentiary source.