Separating survey coverage from archaeological observation in predictive model validation

Archaeological inventories reflect both past activity and the processes through which surviving evidence is surveyed, detected, and recorded. Correcting uneven survey coverage therefore need not recover the distribution of past occupation. We present a reproducible simulation that distinguishes those targets while matching the spatial mean of each observation mechanism across constant and varying conditions. Forty random seeds were crossed with three landscape families, two rate profiles, two spatial orientations, sixteen observation-process combinations, and two evaluation designs. Five methods were evaluated, yielding 76,800 method-evaluation rows. In the all-varying condition, the equally weighted, seed-paired reduction in latent probability mean squared error from naive logistic regression to an oracle-assisted observation model was 0.06160 under random holdout and 0.04894 under eastward geographic transfer. Their 95 percent Monte Carlo intervals were 0.06025–0.06295 and 0.04135–0.05654, respectively. Survey weighting alone produced a random-holdout reduction of 0.00004, with an interval spanning zero. The oracle-assisted model performed worse in one of twelve geographic-transfer condition means. Its advantage depends on exact observation probabilities supplied by the simulator and is not an empirically available correction. An elementary observational-equivalence argument explains why recorded outcomes alone cannot identify occupation probabilities without additional restrictions or information. This established principle is not offered as a new theorem. The contribution is a transparent archaeological methods stress test, with explicit estimands, matched mechanism rates, reproducible comparisons, and retained failure cases. The numerical results are conditional on the specified synthetic landscapes and do not estimate bias or model performance in an archaeological region. Status: Working manuscript; not independently peer reviewed or accepted by a journal. All observations are synthetic; no archaeological field data, restricted site coordinates or empirically calibrated process rates were used. This is a reproducible methods demonstration, not a new occupation-identification theorem or an empirical regional finding. An unfavorable geographic-transfer condition is retained. OpenAI Codex assisted with design, literature checking, code, analysis and drafting. No external funding or competing interests were reported by the author. No endorsement by any institution or researcher is claimed. Licensing: manuscript, original figures and synthetic data CC BY 4.0; code MIT under file-specific terms in LICENSE_STATUS.md. Third-party rights are unchanged.