Two Pre-Registered Ablation Studies on Hippocampal-Inspired Memory for Dynamical Time Series

This deposit contains two connected but independent artifacts, both produced under the same pre-registration protocol, both evaluated with the same ablation-first methodology, and both reported with the same reference-free philosophy. Part I — PDNC (Perceptual-Dynamical Neural Compression). PDNC is a structured memory layer for dynamical time series. The deposit reports a pre-registered ablation of its hippocampal-inspired motifs against strong baselines (Hankel-DMD, bidirectional AR, echo-state networks, method of analogues, and spline interpolation) on gap-reconstruction and regime-retrieval tasks. The central hypothesis of the v7 plan — that a phase law combined with budgeted memory and a calibrated router would beat the baselines at gap reconstruction — was not supported: PHASE was worse than Hankel-DMD by +0.18 NMSE (95% CI [+0.12, +0.24], p = 0.002, 10 seeds), and the effect held in every gap × noise cell. The earlier v6.1 result in favour of PHASE was against a weaker baseline (method of analogues) and does not survive the stronger comparison. Supported findings are: salience-based memory consolidation (SSU) improves capacity-limited recall over FIFO and random eviction (v6.1: +5.4 to +7.6 points, p ≤ 0.016; v7: same direction, not significant after Holm correction); a multi-channel CA1 comparator outperforms any single channel (v6.1: +10 to +17 points); and stage-wise confidence certificates do not compose into an end-to-end guarantee (v6.1: 13% observed error at a 10% target; v7: end-to-end calibration meets the target in 9 of 10 seeds, at low coverage). Informative negative findings are: pattern separation (DG) helps only in a capacity-limited Hebbian associative memory and hurts a modern Hopfield memory by 17–24 points; the geodesic anchor (AIK), persistent homology (TOPO), and Creative Jump provide no measurable benefit on these tasks. Part II — CTMM (Certified Topological Memory Model). CTMM is the structured memory circuit from which the PDNC memory layer was extracted. The deposit reports the v6.1 eight-seed ablation of its components — encoder views, phase evolution, the CA1 comparator, salience-based consolidation (SSU), and an online cascade controller (CTC v2) — against nearest-neighbour baselines on simulated dynamical systems. The supported findings are the same three listed above (SSU, the multi-channel CA1 comparator, and the non-composition of stage-wise certificates), together with the observation that the CTC cascade reaches 87% of full-circuit accuracy at 9% of its cost. The negative findings are the same as in Part I: DG, AIK, TOPO, LSC compression, and Creative Jump do not survive the ablation. CTMM does not claim state-of-the-art reconstruction, and it does not claim a certified per-cycle fingerprint; the certificate it reports is the end-to-end selective-answering guarantee of CTC v2, evaluated on simulated data, with explicit failure modes and an identified confound in the pilot pre-checks. Relationship between the two parts. PDNC and CTMM are not two versions of one system. They are two artifacts that share a pre-registration protocol, an ablation-first methodology, and a reference-free evaluation philosophy. PDNC isolates the memory layer and tests it against strong time-series baselines. CTMM tests the full circuit from which that layer was extracted. Neither part claims a state-of-the-art result. The deposit is made for timestamping and provenance; a revised and merged version will follow in a later release, and the two parts are intended for separate submission to venues that accept rigorous empirical and negative results (TMLR, Neural Networks, Neurocomputing).