HIPPOGRID evidence pack for the preliminary results (2026)

Abstract

HIPPOGRID — evidence pack for the preliminary results (2026) Raw result files, figures and, from version 3, the reproduction code behind the preliminary results cited in the MSCA-PF proposal HIPPOGRID (SEP-211387371, call HORIZON-MSCA-2026-PF-01). Researcher: Álvaro González-Redondo (University of Granada). Each bundle below backs one claim of the proposal; each result file is the unmodified output of the experiment that produced it, and the commit hash anchors it in the (currently private) research repository, scheduled for open release (Apache-2.0) with the corresponding papers. Numbers were produced with 10 random seeds and paired comparisons against non-degenerate null models unless stated otherwise. License: CC-BY 4.0. Contact: alvarogr@ugr.es Bundle Claim it backs (as in the proposal) Key files Commit B1_real_sensor_binding Image and posture written to one shared address are recovered from each other on a real robot's sensor stream (8,346 frames, 10 seeds); writing to a shuffled address destroys the binding asymmetrically while the memory stays intact n562_real.json, figure 5498007e B2_innate_template_binding A one-shot Hebbian binding anchors a fixed grid template at 1.0–1.4 cells median under biological-level velocity noise (10/10 seeds; bounded at 3× that noise; shuffled-binding and shuffled-velocity nulls fail; the free integrator drifts 18–301 cells at the same noise levels). The bound template re-locks on the memory after blind stretches, and the result holds with spatially uncorrelated content and on a continuous attractor network under the same binding. Under room deformation the stored association is kept (a map built from scratch in the stretched room is rejected as a null) and the map rescales partially along the deformed axis only: field spacing follows 63–73 % of the room change at the default gain (48 % in rats, Barry et al. 2007; the shortfall is a rewrite term plus a correction-gain term, both measured, mechanism open); two maps merge by local reorganisation at the old wall that relaxes over the session, with seam displacements an order of magnitude below those of Wernle et al. 2018. n569_*.json (anchoring, nulls), n571_*.json (deformation and merge at 10 seeds), n572_*.json (blind stretches, interpolation, reversal, heading noise), n573_*.json (attractor network), n574_*.json (uncorrelated content), n575_*.json (lattice-vector spacing instrument, return by columns, merge in spacing units), n576_*.json (free-integrator drift, correction-gain sweep, rewrite gain, fixed-projection control, sensory noise); figures of the proposal (n569_*.png, n571_*.png) and of the preprint draft v0.15 (figures_preprint_v015/) 0f20cb40 (n569) … cde48128 (n576) B3_rate_grid_recipe A purely local Hebbian/anti-Hebbian rule yields clean 2-D grid maps with no explicit normalisation: best-unit grid score 0.81 (torus) / 0.71 (walled box), medians over 10 seeds n565_toro.json, n565_caja.json, n565_repro.json, figure 2c6455c7 B4_place_stage_chain The chain closes from raw sensed input through a competitive place stage (grid score ≥ 0.4 in 9/10 seeds; direct input 0/20); small modules of distinct grid units at small grid-layer sizes; a stored-experience place stage performs like a trained one n567_u4096_rejilla.json, n568_10semillas.json, n568_nulos.json, n568_redundancia.json 0243e5b0 B5_blackboard_capacity With learned addresses the blackboard's capacity grows ~N/2 vs ~N^0.45 without learning (n562_escalado.json, keys ley*); the worst alias between two distant places falls from 0.995 to 0.093 at 4,096 cells (n562_aprendido_dos_tablas.json; N-dependent: 0.904 at 1,024 cells) n562_escalado.json, n562_aprendido_dos_tablas.json 16d3322e / 3a22657e B6_local_rule_ceiling A single-layer delta rule with a self-generated teacher ties the exact optimum of its family (+0.71 vs +0.70) and the non-negative constraint improves it (+0.82, 16/16) n541_techo_exacto.json, n542_delta.json, n542_rectificada.json d764a6fd CODE_innate_template Reproduction package for bundle B2 and for the preprint: the simulation chain (walks, multiscale input, place stage, phase template, one-shot binding, correction loop, decoders), every battery n569–n576, the figure scripts and the manuscript build script, plus a copy of the 40 result JSONs under results/. README.md gives dependencies, order of execution and what each script produces; MANIFEST.md lists every file with its role. Two generated inputs (random walks and frozen place-stage prototypes, ≈ 143 MB for 10 seeds) are not deposited; they are deterministic given the seed and the README says how to regenerate them. Known limitation, declared in the README: until n576 the velocity-noise generator was seeded per process, so the n569 rows reproduce statistically but not bit for bit; the seeding is deterministic from n576 on. CODE_innate_template/ cde48128 Figures are included where the proposal reproduces them. JSON field names are in Spanish (the project's working language); each file is self-describing (field "que"), with seeds and conditions as keys. Version history v1 (2 Sep 2026): bundles B1–B6 as cited in the proposal. v2 (2 Sep 2026): adds the room-deformation battery at 10 seeds to B2 (n571_*). v3 (3 Sep 2026): adds to B2 the results of the second and third rounds of review of the preprint (n572_* to n576_*: blind stretches, interpolated deformation, return and compression, merge at 400k steps with its null, attractor-network replication, spatially uncorrelated content, lattice-vector spacing instrument replacing the 1-D autocorrelation one, free-integrator drift, correction-gain sweep, rewrite gain from the first visit, fixed-projection control, sensory noise per visit) and the figures of the preprint draft; adds the reproduction code (bundle CODE_innate_template). Cited by the preprint "A one-shot binding of arbitrary sensory content to a fixed grid template defends its frame under environmental deformation" (bioRxiv, September 2026; title as of draft v0.15). v4 (9 Sep 2026) preserves the scientific files from version 3 and adds bundle B7_state_dependent_motion, documenting the earlier controlled simulation that motivates the generalisation question in Section 1.1 of HIPPOGRID. B7 contains the original experiment source and historical report from commit 8a50844ab7f4795cf49847cb5a7139701c2028fd, a subsequent scope correction, and an explicitly labelled CSV transcription of rounded aggregate results. In this three-seed planar simulation, a state-dependent multilayer perceptron with supplied coordinates retains its central-range fit while displacement-field extrapolation worsens as command effects become more position-dependent. This study does not test the proposal's constant-generator operator on perception-anchored codes and does not measure Riemannian curvature. No new experiments were run for this release. The historical addition contains reported aggregates rather than raw per-seed results; it does not include the original execution environment or a newly verified reproduction. Updated documentation explains these limits and supersedes broader claims in the historical report. All previous scientific files remain unchanged; the previous root README is archived, and file manifests provide SHA256 hashes for integrity checks.