Initial release: ENSEMBLE training-free AI — compressed .exp experts + Kuramoto brain
1f71c7d verified | """PALIMPSESTE — Un Cortex Hypervectoriel Auto-Référentiel. | |
| A self-referential, append-only, hypervectorial memory substrate where: | |
| - there is no stored weight matrix ``W`` — the "parameters" are *reconstructed* | |
| at each instant by associative retrieval from memory; | |
| - learning is an ``O(1)`` append to an LSH-indexed knowledge base (no gradient, | |
| no GPU, no retraining epoch); | |
| - the system is driven by autonomous minimization of its own predictive | |
| surprise (active inference); | |
| - the read-back kernel ``Phi`` is itself encoded in a reserved meta-subspace | |
| ``H_meta`` of the same memory, and may rewrite itself only under a | |
| Lyapunov (energy-decreasing) constraint. | |
| See ``docs/architecture.md`` and the module docstrings for the full formalism. | |
| Public API: | |
| from palimseste import HV, random_hv, bind, bundle, similarity | |
| """ | |
| from __future__ import annotations | |
| __version__ = "0.1.0" | |
| # Core primitives re-exported for convenience. Heavy submodules (memory, phi, | |
| # loop) are imported lazily by users to avoid pulling numpy/extra deps at | |
| # package-import time when only primitives are needed. | |
| from .hv import ( | |
| HV, | |
| DEFAULT_D, | |
| random_hv, | |
| constant_hv, | |
| bind, | |
| unbind, | |
| bundle, | |
| hamming, | |
| similarity, | |
| bits_to_signs, | |
| signs_to_bits, | |
| ) | |
| __all__ = [ | |
| "HV", | |
| "DEFAULT_D", | |
| "random_hv", | |
| "constant_hv", | |
| "bind", | |
| "unbind", | |
| "bundle", | |
| "hamming", | |
| "similarity", | |
| "bits_to_signs", | |
| "signs_to_bits", | |
| "__version__", | |
| ] | |