Three provisionals filed May 7

Three provisional applications were filed at the United States Patent and Trademark Office on May 7. Mechanism descriptors and counsel posture for each are now surfaced on the per-patent pages.

K-Pool LoRA (App. 64/060,315)

A software method for continual supervised fine-tuning of frozen pre-trained language models. The method maintains a fixed-size pool of LoRA adapter snapshots over a frozen base, selects the active slot per training batch and per inference query through a gradient-detached content-aware routing module, opens unfitted slots via a bootstrap-and-novelty rule with a deterministic claim window, and updates only the active-slot parameters via a sign-quantized momentum optimizer. Sole inventor. Patent page.

MRRO (App. 64/060,392)

A method of optimizing a non-convex objective by iterative application of a renormalization-group-derived smoothing operator at a finite descending schedule of scales, with warm-starting between scales and a deterministic inverse-renormalization refinement step at the terminal scale that maps the converged coarse-grained coordinate back to the fine-grained continuous manifold by deterministic gradient-based descent. The unified operator family covers a frequency-domain low-pass cascade variant, a Gaussian convolution variant with a Monte Carlo estimator, and a block-decimation renormalization-group flow variant adapted from spin-lattice physics to continuous-objective optimization. Joint inventors. Patent page.

CMLGS (App. 64/060,404)

A computer-implemented zeroth-order optimization method for non-convex objective functions with regular spatial structure or periodic topology. At each step the method draws perturbation directions from a fixed-magnitude distribution, applies a one-dimensional coordinate-aligned coupled-map-lattice discrete-diffusion chain to each direction, evaluates a symmetric finite-difference at the perturbed coordinates, aggregates into a coupled gradient estimator, and applies gradient descent. Per-step computational cost is O(D) in problem dimension. The independent claim is scoped to objectives with regular spatial structure, periodic topology, or physical-lattice geometric origin, with the tiny-neural-network training workload explicitly excluded by the workload-class scope demarcation. Joint inventors. Patent page.

Where this leaves the portfolio

The May 7 filings bring the workspace’s filed-provisional count to six, up from three at the end of April. The portfolio index lives at tsugilabs.ai/patents. Twelve-month conversion and PCT deadlines for all three of these are 2027-05-07.