Claude Skills by VectorSpaceLab

Build deterministic one-to-one target-to-source label mappings from source prediction frequencies.

Run reduced or full ILM-VP alternating label-remapping and prompt-optimization loops.

Diagnose visual-prompt label-mapping instability, precision, and concept-overlap explanations.

Apply and audit universal visual prompts for frozen source-model adaptation experiments.

Build and train a LoRA linear layer with frozen base weights and low-rank trainable factors.

Merge LoRA low-rank factors into a dense weight and verify deployment-equivalent inference.

Compute LoRA trainable parameter budgets and reduction ratios for adapted Transformer projections.

Run a bounded soft-mode LoRA recovery protocol with executable mechanism checks and source-boundary logging.

Use this skill when building a recovery experiment that trains only model-reprogramming parameters around a fixed source model. The skill focuses on deterministic small-data optimization traces proving source immutability, parameter updates, loss reduction, and numeric target metrics.

Use this skill when converting lower-dimensional target vectors into a frozen source model input space by zero-padding and applying a masked, trainable additive reprogramming pattern. It preserves target coordinates and exposes metadata for downstream training checks.

Use this skill after a model-reprogramming recovery run to check whether the observed metric came from the intended mechanism: frozen source model, updated reprogramming parameters, reduced target loss, compatible target metadata, and optional alignment-distance decrease.

Use this skill when converting frozen source classifier outputs into target-class probabilities through non-overlapping many-to-one label groups or a small trainable linear head. It is appropriate for model reprogramming recovery when source weights must remain unchanged.

Compute VPT recovery accuracy, parameter efficiency, and mechanism report fields.

Audit VPT trainable scope so only prompts and heads change while backbone stays frozen.

Build shallow or deep visual prompt token insertions for VPT-style Transformer inputs.

Run a bounded tiny VPT optimization loop with frozen backbone and trainable prompts.

Apply policy-iteration style actor updates and verify objective improvement under compatible gradients.

Fit and validate compatible advantage critics whose residuals are orthogonal to policy score features.

Compute exact policy-gradient theorem checks for finite discounted MDPs with differentiable softmax policies.

Run a complete mechanism-faithful recovery experiment for the Sutton et al. policy-gradient theorem paper.

Validate IMPALA actor learner unrolls with behavior and learner policy metadata for V-trace recovery.

Run a bounded soft-mode IMPALA V-trace recovery harness and emit validator-compatible evidence.

Run a deterministic V-trace actor-critic optimizer step with value loss and policy-gradient diagnostics.

Compute IMPALA V-trace targets, clipped ratios, and policy-gradient advantages from off-policy trajectories.

Build and validate independent batched robotics environments for Isaac Gym style reduced recovery experiments.

Execute a deterministic batched simulation policy reward reset loop for Isaac Gym mechanism recovery.

Run a deterministic PPO style clipped scalar update over batched rollout evidence for recovery validation.

Validate direct state action observation and reward buffer contracts for Isaac Gym style pipelines.

Compute P3O effective sample size and derive adaptive clipping and KL coefficients from replay policy probabilities.

Build decomposed P3O surrogate loss terms with clipped replay gradients and behavior-to-target KL regularization.

Execute a bounded soft-mode proxy experiment that validates P3O mechanisms with generated skills and numeric training evidence.

Model P3O iteration ordering by appending current rollouts before bounded sequential replay mini-batch updates.

Assign mixed Gaussian exploration scales across parallel actors and produce reproducible noisy bounded actions.

Build and validate the process topology for Parallel Q-Learning with actors, replay, policy learning, and value learning.

Run a mechanism-faithful reduced Parallel Q-Learning proxy experiment with replay, mixed exploration, and optimizer evidence.

Compute speed-ratio and replay-overwrite diagnostics for massively parallel off-policy Q-learning configurations.

Compute and validate PPO clipped probability-ratio surrogate losses and diagnostics for policy update recovery or implementation.

Run a deterministic reduced PPO minibatch update loop with frozen old log probabilities and auditable optimizer traces.

Validate reduced PPO recovery results for target consistency, source-boundary compliance, numeric metrics, and mechanism-faithful proxy evidence.

Compute bootstrapped returns and generalized advantage estimates for fixed-horizon PPO rollout batches.

Evaluate soft-mode REINFORCE proxy recoveries for target consistency, numeric metrics, and mechanism-faithful evidence.

Run bounded REINFORCE stochastic-policy training loops with sampled actions, scalar rewards, baselines, and optimizer traces.

Compute REINFORCE score-function update terms for sampled stochastic policy actions with scalar rewards and baselines.

Compute SAC maximum entropy objective diagnostics for rewards, log probabilities, discounts, and temperature-scaled entropy bonuses.

Execute a bounded reduced SAC recovery experiment that combines generated objective, backup, and actor-update skills into validation artifacts.

Construct SAC soft state values, Q targets, and Bellman residuals from replay rewards, log probabilities, and critic estimates.

Run a deterministic SAC-style stochastic actor update with reparameterized Gaussian action, log probability, and policy-gradient diagnostics.

Build and optimize the finite-atom APT contrastive loss with proposal posterior correction.

Apply APT proposal-prior log-density corrections to candidate posterior scores for likelihood-free inference.

Evaluate soft-mode APT recovery records for source-boundary, metric, and mechanism-faithfulness checks.