Claude Skills by VectorSpaceLab

Run a bounded BridgeStan score-model proxy recovery that invokes generated module scripts, logs command evidence, and writes validator-ready recovery artifacts.

Apply scalar Stan-style constrained and unconstrained transforms with support checks and log-Jacobian values for bounded real parameters.

Evaluate log density, gradient, Hessian, and finite-difference checks for a Bernoulli-Beta Stan score-model proxy in unconstrained coordinates.

Extract a structured contract from simple Stan programs so score-model workflows can identify data, constrained parameters, transforms, model terms, and generated quantities.

Select and validate scalar diffusion factors for robust or heavy-tailed minimum Stein discrepancy estimators.

Minimise empirical Stein discrepancy objectives with bounded searches and auditable loss traces for recovery experiments.

Assemble executable soft-mode recovery evidence for Minimum Stein Discrepancy Estimator proxy experiments.

Compute one-dimensional kernel Stein and diffusion kernel Stein losses for unnormalized model recovery experiments.

Use this skill to score Particle Mirror Descent density approximations against reduced posterior targets with total variation, cross entropy, and symmetric mode coverage diagnostics.

Use this skill to run the weighted kernel-density Particle Mirror Descent prox update loop with stochastic likelihood factors, normalized weights, Gaussian kernels, and trace logging.

Use this skill to construct the synthetic tied Gaussian mixture posterior target from the Particle Mirror Descent paper, including data generation, log prior/likelihood functions, and grid posterior evaluation.

Use this skill to compose PMD mixture protocol, KDE update, and density metrics modules into an executable bounded soft-mode recovery experiment with auditable evidence.

Assemble executable posteriordb reduced benchmark recovery runs with provenance, metric, and mechanism evidence.

Compute posterior moment accuracy metrics for approximate draws against reference posterior summaries.

Validate and normalize posteriordb posterior, model, data, and reference object contracts for benchmarking workflows.

Load posteriordb reference posterior summaries and validate numeric quality fields for benchmark targets.

Run a bounded 1D Gaussian-mixture KSD-U recovery experiment for the Kernelized Stein Discrepancy paper.

Calibrate a KSD U-statistic goodness-of-fit test with the centered multinomial bootstrap for degenerate U-statistics.

Compute Kernelized Stein Discrepancy pairwise Stein kernels and U/V-statistic estimates from samples, model scores, and an RBF kernel.

Generated robust stochastic VI module skill for arXiv 2009.00666 recovery.

Generated robust stochastic VI module skill for arXiv 2009.00666 recovery.

Generated robust stochastic VI module skill for arXiv 2009.00666 recovery.

Run bounded constant-step noisy stochastic optimization chains that model stochastic VI iterates as a Markov process around an optimum.

Run a compact AEVB/VAE training step with encoder statistics, reparameterization, decoder reconstruction, analytic KL, and optimizer evidence.

Create deterministic tiny image batches for bounded VAE/AEVB recovery experiments when full image datasets are unavailable or too expensive.

Convert VAE/AEVB proxy training traces into auditable recovery_result.json artifacts with target matching and mechanism checks.

Extract and normalize final answers from chain-of-thought outputs without scoring intermediate reasoning text.

Safely check and repair simple arithmetic equations inside chain-of-thought reasoning traces.

Build standard, chain-of-thought, and ablation few-shot prompts with auditable prompt structure metadata.

Run a bounded standard-versus-chain-of-thought proxy recovery experiment and emit auditable recovery artifacts.

Compute binary conditional NCE objectives and diagnose self-normalization failures in finite protocols.

Build finite conditional NCE protocols with noise distributions, adjusted scores, and paper counterexample fixtures.

Evaluate conditional NCE recovery with ratio errors, KL metrics, and explicit mechanism checks for proxy runs.

Optimize finite ranking NCE objectives and recover conditional distributions when partition functions vary by input.

Compute and test DeBERTa style disentangled relative attention terms for bounded recovery experiments without relying on the original repository.

Apply DeBERTa enhanced mask decoder scoring that injects absolute position evidence only at candidate decoding time.

Build and evaluate RACE style multiple choice candidate records for DeBERTa recovery experiments.

Orchestrate a source boundary clean DeBERTa reduced recovery harness that records executable evidence and mechanism checks.

Run deterministic SiFT style normalized perturbation and optimizer traces for DeBERTa reduced recovery evidence.

Extract GSM8K final answers, validate calculator annotations, and label candidate solutions by answer correctness.

Build diverse GSM8K candidate solution records and label them by final-answer agreement for verifier training.

Run a bounded GSM8K verifier recovery harness with source provenance, generated-skill invocation logs, and solve-rate metrics.

Select GSM8K answers by ranking verifier-scored candidate solutions or voting among top-ranked answers.

Train a deterministic lightweight verifier on correctness-labeled GSM8K candidate solutions and record loss traces.

Train a tiny residual sequence-energy model with the paper's conditional NCE objective on positive and fixed-LM negative continuations.

Estimate residual EBM partition terms and select text continuations by importance reweighting fixed-LM proposal samples.

Build and validate soft-mode reduced recovery artifacts for residual EBM text generation without reading the original repository.

Compute residual EBM sequence scores and normalized importance weights from fixed LM proposal log probabilities and scalar energies.

Compute direct-RLAIF rewards from one-to-ten score-token logits without training a separate reward model.

Run a bounded REINFORCE-style policy update with terminal RLAIF rewards, a value baseline, and KL regularization evidence.