Eight original technical reports from graphwiz.ai — LLM evaluation, neural architectures, distributed systems, VR interaction design, and alignment research.
Eight original technical reports written and published by graphwiz.ai — the long-form research that backs the articles on this site, collected as print-ready PDFs with full references.
These are our own reports, not reprints: each one takes a question the articles could only summarise and works it through in depth, with methodology, data, and a bibliography you can follow to the primary sources.
LLM Evaluation & Benchmarking — Why static benchmarks saturate, how contamination inflates scores, and what dynamic evaluation protocols measure instead. Covers the eval loop from dataset construction to human-judgement correlation.
Advanced Mathematical Frameworks — The linear algebra and probability foundations that modern ML systems actually use: embedding geometry, attention as matrix operations, sampling strategies, and where the textbook version diverges from the shipped version.
VR Affordances & Interaction Design — How interface affordances transfer (and fail to transfer) between flat screens and head-mounted displays. Input precision, spatial memory, and motion-sickness-safe locomotion patterns.
AI Alignment & Safety — A survey of alignment techniques from RLHF through constitutional methods to interpretability research, with the failure modes each technique addresses and the ones it introduces.
Probabilistic Programming — Building models where uncertainty is a first-class citizen. Graphical models, inference algorithms, and the practical gap between PPL syntax and what the sampler actually does.
Neural Network Architectures — The architecture zoo with selection criteria: transformers, state-space models, mixture-of-experts, and retrieval-augmented designs — organised by the constraints (latency, memory, data volume) that should drive the choice.
Distributed Systems Theory — Consensus, consistency models, clocks, and failure detectors — the theory that makes coordination services work, explained through the systems that implement it.
Human-Computer Interaction — Research methods for evaluating interfaces: study design, effect sizes, and the replication problems the field is working through.
The core problem with static evaluation is not that models memorise answers — it is that memorisation is indistinguishable from competence on any benchmark whose contents have leaked into training data. A model can score near-perfectly on a benchmark it has effectively seen while failing the same tasks rephrased. Dynamic protocols that regenerate test instances per round trade a small amount of comparability for a large amount of validity.
Failure detectors are the quiet foundation of every consensus protocol. The FLP result tells us consensus is impossible in a fully asynchronous system with even one faulty process — so practical systems weaken the model. A detector that is eventually perfect (accurate after some unknown time) is enough to build a protocol that terminates in practice, even though no protocol can guarantee termination in the worst case.
Each report stands alone, but they reference each other: the evaluation report presupposes material from the architectures report; the alignment survey cites the probabilistic programming chapter on uncertainty. The bundle costs less than three individual reports.
The reports extend the site's article coverage — start with the free versions if you want the short treatment first:
$29.00
Are these peer-reviewed papers? No — they are original technical reports by graphwiz.ai, written for practitioners. Each cites peer-reviewed primary sources, so you can trace every claim to its origin.
What citation style do they use? Numbered references with a full bibliography, suitable for following up primary sources.
Are the reports updated? Reports are revised when the underlying research moves. Buyers get re-download access to new revisions through the library page.