Research
AI and security research notes with methods, controls and measured limits. Proposed study means a plan; local implementation means working code; bounded pilot means a saved experiment with a stated scope.
Agents & assistants 5
- Swarm Sequence Generation: Neighbour Coupling TestsTest neighbour coupling in a small finite grammar, with baseline coverage and clear boundaries around untested communication and routing proposals.
- Multi Agent Orchestration: Streaming, Failures and ReplayTrace streamed output, cancellation and injected failures through a replayable orchestration fixture before evaluating real model answer quality.
- Mixture of Perspectives: Decision Support with DisagreementSeparate hard constraints from competing preferences and retain useful disagreement in a synthetic decision fixture with explicit utility assumptions.
- Local AI Memory: Retrieval and Temporal ValidityCompare current and historical assistant memory using ownership, validity and recording time, with a fixed word budget and bounded synthetic results.
- Autoresearch: Citation Tracking for Research AgentsTrack exact citations through revised sources without rewriting earlier answers, and distinguish citation preservation from evidence supporting a claim.
Knowledge & memory 4
- SL5: Thread Aware Retrieval and Quote AttributionPreserve nested quote attribution, source timestamps and thread context, while separating exact attribution tests from forecast scoring and retrieval quality.
- OpenAbstract: Versioned Claims and CorrectionsFollow a claim correction without erasing its history, separating replay correctness, support labels and the rules for the default reading view.
- Graph Based Research Memory with Source ProvenanceCompare graph and metadata retrieval on versioned sources, keeping source history visible and accounting for cases where the simpler baseline matches it.
- CloudVec: Vector Index Freshness with a SQLite OutboxFollow updates, deletion and crash recovery through a SQLite outbox, with freshness and replay checks kept separate from semantic search quality.
Mathematics 2
- Adaptive Range Summaries and Interval SearchSeparate multidimensional range summaries from sorted interval search, with exact overlap checks, scan baselines and construction costs.
- 4D Range Queries with Sparse Summary TreesCompare sparse summary trees, scanning and spatial search on synthetic 4D queries, including construction costs and distribution dependent results.
Security 3
- Defensive Security Research ProjectsBrowse defensive research ideas, local implementations and unresolved sources, with practical next checks for incident review and model evaluation.
- Synthetic Signal Detection Under Distribution ShiftCompare synthetic signal detection and false alarms under session shift, including a detector whose ranking score hides a failed operating threshold.
- Protocol Resource Limits and Amplification ModellingInspect byte, work, connection and time budgets in local protocol simulations, without treating bounded fixtures as evidence of live vulnerabilities.
Training & evaluation 3
- Self Rewarding Training: DPO and EvaluationCheck preference loss direction, answer masks and judge bias, and distinguish objective tests from a toy training run that failed to generalise.
- LLM Self Correction: Evaluation and StoppingEvaluate helpful and harmful answer revisions, frozen stopping rules and baseline competence before drawing conclusions about LLM self correction.
- Glow Worm: A Small Byte Level Language ModelA small byte level language model tested on WikiText 2, with bits per byte, a unigram baseline and an explicit account of training exposure.
Transformers 12
- SVD Compression: Storage and Functional ErrorCompare SVD factor storage with held out output error on a GPT Neo projection, and see why weight similarity alone cannot establish useful compression.
- Temporal Attention for Irregular ObservationsCompare learned time decay, timestamp features and a last observation baseline across synthetic switching, periodic and shifted arrival tasks.
- Adaptive Neural Architectures: Experiments and QuestionsExplore memory, routing, repair and compression experiments with a clear distinction between local implementations, measured results and proposals.
- Sparse Autoencoders: Reconstruction vs Feature RecoveryA synthetic sparse autoencoder comparison separates reconstruction error from feature recovery using cosine matching, support scores and correlated features.
- Self Healing Tokens: Selective Denoising ExperimentsCompare learned token repair with majority repair, including recovery of corrupt tokens and harm to clean inputs in a controlled synthetic task.
- Transformer Memory: Causal Retrieval and AdaptationTest delayed memory retrieval, episode reset and batch isolation, with explicit limits on comparisons between memory systems of different capacities.
- Transformer Activation Maps and Token TrajectoriesTrack individual token occurrences, model revisions and shared projection bases so activation maps remain comparable and their limits stay visible.
- Judge Head Attention: Contextual Head GatingCompare contextual attention head gating with standard attention and a token MLP control, separating initial failures from a longer recall comparison.
- Hydraform: Adaptive Attention Head MutationExamine attention head mutation, optimiser replacement and checkpoint recovery alongside fixed and random mutation controls on a synthetic task.
- Face Based Attention: Gated Feature MixingCompare gated feature mixing with dense and static controls, including parameter costs and results that do not establish an advantage for the gate.
- BitNet Style Quantisation: Storage vs ExecutionSeparate float conversion, ternary simulation, packed weight storage and measured execution time, including the limits of a scalar Python decoder.
- LLM Arithmetic: Digits vs Number WordsA protocol for comparing digit and number word arithmetic circuits, including the failed answer generation gate that blocks causal interpretation.