Neelansh's Blog

Release prep for the retrieval saturation paper and a look back at six months of independent research across four projects.

Setting up local LLM concept extraction, scoping a new translation direction for DGG models, and building both papers to PDF.

First real-data results for concept-guided retrieval, FoxFlow output that finally parses, and two papers prepared for a public release.

The DGG agents complete their first cross-agent conversion, a fourth research project begins, and both papers get tighter claims.

Hardening passes across all three projects uncover a resampling default, a prompt-formatting bug, and a five-minute timeout.

Integration bugs between the DGG agents and a live transpiler: failures disguised as successes, mismatched embeddings, and a reserved word.

Connecting the DGG agents to a live FoxFlow transpiler, writing a deterministic emitter, and packaging two papers for release.

Hardening the DGG pipeline around a typed intermediate representation, scaling the Lost in the Middle dataset, and correcting the RAG paper's limitations.

Running Llama 3.2 3B on the Lost in the Middle tasks, and a clean re-run that settles the retrieval saturation numbers.

A reproducibility audit finds my main results table was assembled from incompatible runs, and the Lost in the Middle harness gets a local model.

Committing every output the paper relies on, auditing citations, and walking back claims that went beyond the evidence.

A mocked test suite for the DGG agents, a two-line citation fix, and a '100k' run that turned out to be mislabeled.

A self-audit retracts an unsupported stability number in the RAG paper, the Lost in the Middle paper gets a reviewer checklist, and the DGG agents can run on local models.

Adding OpenAI, Hugging Face and context-length sweep support to the Lost in the Middle harness, plus a harder semantic-noise dataset.

Fixing a mismatch between my paper's score-margin definition and its implementation, and adding related work to the Lost in the Middle paper.

A full draft of the Lost in the Middle paper, and an expanded retrieval saturation manuscript with related work and formal metric definitions.

Porting the retrieval saturation manuscript to LaTeX and outlining what a reproducibility section needs to cover.

The 30k-document run lands, the repo gets tests and a license, and the reproduction report gains an honest limitations section.

Kicking off an agentic system for converting Dynamical Graph Grammar models between representations, plus a lesson about writing results before the runs finish.

SciQ-aligned evaluation gives the RAG study its first real Recall@5, and I start drafting the manuscript.

Moving the RAG study onto a real, unique corpus (SciQ) and getting my first meaningful, if mismatched, signal.

Noise injection and longer contexts, and the realization that my toy corpus design is now the bottleneck.

Redundancy experiments on both tracks, and why a mitigation that cannot fail does not tell you much.

A three-scale RAG sweep that returned identical numbers, and an honest look at what my U-shaped curve actually proves.

Reproducing the 'Lost in the Middle' U-shaped performance curve and probing query stability in retrieval systems.

Kicking off two parallel research tracks: a RAG baseline and a position-controlled reproduction of long-context evaluations.

A quick note on why this blog exists and what I plan to write about.