Blog
Practical guides, measured experiments and architecture notes on research, agent skills and useful LLM context.
Start with a useful question
These articles explain a decision you can apply, show their supporting work, and state the limits of their evidence.
Controlled experiment
What Good Docs, llms Files and Indexes Actually Save: 256 Agent Runs
Decide what to put in an agent's context before investing in indexes. Compare cost, correctness and asset use across 256 runs on two small repositories with one model.
Read the article →Architecture whitepaper
State That Survives the Session
Design a short session handoff and a separate store for durable recall. This implementation account explains the file roles and states which benefits still need measurement.
Read the article →Recorded retrieval walkthrough
Semantic indexing: two legs and a fusion
See where keyword search, embeddings and rank fusion succeed or fail on eleven questions. Follow the retained hits and timings from one docset and one recorded run.
Read the article →Technical analysis
Markdown as the LLM Output Gateway
Choose where Markdown fits in an LLM interface. Walk through parsing, streaming, safe rendering and the limits that call for a structured data format.
Read the article →
All articles
Newest first. The archive includes experiments, implementation notes, technical reviews and announcements.
- Comparative Review — Case-Resolution Diagnosis Strategies
Case-diagnosis strategies scored blind on real Atlas support cases: skill knowledge leads on 244 cases, but the ranking flips with the rubric and the quality of the ground truth.
- Case Study — Keeping a Customer's Feature-Request Tracking in Sync with an Agent
Claude Code and Glean kept a customer's feature requests in sync across Aha!, Monday.com and Google Sheets with idempotent writes, no deleted customer data and no pasted tokens.
- Choosing a local LLM for a 16 GB RTX 5080, part 2: the winner and how we serve it
Qwen3.8-27B with its built-in speculative-decoding head won coding and research on one 16 GB card at about 90 tokens per second. Here is the evidence, the launch command, a correction to part 1, and the harness scripts.
- Choosing a local LLM for a 16 GB RTX 5080, part 1: what the bake-off shows so far
Ten open-weight models, six inference engines and one 16 GB GPU, tested on real coding and research work. Two models lead, neither wins outright, and several results flipped once we stopped trusting our own first-pass metrics.
- Case Study — Gamifying Incident-Response Training with a Rate-Limiting Fire Drill
A training case study of an Atlas API rate-limiting fire drill: how its scorecard rewards restraint, and why team support multiplies, but cannot replace, individual expertise.
- TSE Strategy Backtest Scoreboard — Project Briefing
A first-person briefing on a git-backed scoreboard that backtests case-diagnosis strategies blind against resolved support cases, with how to read its scores and add a strategy.
- MDB Context Hub — Project Briefing
A project briefing on the MDB Context Hub, a local-first MCP server for TAM skills, prompts and catalogs, covering its features, architecture, security model and setup.
- Customer Dashboard — Project Briefing
A briefing on the Customer Dashboard, a local-first Chrome extension that gathers a TAM's account context: its features, scope, security model and architecture.
- A Real RTX5080 Response at 59 Tokens/sec—and a Failed Numerical Gate
One matched CPU and eGPU request, preserved failure receipts, and a physical F32 prefill trial that remained numerically outside the original tolerance.
- Getting local models to do real work: MLX, eGPU coding, and research failures
The complete investigation so far: useful Gemma MLX research phases, RTX 5080 coding sessions, failed standard research, context measurements, and the Qwen3.6 numerical discrepancy.
- The Deep Document Optimizer — A Live Run
A trimmed Claude Code transcript of /ddo cleaning a corrupted Slack export in one iteration, from contract to close-out report, with a rollback command for its one judgment call.
- 54 tokens/sec and real coding tools: what a local eGPU agent can do
Two complete Qwen3.5-27B coding sessions on an RTX 5080, the failures that preceded them, and why standard research still needs more work.
- Getting Gemma 4 MLX to do real research on a Mac
A measured local-agent case study: failed candidates, real tool calls, a corrected research worker, a ten-claim verification gate, and the limits of the Gemma 31B result.
- Local Model Performance: An Unverified Reported MLX, RTX 5080 eGPU and Ollama Comparison
An unverified MLX, RTX 5080 eGPU and Ollama comparison with withdrawn conclusions, preserved transcripts, and documented benchmark-script limits.
- Running Ollama on an RTX 5080 eGPU via Thunderbolt 5 on Apple Silicon M5 Max
A partial RTX 5080 compute guide for Apple Silicon, covering recorded failures, tinygrad prerequisites, corrected KV-cache math and compatibility limits.
- Testing speculative decoding on Apple Silicon and an RTX 5080
A working native MLX verifier, a cross-device draft adapter, and the tests that rejected the configuration: low acceptance, slower output and a reproducible token-parity failure.
- Every Token-Saving Strategy in My Stack, With Sources and Numbers
A dated inventory of token-saving techniques in my stack, with implementation sources, recorded savings and limits on the private benchmarks.
- What Good Docs, llms Files and Indexes Actually Save: 256 Agent Runs
In 256 runs on two small repos, loaded docs reduced lookup cost; llms files and indexes showed no measurable savings, with limits on model and interval reproducibility.
- An RTX 5080 over Thunderbolt on a Linux NUC: the tunnel the kernel threw away
A recorded RTX 5080 troubleshooting session on a Linux NUC, separating the working kernel profile from unverified causality and reboot guarantees.
- Fallen Off the Bus: the Resets That Failed, the Manuals, and the Concept Tree
A second investigator’s RTX 5080 eGPU recovery account: failed resets, bridge memory decoding, ASUS specifications, and unresolved incident evidence.
- A Closed-Loop System for Autonomous Skill-Knowledge Acquisition
A technical report on how concept-family-explorer, the concept tree, and /dr compose into a self-expanding expertise pipeline, demonstrated with real TypeScript/Node.js/MongoDB build runs.
- Semantic Skill Discovery and the Optimizer Family
A technical paper on mdb-context-hub's relevance-ranked discovery, role-based auto-skills, peer-deferral seeding, and token-budget defenses, closing with a comparison of the sko/pdo/ddo/cdo optimizer family.
- Building Codebases for Machine Collaborators
A technical review arguing LLM agents need codebases built for legibility and verifiability, showing mdb-tam's generated docs, agent-facing contracts, indexes, logging, tests, and dual-moded surfaces.
- Automated Auditing and Security-First Software Design
A whitepaper on pairing security-first design with automated CI/CD auditing, using a candid case study of advisory-not-blocking dependency scanning to argue a staged, ratcheted adoption path.
- Legible by Construction — Automated Docs, Indexes, Logging, and Test-Centric Design
A practitioner review arguing automated docs, doc-as-contract conventions, retrieval indexes, structured logging, and test-centric dual-moded design make a codebase legible/verifiable, grounded in mdb-context-hub.
- Markdown as the LLM Output Gateway
A technical review of Markdown as an LLM presentation layer, with sources for parsing, streaming, security and the limits of structured-data conversion.
- Hub-and-spoke indexes
How saved llms.txt exports split into nested indexes, what the generator preserves, and which historical lint results remain unverified.
- State That Survives the Session
The June mdb-tam memory design separates session handoffs from indexed recall, with limits on persistence, mirroring, and measured benefit.
- Implementing a Native macOS Meeting-Intelligence System
A proposed macOS meeting pipeline using Core Audio taps and on-device transcription, with capture permissions, lifecycle handling, and reviewed analysis.
- The lint that gates the estate
The August 2026 lint run reported 0 High findings across 15 docsets; source checks explain its current thresholds and limits on credentials, rights, and steering.
- Keyword plus vector: the cheap path
FTS5 adds token-phrase search without new embedding calls, while reciprocal-rank fusion combines keyword and vector rankings without independently verifying facts.
- Reducing LLM Cost and Latency Without Losing Context
A June 2026 review of mdb-tam’s context reduction and Anthropic caching controls explains partial telemetry and the unmeasured cost, latency, and quality tradeoffs.
- Anchors that point nowhere
How raw-mirror heading checks corrected recorded anchor failures, and why those checks do not prove that every link works on the live rendered site.
- Comparing Output Quality Across Claude Model Tiers and Effort Levels
A reported June Claude comparison with one response per condition; raw outputs are missing, and its score differences do not establish causal effects.
- Two thousand artifacts, three taxonomies: crawling a customer engagement
Three anonymized engagement crawls report 2,011 artifacts and 416 resolved Drive stubs, while exposing context-selection and confidentiality-check limits.
- A third axis for repo dossiers: every command, statically read
How a static executable inventory extends a repo dossier, illustrated by recorded private-run counts and three reported documentation discrepancies.
- Six months of hand-made llms files
A recorded survey of 608 downloaded llms files and the hub’s V1-to-V2 pilot, with sourced adoption evidence and limits on retrieval results.
- StickySites — Project Briefing
A June 2026 briefing of StickySites v1.10.0 covers its six note types, Chrome page access, local storage, and opt-in AES-256-GCM encryption.
- Abstracting one concept out of many docsets
Two August 2026 /lca evaluations build indexing and prompt-caching packs with lexicons and local embeddings, reporting task grades, runtime tradeoffs, and later cleanup.
- MDB Case Assistant — Project Pitch
A dated pitch for MDB Case Assistant, covering case triage, incident drills, operator-triggered enrichment and the local integration trust boundary.
- A topical llms file from a pool of facts
The August topical llms.txt pilot: keyword and file-affinity scoring, optional embeddings, assignment counts, and unresolved historical evidence.
- Orientation — Worked Examples: The Tooling Explains Itself
Two historical chat transcripts explaining the optimizer family and skill-guided development, with their recorded claims preserved as conversation.
- Turning a customer's docs into an llms family
An August docset acquisition and export case study, with stored llms.txt families, dated figure snapshots, and the limits of their verification.
- Conceptual vs proprietary llms files
A proposed governance design for conceptual llms files compares scoped claims, retains disagreements, and distinguishes source provenance from proof of correctness.
- Semantic indexing: two legs and a fusion
How keyword and vector retrieval differ, how optional rank fusion combines them and what one recorded eleven-question run can establish.
- The vocabulary file
How llms-vocabulary.txt defines family terms, labels model-written definitions and supports aliases, with a clear boundary between shipped behavior and proposed senses.
- V2 vs V1
Compare the llms.txt proposal’s August 2026 v2 changes with the hub pipeline’s V2 artifacts, migration steps, and limits of the compatibility assessment.