https://stoneytech.net/ Citation-first learning notes from StoneyTECH — AI Architecture Applied. Two series: Determinism Ladder (architectural, opinionated, for senior practitioners) and Demystify AI (primers for technical generalists). Public sources, reference builds, and no originality claim. en-us Sat, 23 May 2026 04:28:46 GMT stoney@stoneytech.net (StoneyTECH) stoney@stoneytech.net (StoneyTECH) https://stoneytech.net/learn/2026-05-17-graph-data-fabric https://stoneytech.net/learn/2026-05-17-graph-data-fabric Sun, 17 May 2026 12:00:00 GMT Graph-first architecture does not mean one database for everything. The semantic graph owns meaning while persistence categories earn their roles by workload. stoney@stoneytech.net (StoneyTECH)graphsdata-fabricpersistencearchitecturedeterminism-laddermcpseries:learn https://stoneytech.net/learn/2026-05-17-prompt-context-fine-tune-gate https://stoneytech.net/learn/2026-05-17-prompt-context-fine-tune-gate Sun, 17 May 2026 12:00:00 GMT The Determinism Ladder moves AI behavior from probability layers into authority layers as consequence rises. stoney@stoneytech.net (StoneyTECH)determinism-ladderpromptingragfine-tuningloraevalsgovernanceseries:learn https://stoneytech.net/demystify/2026-05-17-how-llms-are-built-and-where-lora-fits https://stoneytech.net/demystify/2026-05-17-how-llms-are-built-and-where-lora-fits Sun, 17 May 2026 12:00:00 GMT A language model moves through stages: pretraining, supervised tuning, preference tuning, evaluation, serving, retrieval, and adapter training. LoRA enters as a compact adaptation layer after the expensive base model exists. stoney@stoneytech.net (StoneyTECH)demystifyllmtraininglorafine-tuningrlhfragtransformersprimerseries:demystify https://stoneytech.net/learn/2026-05-11-deployment-context-first https://stoneytech.net/learn/2026-05-11-deployment-context-first Mon, 11 May 2026 12:00:00 GMT Deployment context comes before model choice. Three contexts, changing levers, and shippable architectures make axiom #18 concrete. stoney@stoneytech.net (StoneyTECH)deployment-contextarchitecturedeterminism-ladderaxiom-18data-residencyseries:learn https://stoneytech.net/demystify/2026-05-09-ai-ml-llm-agents-sorting-out-the-words https://stoneytech.net/demystify/2026-05-09-ai-ml-llm-agents-sorting-out-the-words Sat, 09 May 2026 12:00:00 GMT Four different words often collapse into one marketing pitch. A nested mental model makes the buying, building, and risk questions sharper. stoney@stoneytech.net (StoneyTECH)demystifyaimlllmagentsprimervocabularyseries:demystify https://stoneytech.net/learn/2026-05-06-local-graphs-first https://stoneytech.net/learn/2026-05-06-local-graphs-first Wed, 06 May 2026 12:00:00 GMT A public pattern repo should not begin with hosted graph gravity. File-backed graphs earn the first version because they stay inspectable, portable, and legible to agents. stoney@stoneytech.net (StoneyTECH)graphsmcptemplatesagentsdeterminism-ladderarchitectureseries:learn https://stoneytech.net/learn/2026-05-06-portable-agent-pattern-kits https://stoneytech.net/learn/2026-05-06-portable-agent-pattern-kits Wed, 06 May 2026 12:00:00 GMT A useful public agent repo should not ask for one blessed model or one hidden control plane. A reader should be able to bring a model, read the local graph and MCP, and get a bounded system working. stoney@stoneytech.net (StoneyTECH)agentsmcptemplatesgraphsdeterminism-ladderarchitectureseries:learn https://stoneytech.net/learn/2026-05-06-shadow-tribunals https://stoneytech.net/learn/2026-05-06-shadow-tribunals Wed, 06 May 2026 12:00:00 GMT A strong agent system does not need one louder voice. It needs a primary path, bounded shadow judges, and a clear rule for what disagreement can and cannot do. stoney@stoneytech.net (StoneyTECH)agentsgraphsmcpevaluationdeterminism-ladderarchitectureseries:learn https://stoneytech.net/learn/2026-05-05-three-repos-one-thesis https://stoneytech.net/learn/2026-05-05-three-repos-one-thesis Tue, 05 May 2026 12:00:00 GMT One thesis now lives in three codebases. Each repo pushes determinism into a different layer: loop boundary, evidence boundary, or graph boundary. stoney@stoneytech.net (StoneyTECH)agentsdeterminism-ladderproof-of-workanthropicopenailanggraphseries:learn https://stoneytech.net/learn/2026-05-05-three-sdks-three-jobs https://stoneytech.net/learn/2026-05-05-three-sdks-three-jobs Tue, 05 May 2026 12:00:00 GMT Three popular agent stacks solve three different jobs. The useful question is not which SDK wins. The useful question is which job sits on the desk. stoney@stoneytech.net (StoneyTECH)agentssdklanggraphopenaianthropicdeterminism-ladderarchitectureseries:learn https://stoneytech.net/demystify/2026-05-05-what-is-mcp https://stoneytech.net/demystify/2026-05-05-what-is-mcp Tue, 05 May 2026 12:00:00 GMT MCP is a standard way for an AI agent to ask another system for context or tools. Think less magic brain, more well-labeled port. stoney@stoneytech.net (StoneyTECH)demystifymcpagentstoolscontextprimerseries:demystify https://stoneytech.net/learn/2026-05-04-published-content-mcps https://stoneytech.net/learn/2026-05-04-published-content-mcps Mon, 04 May 2026 12:00:00 GMT A public MCP should not become a workspace wormhole. It should project intentionally published material through a contract-bound interface. stoney@stoneytech.net (StoneyTECH)mcppublic-contentcloudflaresecuritydeterminism-ladderaxiom-21series:learn https://stoneytech.net/learn/2026-05-04-threat-surface-layer-by-layer https://stoneytech.net/learn/2026-05-04-threat-surface-layer-by-layer Mon, 04 May 2026 12:00:00 GMT Threat surface belongs beside every agentic lever. Seven layers, entry paths, and mitigations make axiom #17 concrete. stoney@stoneytech.net (StoneyTECH)securityagenticdeterminism-ladderthreat-modelingaxiom-17series:learn https://stoneytech.net/learn/2026-05-03-graph-constrained-execution https://stoneytech.net/learn/2026-05-03-graph-constrained-execution Every agentic system has a graph. The real choice: draw it before the incident or reconstruct it from the postmortem at four in the morning. stoney@stoneytech.net (StoneyTECH)graphdagorchestrationconcurrencyintegrityaxiom-2axiom-4series:learn https://stoneytech.net/demystify/2026-05-03-tokens-context-attention-no-math https://stoneytech.net/demystify/2026-05-03-tokens-context-attention-no-math A working mental model for the path from prompt to returned text: tokens, context windows, and attention without a single equation. stoney@stoneytech.net (StoneyTECH)demystifyprimertokenscontext-windowattentionmodel-mechanicsseries:demystify https://stoneytech.net/demystify/2026-05-03-why-llms-hallucinate https://stoneytech.net/demystify/2026-05-03-why-llms-hallucinate Hallucination comes from the same retrieval looseness behind useful LLM answers, with a different consequence. stoney@stoneytech.net (StoneyTECH)demystifyprimerhallucinationreliabilitymental-modelsseries:demystify https://stoneytech.net/demystify/2026-05-02-llms-as-a-loose-database https://stoneytech.net/demystify/2026-05-02-llms-as-a-loose-database Sat, 02 May 2026 12:00:00 GMT A practical mental model for LLMs: word-based queries over learned patterns, refined with the looseness behind iteration, useful surprises, and confident wrongness. stoney@stoneytech.net (StoneyTECH)demystifyllmprimermental-modelsseries:demystify https://stoneytech.net/learn/2026-04-27-cheaper-alternatives-to-mcp https://stoneytech.net/learn/2026-04-27-cheaper-alternatives-to-mcp Mon, 27 Apr 2026 12:00:00 GMT MCP fits wide tool surfaces. For narrow surfaces, mature CLIs and single REST endpoints often win on cost, latency, and debuggability. The break-even point matters, along with the threat model under it. stoney@stoneytech.net (StoneyTECH)mcptoolsfunction-callingclideterminism-laddersecuritydeployment-contextseries:learn https://stoneytech.net/learn/2026-04-27-eighth-lever-eval-and-observability https://stoneytech.net/learn/2026-04-27-eighth-lever-eval-and-observability Mon, 27 Apr 2026 12:00:00 GMT The seven levers need a feedback loop. Evaluation and observability become the determinism ladder's load-bearing rung, plus the trace store creates a PII/PHI surface. stoney@stoneytech.net (StoneyTECH)agenticevaluationobservabilitydeterminism-laddersecuritydeployment-contextseries:learn https://stoneytech.net/learn/2026-04-27-lora-plus-rag-composition https://stoneytech.net/learn/2026-04-27-lora-plus-rag-composition Mon, 27 Apr 2026 12:00:00 GMT LoRA and RAG compose because they live at different layers: brand voice in weights, live facts in retrieval, plus composition-layer costs and threat model. stoney@stoneytech.net (StoneyTECH)loraragcompositionworked-exampledeterminism-laddersecuritydeployment-contextseries:learn https://stoneytech.net/learn/2026-04-27-model-portability-exceptions https://stoneytech.net/learn/2026-04-27-model-portability-exceptions Mon, 27 Apr 2026 12:00:00 GMT The inaugural piece said don't agonize over model choice early because most architectures are model-portable. True for most. Here are the cases where the model is the architecture — and skipping over them costs months. stoney@stoneytech.net (StoneyTECH)model-selectionregulated-industriesconstraintsdeterminism-laddersecuritydeployment-contextseries:learn https://stoneytech.net/learn/2026-04-26-the-stack-matrix https://stoneytech.net/learn/2026-04-26-the-stack-matrix Sun, 26 Apr 2026 12:00:00 GMT Each lever swaps model autonomy for determinism. The seven — Model, API, LoRA, RAG, Skills, MCP, Agents — sit in build order and reveal purchased capability. stoney@stoneytech.net (StoneyTECH)agenticarchitecturedecision-makingseries:learn