The Nine-Stage Mind

Setup

Everyone shopping for intelligence is shopping for parameters. Bigger model, higher bill, better answer — that is the assumption the whole market runs on. I built the opposite bet, and I run it every day.

Gizmo’s deep-research query does not hit one model and wait. It runs nine stages before the language model says a word. A 27-billion-parameter model sits at the center. Not a trillion. Twenty-seven billion, local, on my own hardware, pointed at my own knowledge store — a note-based store, no vector database, no embeddings pipeline underneath it. Every stage is engineering I can name and change. None of it is a weight I have to pray to.

Observation

Nine stages, in order, each doing exactly one job.

Decomposition splits the query into three to five diverse sub-queries plus a set of themes — an on-device pass reads intent and plans the angles before anything searches. Parallel search fires all of them at once against the unified store, pulling overlapping result sets instead of one ranked list. Five-factor reranking scores every candidate: brief relevance 40%, sub-query relevance 25%, recency 15%, coverage bonus 10%, title match 10%. Weights I chose, with a rationale — not weights a training run guessed. Diversity enforcement runs MMR, maximum marginal relevance, so the final set does not cluster on one angle; under-represented sub-queries get boosted. AI quality review is a second pass that flags candidates that are not genuinely relevant, with a safety net: if fewer than two notes would survive the filter, it restores all of them rather than starve the answer. Contextual synthesis is a third session that reads the survivors and streams a report, scaling its scope down instead of failing when the notes exceed model capacity. Human curation stops the machine — I read the report and the ranked note cards, and nothing proceeds without my approval. Targeted summarization writes a per-note reason it matters. Structured packaging bundles the whole thing for progressive disclosure: overview, then details, then full sources.

Nine stages. One 27B model doing the language work at the very end.

Here is what I live with, not what I benchmarked. On tasks grounded in my own knowledge, this pipeline around a 27B model routinely beats a trillion-parameter model answering alone. No eval suite, no score — daily use, my tasks, my data. The big model knows more about the world. It knows nothing about mine, and it has no pipeline deciding what matters before it speaks.

Take

By the time the model synthesizes, it is not reasoning. It is narrating. The reasoning already happened — across decomposition, reranking, diversity, review. The model is the voice. The pipeline is the mind.

That inversion is the whole position. A ten-percent gain in decomposition does not stay in stage one. It propagates through the eight stages downstream, because every later stage works on what the earlier one handed it. Sharpen the front of the chain and the whole chain gets sharper. Architecture compounds. Parameter count just sits there being large, and you pay for the largeness on every single call.

And the stages are swappable. When Mamba-3, RWKV-7, or Apple’s on-device Foundation Models are the right tool for a stage, they drop in without a redesign. The pipeline outlives every model that passes through it.

Stop buying bigger amps. Build the signal chain. One pipeline, nine stages, any model.