Hosho Research Preview
How DTS works
Hosho proprietary tech
We measure the model's internal state, layer by layer
As a model processes your prompt, we probe every layer of the forward pass — where it is uncertain, what it is holding at once, how it groups your instructions, where the answer is still being assembled. Signals the model emits itself, not inferences from the wording.
In parallel
LLM analysis for what the probe can't see
Some demand is semantic — which facts a prompt assumes, how strict its output contract is. We read that with LLM calls, scored against demand rubrics from peer-reviewed evaluation research, then integrate with the probe signals.
The output
Four axes fundamentally drive demand
Hosho research shows four axes drive the demand a prompt places on a model — reasoning, format, parallel and knowledge. We score your prompt's load on each, and identify how that load can be restructured — enabling a lower-cost model by reducing the demand placed on it.
The match
Every model scored on those same four axes
Hosho maintains ability profiles for an extensive model library, built from public benchmarks and our own proprietary evals. Demand and ability in the same units is what lets us name the lowest-cost model that will actually carry your prompt — as it stands, or once restructured.
0.878 correlation with real model performance · measured across 12 model families
Breakdown appears after analysis.
Recommendations appear after analysis.
Structural improvements appear after analysis.
DTS reads a prompt and rates how hard it is to answer well, on four independent axes. Each axis catches a different way prompts go wrong — so a prompt can be easy on three and brutal on the fourth, and that’s the one that decides which model you need.
How many logical steps the model must work through to get the answer right. A lookup question (“what’s the capital of France?”) is low. A multi-step proof, a nested decision tree, or “compare these 5 options against 3 criteria” is high.
High reasoning prompts are where smaller models silently fail — they answer fluently but skip steps. The fluent surface fools you; the missing step is the bug.
How rigid the output structure has to be. “Write a paragraph” is low. “Return JSON with exactly these 7 fields, no markdown, ISO-8601 dates, never use the letter e” is high.
High format prompts are where models follow the spirit but break the letter — and that’s the kind of failure that breaks downstream code, not the kind that looks wrong to a human.
How much simultaneous bookkeeping the model has to do. “Summarise this paragraph” is low. “For each of these 12 customers, apply rules A through D and flag conflicts” is high.
The model has to carry every item plus every rule in working memory at the same time. Past a certain count, things start dropping silently — usually the items in the middle of the list.
How much specialised knowledge the answer relies on. Common-domain prompts are low. Specialised case law, recent papers, internal company terminology, or fast-moving fields (regulation, infosec, current events) are high.
High-knowledge prompts are where retrieval-augmented generation beats raw model size — feeding the right document in matters more than picking a bigger model.
Each axis runs 0–100. Anything above 50 on an axis means that axis is contributing real difficulty, not just being a number. The recommended model is picked so it can comfortably handle every axis you scored high on — not just the average.
That’s why a prompt with a moderate average can still get routed to a stronger model: one axis that’s out of range decides the pick, even when the others are easy.
The four axes measure what your prompt demands. To pick a model we need the other half: what each model can give. Both sit on the same 0–5 ruler, so a 3 means the same thing on either side.