RESEARCH

Mid-query, token-based decisions

A Prizmal answer is written by several models, moved between them without loss, and governed token by token. Here is what we can show, and what we keep.

An agent's answer is thousands of tokens. Most are routine. A few carry the whole thing, and you cannot tell which from the outside. Roughly 85% of tokens are trivial (internal estimate across production token streams). Paying frontier price for all of them is the default. It should not be.
Fig. 01 · Token field. Most marks are routine, the tall ones carry the answer. Internal estimate across production token streams: roughly 85% of tokens are trivial. The last 15% is why quality exists.

The pieces

01 · The recipe
No single model is best at every span of an answer. Prizmal curates a small set, open weight and frontier, and composes each answer from them in measured proportions. Recipes are tuned per workload and retuned as models ship: when a better one appears, your mix updates and your code does not.
02 · Cheaper tokens that keep the thread

RASP makes open models radically cheaper per token without losing the reasoning. Training free: the principal structure stays dense, the residual is pruned.

Reasoning kept
Dense100%
RASP99%
Compute per token
Dense100%
RASP40%
Energy
Dense100%
RASP74.9%
99% of the reasoning for 40% of the compute. Dense is 1.00x; R/E lands at 3.51x. Measured on Llama-3 8B, Gemma-2 9B, DeepSeek-R1 7B; energy at device level.

These are 7B to 9B class results. The bet, and the patents, are about what happens above.

03 · The moment a token requires frontier
Token by token, the layer estimates whether the next span still belongs to the model writing it. When the work outgrows the writer, frontier steps in for exactly as long as it is required. How the estimate is made is not published.
04 · The move without loss

Mid-answer, the writing model can change. The full working context moves with it: every fact, every constraint, every open thread of the reasoning. Nothing is summarized away, nothing is re-asked, and the sentence lands as if one mind wrote it. In a routed answer the model can change between one word and the next, and the seam is not visible.

tok 0128256384512SignalToken traffic by typeEscalation thresholdarithmetic chainnovel domain shiftambiguous coreference
1 · arithmetic chain2 · novel domain shift3 · ambiguous coreference
Local open · 41%Large open · 36%X-large open · 12%Frontier sub · 5%Frontier flagship · 3%Specialized · 3%
One 512-token span of a routed answer, representative trace. The signal is real. What computes it is not published.

Moving a live context between two different models without loss is the hardest problem on this page, and the one we say the least about.

05 · The proof
Every decision the layer makes is logged: which model wrote which span, what it cost, which policy allowed it. The mix is invisible in the answer and fully visible in the audit.

What we publish and what we keep

We publish results, benchmark scores, recipe outcomes, and audit formats. We keep the estimator, the switch policy, and the transfer method.

Six patent families

FAMILY 01 · US PROVISIONAL · DEC 2025

Less compute, same answer

Identifies, per token, computation whose removal leaves the output unchanged at zero divergence. The removable set is knowable before the work is done.

FAMILY 02 · US PROVISIONAL · DEC 2025

How much model a token gets

A generating model rarely needs all of itself. This family sets, live, how much of the network each token engages. Deciding costs less than the compute it saves.

FAMILY 03 · US PROVISIONAL · DEC 2025

The recovery reflex

Aggressive efficiency requires a failure detector. When output confidence degrades, the compute envelope widens inside the same answer, then narrows once the risk passes.

FAMILY 04 · US PROVISIONAL · DEC 2025

Error, bounded by design

Approximation error compounds with depth. This family caps the error each layer may contribute, so the total stays engineered rather than emergent.

FAMILY 05 · US PROVISIONAL · DEC 2025

Speed on the metal

Sparsity pays only when hardware can skip what theory removed. This family turns pruned structure into kernels that keep their speedup on real accelerators.

FAMILY 06 · US 64/072,126

The move without loss

A live generation carries state: facts, constraints, the open threads of a reasoning. This family moves that state across models mid-answer, nothing summarized and nothing re-derived. The seam does not appear in the text.

Six USPTO patent families, the first filed December 2025. Legal titles publish with the filings.

Application serials on file include US 63/928,790, US 63/928,818, and US 64/072,126.

If you work on these problems, the interesting question is not which model to call. It is when a token stops being cheap, and how a thought survives a change of mind.

BenchAlign note

BenchAlign v5 is benchlm.ai's reasoning suite, scored 0 to 100. A routed Prizmal answer is scored the way any single model is: the suite sees answers, not the mix. Prizmal's figure is 84, July 2026.

The rest is in the audit trail.