LQM vs. LLM: Verifiable Numbers Over Empty Rhetoric

Predicting a lifespan or simulating an interaction requires more than just a well-crafted response. LQM (Quantitative Models) claim to be models for calculating and estimating with margins of error and domains of validity. The example of SandboxAQ's integrations into Claude illustrates a possible collaboration with LLMs (Large Language Models), without negating the necessity to compare performances on identical bases.

Comparing on Equal Terms and Maintaining Traceability

In October 2024, SandboxAQ announced a 95% reduction in the time required for battery lifespan prediction during an experiment with NOVONIX. This corporate statement does not prove that LQM consistently reduces the duration of all industrial projects by 95%, nor that they always outperform proven statistical models. To evaluate a solution, it is essential to analyze the test sets, the reference model, the amount of data required, the accuracy achieved, the cost of simulations, and the reproducibility. A decrease in computation time does not imply that the progress of research follows the same trend. Valid comparisons pit systems performing the same task under the same constraints against each other. In this perspective, the traceability of calculations and data takes precedence over any alleged universal superiority.

Accessing Calculators from an LLM Without Confusing Roles

On May 18, 2026, SandboxAQ announced the integration of some of its quantitative models into Claude via the MCP protocol. When a user poses a question in natural language, Claude has the ability to request a specialized scientific tool and then provide an explanation. The LLM serves as an interface while the quantitative engine handles the task in its own domain. On August 19, 2026, the general availability of AQCat on Claude was announced for research on catalysts. However, the Anthropic connector's documentation requires a connection and lists distinct tools, which does not imply free access to the entire SandboxAQ offering. This integration highlights a distribution of functions rather than a combination of technologies, and may allow researchers to use these tools without manipulating each technical interface separately. This ease of use does not replace experimental verifications or a critical analysis of results, even if an official demonstration shows its use on the provider's side.

From Verifiable Prediction to Industrial Tool

Building an actionable prediction begins with usable observations derived from laboratory measurements, sensors, operational histories, or simulations. Units, missing values, and representativeness must be controlled, which gives data engineers a key role, as inadequate preparation can compromise the entire process. On the modeling side, neural networks, regressions, physical simulators, or hybrid approaches are employed, some incorporating known constraints such as energy conservation. The number of parameters alone is not sufficient to judge reliability. Predictions must then face independent results, evaluating errors, stability, and conditions of irrelevance. An estimate is only actionable if its domain of validity is established, a step often less visible than conversational demonstration but crucial for transitioning from prototype to industrial tool.

Batteries, Molecules, Risks: What the Numbers Allow

For lithium-ion batteries, a quantitative model must predict degradation based on indicators such as voltages, internal resistances, or charge cycles. SandboxAQ has presented a method aimed at anticipating end-of-life and indicates an average absolute error of 150 cycles on manufacturers not included in the training, a performance specific to its protocol that does not constitute a universal guarantee. In drug discovery, simulations and learning are used to analyze interactions to guide experiments, while an LLM facilitates monitoring and query formulation. A predicted molecular property does not constitute proof of a treatment's efficacy or safety, which require biological and clinical trials. Quantitative models are also used to assess risks, anticipate demand, or optimize operations, applications that predate the emergence of the term LQM. Their performance depends on the available data and their ability to remain reliable in changing contexts: a financial model trained in stable times may fail during a crisis, and a validated prediction in one factory does not automatically apply to another site.

What LLMs and LQMs Can Do, and Their Own Limits

LLMs excel at understanding questions, leveraging documents, producing text, and operating tools, far beyond mere writing, even solving certain mathematical problems with the help of calculators. In contrast, LQMs target problems formulated in measurable quantities, with numerical, probabilistic, or simulated outputs, combining learning, physical equations, and numerical methods. However, LQM is not as clearly defined a category of architectures as LLM; SandboxAQ uses this label for its scientific models without establishing a novel technology per forecasting engine, and it does not necessarily involve quantum computing. LLMs can produce convincing but numerically inaccurate responses; well-crafted reasoning does not prove hypotheses or sources. Quantitative systems are evaluated based on their predictions, including in changing contexts, with error measurement and, if possible, uncertainty. The boundary is not impermeable: an LLM can call a reliable calculator, and a quantitative model can mislead if the data is biased.

Choosing the Architecture and Controlling Production Deployment

To summarize documents, produce a report, or query a knowledge base, an LLM is often appropriate. To predict a failure, estimate a risk, or simulate a molecule, one must evaluate a suitable quantitative engine, whether marketed as LQM or falling under classical statistics; the label never replaces the test of facts. The choice relies on data quality, acceptable error, and total cost, with requirements for reproducibility, access rights, and monitoring in production. An AI experimentation environment helps compare multiple approaches before commitment. Hybrid architectures combine natural language for collection and presentation with an auditable calculator for the numerical result and its limits, to be tested like any software, especially in health, safety, or finance. More broadly, the ambition of LQMs is to refocus AI on measurable properties without erasing biases or uncertainties; the NIST framework emphasizes continuous performance measurement and post-deployment monitoring. In the face of a precise claim, such as the announced 800 cycles without a verifiable method, the priority remains how the figure was obtained and validated, with the possibility, on the LLM side, of calling a specialized tool but without substituting for its validation.