From summarizing information to updating judgment

01 / A useful summary needs context beyond the document

After an earnings call, revenue, profit and management guidance can be accurately summarized. A researcher also asks: what changed relative to prior expectations? Which piece of information changes an existing view? Which conditions remain unverified?

The same statement can carry very different weight in different research contexts. An expansion plan raises questions about delivery capacity, financing, customer demand and future returns. Understanding it requires connecting new information with historical evidence, industry constraints and competing interpretations.

Context therefore includes the questions we are tracking, the views we already hold, the conditions those views depend on, and the evidence we still need to find. These relationships give source material its meaning.

02 / Research itself is a Context Loop

Research continues after a report is generated. A signal is compared with an existing view, leading to a question, a counterargument or a revision. Verification then creates the context for the next round of research.

The loop needs a record of how a view changed: what was knowable at the time, which assumptions were used, and which evidence prompted a revision. Keeping that history gives people and agents a place to resume their work.

03 / Three kinds of problem call for different evaluations

The first is execution. Calculating a valuation from a specified formula, organizing a table or completing a search can be assessed against the task’s requirements.

The second is financial discipline. Periods, currencies, units, metric definitions and sources must be consistent. Correct arithmetic can still lead to an incorrect conclusion when the underlying contexts are mixed.

The third is judgment. Which information deserves attention? What would support or overturn the current explanation? Evaluating these choices requires the research context available at the time, expert feedback and subsequent evidence.

Evaluation should consequently examine how the model, tools, workflow and information environment work together. Calling a tool is an execution skill. Knowing when to call it, which question it serves and how to interpret the result adds another requirement.

04 / Signal quality and time give context its coordinates

Useful information has both relevance and an effect on what we know. A company mention may add little to a research view, while a development in its supply chain, financing environment or regulation may change a central assumption.

Time changes the weight of information. Expectations before a call, new statements during it, questions afterward and later verification address different problems. A research system should connect these stages while preserving sources, timestamps and the conditions under which a judgment holds.

Attention helps locate questions worth investigating, but popularity is not evidence strength. We need to track both what people are discussing and the evidence supporting their views.

05 / Design agents around a state that keeps evolving

We are exploring a design in which research context accumulates across interactions. The interface lets people participate; persistent state retains facts, expectations, disagreements and unresolved questions.

In this design, scenario agents handle events such as earnings calls and research reports. State agents maintain evidence and versions around companies and industries. User agents connect retrieval, feedback and requested actions. Market agents review changes in consensus and attention to identify what needs further investigation.

Shared state connects these roles. New material should make clear what it adds, challenges or changes, and every update should leave a traceable reason. The long-term aim is useful context that can keep accumulating and serve different people and AI systems.

06 / Communities contribute judgment; markets bring new tests

A professional community contributes questions, missing conditions, corrections and disagreements as well as information. How people participate in research is itself part of the context.

Subsequent market developments provide feedback, but a price move alone cannot prove an explanation correct. We still need to revisit the evidence and its conditions, separating new facts from changing opinions and retrospective explanations.

Frontier’s ambition is to make this ongoing research process something people and AI can carry forward together: a shared understanding with a traceable history, a clearer view of the present, and a concrete question to investigate next.