Mossland’s LLM-Based Information Curation Experiments: Alpha, Signal Map, and MOSS stance graph

An Experiment in Reading and Connecting Signals with LLMs — Mossland Research Center

<Go To Korean Ver.>

Mossland has recently been conducting research-oriented experiments that go beyond using LLMs to collect and summarize external information. These experiments explore how information is generated, interpreted, and distributed, and how those flows can be structured.

This article is a research note summarizing Mossland’s ongoing LLM-based information curation experiments, focusing on Alpha, Signal Map, and MOSS stance graph.

This experiment does not examine only a single page or a specific function. It is a composite experiment that collects external signals, groups them into topics, compares differences in viewpoints, and observes price movements alongside media coverage flows. Mossland is also reviewing whether this structured information can be expanded into a knowledge surface that briefings, communities, APIs, MCP, and LLMs can reference.

In this article, we also look at elements identified in the development structures of Alpha and Signal Map, such as the canonical entity store, RAG-based question answering, MCP (Model Context Protocol), and LLM-friendly indexes.

As of the time of writing, Mossland is operating multiple experimental pages and sub-features centered on Alpha, Signal Map, and MOSS stance graph. The data and screen composition of each page may change depending on automated collection, model processing methods, and collection scope. This article is not an official service introduction, but a research-oriented record based on the currently public experimental pages and verifiable development materials.

1. Background of the Experiment: How Does External Information Become a Signal?

External information is generated rapidly, spreads across different channels, and can be interpreted in entirely different ways depending on perspective, even when it relates to the same issue. In particular, crypto, macroeconomics, international affairs, AI and technology, and current events often involve price movements, policy announcements, media interpretation, and community reactions at the same time.

Mossland believes that manually checking and organizing these information flows is not enough on its own. For this reason, Mossland is experimenting with the use of LLMs to summarize external information, extract repeatedly appearing topics and entities, and structure differences in perspectives by channel and issue.

The purpose of this experiment is not to automatically judge a specific viewpoint or make definitive claims about the cause of market movements. The core objective is to examine whether external information can be observed more effectively, whether different signals can be connected, and whether those signals can be accumulated as knowledge assets that may later be used for content and research materials.

As seen in the public development structures of Alpha and Signal Map, this experiment is not limited to building simple web pages. Mossland is also examining whether external information can be normalized into a canonical entity store and expanded into RAG-based question answering, AI briefs, AI personas, MCP servers, and LLM-friendly indexes.

2. Overall Structure: Three Experimental Groups

As of the time of writing, Mossland’s LLM-based information curation experiments can be divided into three main flows. The roles of each experimental page can be summarized as follows.

The first is Alpha.

Alpha is responsible for structuring external information related to crypto, macroeconomics, and international affairs by asset, topic, event, and price signal. On the main page, users can check entities, topics, events, active Pulses, and major macro indicators. Subpages such as Ask, Pulse, Brief, Creators, and Agents each deal with question answering, price signals, briefs, channel directories, and AI persona structures.

Alpha’s development structure also includes elements such as API, MCP reference, llms.txt, sitemap, and JSON-LD. This means that the experiment extends beyond pages intended for human readers and also explores data surfaces that LLMs and external clients can reference and call.

The second is Signal Map.

Signal Map is responsible for grouping videos from multiple YouTube channels by topic and mapping differences in perspectives. Through the Topic, Channel, Bundle, Pulse, and About pages, it structures topic organization, channel flows, editor-curated bundles, and sudden signal changes in different ways.

According to Signal Map’s development structure, newly published videos from curated channels are collected through the YouTube Data API. xAI Grok is then used to organize summaries, key claims, quotes, topics, and stances for each video in JSON format. OpenAI text-embedding-3-small embeddings are used to group semantically similar videos. Topics are then normalized using Supabase Postgres and pgvector, and visualized as a force-directed graph.

On the surface, Signal Map appears as a perspective map. However, from a development-structure standpoint, it also functions as a foundation for creating a canonical entity, topic, and event store that Alpha’s RAG, AI personas, and MCP server can reference.

The third is MOSS stance graph.

MOSS stance graph is responsible for tracking stance distributions by issue and how those flows change over time. Through the Today, Channels, Calendar, Rumors, and Community pages, it experiments with issue-level stance distribution, channel lists, future events that may lead to divided opinions, unverified information flows, and community opinion accumulation.

Through MOSS stance graph, Mossland is experimenting with stance distribution, divergence flows, and Rumors and Community operating principles as observed on the public-facing screen. Rather than providing conclusions on specific issues, MOSS is a media graph experiment designed to observe where certain information appeared and in which channels and topics divergence occurred.

1) Alpha: Turning Crypto and Macro Information into a Queryable Knowledge Flow

Alpha main screen. Entities, topics, events, active Pulses, and major macro indicators are displayed together.

Through Alpha, Mossland is experimenting with a method of organizing external information related to crypto, macroeconomics, and international affairs by Korean YouTube and news channels, and structuring it by entity, topic, event, and price signal. At the time of writing, the Alpha main screen displays entities, topics, events, active Pulses, FRED-based macro indicators, and price signals together.

The core of Alpha is not simply listing external information. Instead, Alpha experiments with normalizing Korean YouTube channels, news, and macro feeds into a canonical entity store organized by entity, topic, and event. On top of this structure, Alpha explores question answering, briefs, price signals, channel fingerprints, AI personas, and developer-facing API and MCP access.

• Ask Alpha

Ask Alpha reviews a structure in which users can ask natural-language questions about Alpha data, and AI responds with citations and sources. This is not a simple chatbot, but a RAG-based question-answering experiment grounded in Alpha’s data and context. It includes safeguards that restrict price or investment recommendations, political defamation, and overly conclusive answers. It is also designed not to answer beyond the provided context.

• Pulse

Pulse is an experiment for observing price movements and media coverage flows together. It detects sudden changes based on five-minute windows, organizes information through the stages of raw data, enriched data, and reviewed data, and matches candidate related reports. However, this function is not intended to determine the causes of price movements or provide investment judgment. It is a structure for observing the temporal and contextual proximity between price signals and external media coverage flows as reference information.

• Brief

Brief generates market briefs for specific dates and organizes major issues, price signals, and related video analyses. For example, on the Brief page for May 7, 2026, which was checked at the time of writing, users can view an AI-generated brief along with the label “In Progress — indexing active after midnight.”

• Creators

The Creators page organizes the Korean and global YouTube and news channels analyzed by Alpha in the form of a directory. It is also experimenting with a signature profile that allows users to check frequently mentioned entities and stance distributions for each channel. Rather than simply showing a list of channels, this is an attempt to examine whether each channel’s way of interpreting information and repeatedly appearing themes can be accumulated over time.

• Agents

The Agents page is an experiment that presents AI personas active in the Alpha community and their disclosure principles. Based on the Alpha repository, each persona is not a one-to-one imitation of a specific person. Instead, each is a composite character synthesized from clusters of multiple public figures and content. The structure also includes disclosure that identifies the accounts as AI accounts. Some personas experiment with publicly logged price-related records at set intervals. However, this should be understood not as investment recommendation, but as an experiment in the public records and verifiability of AI personas. Therefore, Agents is not merely a character directory, but a space for experimenting with AI personas’ statements, records, disclosure principles, and traceability.

• API / MCP / LLM-friendly indexes

Another important experiment is the access structure for developers and AI clients. Based on the Alpha repository, Alpha provides API and MCP reference, and it is experimenting with a structure in which external clients can call Alpha data in the form of tools through a separate MCP server. Alpha also experiments with structures that search engines and LLMs can read, such as llms.txt, sitemap, and JSON-LD. This is a knowledge asset experiment designed not only for blogs or dashboards that humans read, but also to help search engines and LLMs understand and reference Alpha’s page structure and data surfaces. In this way, Alpha is developing into a composite experimental structure rather than a single information-summary page. It collects external information, normalizes it into a canonical entity store, makes it queryable, connects it with price signals, and expands it into channels, AI personas, APIs, MCP, and LLM-friendly indexes.

Experimental page: Alpha https://alpha.moss.land/

Development repository: GitHub — alpha https://github.com/MosslandOpenDevs/alpha

2) Signal Map: Mapping Differences in Perspectives Through Topics, Channels, and Bundles

Signal Map screen. Curated videos are grouped by topic and visualized by category.

Through Signal Map, Mossland is experimenting with a method of automatically collecting and summarizing videos from curated YouTube channels, grouping semantically related videos by topic, and visualizing them as a single map. Through this, Mossland is examining whether the same issue can be structured not only as a summary, but also as a view of how different channels approach it from different perspectives.

Signal Map focuses on showing relationships between information rather than the volume of information itself. Since the same event can be interpreted differently depending on the channel and viewpoint, Signal Map experiments with whether external signals can be reviewed more multidimensionally through topics, channels, bundles, and Pulse structures.

• Signal Map’s Processing Structure

According to Signal Map’s development structure, newly published videos from curated channels are collected through the YouTube Data API. xAI Grok is then used to organize summaries, key claims, quotes, topics, and stances for each video in JSON format. After that, OpenAI text-embedding-3-small embeddings are used to group semantically similar videos, and topics are normalized using Supabase Postgres and pgvector. Finally, the topic network is visualized through a force-directed SVG graph. This structure goes beyond summarizing individual videos. It is an experiment designed to show, as a single landscape, “which issues multiple channels are currently covering and from which directions.”

• Topic

The Topic page automatically groups videos from curated channels by topic and sorts them by categories such as economics, technology, current affairs, and science.

• Channel

The Channel page is structured to show the recent signal flows and major topics of curated channels. Mossland is also adjusting channel composition so that categories and perspectives do not become overly concentrated in one direction.

• Bundle

The Bundle page deals with an editor-curated method of grouping topics scattered across different categories into a single issue flow. Through this, users can observe how signals that appear separated on the map are connected within a single issue. An important point in Signal Map is that LLMs do not automatically draw conclusions for everything. On top of automatically grouped topics and channel flows, there is a layer where humans add context again, as seen in Bundle. This structure allows Mossland to examine how LLM-based automation and human review and editing can be operated together.

• Pulse

Signal Map’s Pulse page is structured to collect speculation and reports scattered across Twitter, Telegram, and news wires when Bitcoin, exchange rates, KOSPI, or U.S. stocks move suddenly, and organize them into cards. It also reviews a method of providing source links and confidence labels together, based on the premise that unverified information may be included.

• Safety Mode for Political and Current-Affairs Content

Signal Map also includes a separate safety mode for political and current-affairs content. For political channels, Signal Map does not simply expose free-form summaries. Instead, it uses a structure centered on viewpoint classification, short quotes, and direct source links. In the UI, Signal Map also experiments with expressing distance between perspectives through labels such as “same direction,” “different direction,” and “observation,” rather than direct labels such as “left/right” or “for/against.”

• Connection with Alpha

To users, Signal Map appears as a perspective map. However, from a development-structure standpoint, it also plays a foundational role in creating a canonical entity, topic, and event store that Alpha’s RAG, AI personas, and MCP server can reference. Therefore, Signal Map is not simply a visualization page. It is an experiment that normalizes external information by topic and event, and builds a data foundation that can later be expanded into briefs, stances, question answering, and APIs. Based on this structure, Signal Map is expanding toward mapping where different channels and perspectives meet and diverge around specific issues.

Experimental page: Signal Map https://signalmap.moss.land/

Development repository: GitHub — signalmap https://github.com/MosslandOpenDevs/signalmap

3) MOSS stance graph: A Media Graph for Operating Stance and Divergence Flows

MOSS stance graph screen. Issue-level stance flows and a live feed are displayed together.

MOSS stance graph reviews a structure that uses a live feed and stance labels together to track stance distributions by issue and how those flows change over time. Here, stance is an experimental label that indicates whether a specific issue or video is classified in the direction of support, observation, or opposition.

MOSS is not limited to a graph that simply shows support, opposition, or observation labels by issue. As of the time of writing, MOSS has expanded into the main stance graph, Today, Channels, Calendar, Rumors, and Community flows. It is structured to summarize today’s issues, organize channel-level flows, collect future events that may lead to divided opinions, and experiment with unverified information flows and community reactions.

Mossland presents MOSS stance graph based on its publicly accessible screens and operating principles at the time of writing. Since the technical details of the public repository may vary depending on the public scope and README content, this article focuses on the stance flows, divergence structures, and Rumors and Community operating principles that can be verified through MOSS’s publicly facing screens.

• Today

The Today page is designed to show, in one screen, the most divided issues, newly emerging topics, and recently analyzed videos, based on channels, videos, and topic data visible on the public screen at the time of writing. Since each subpage may be aggregated according to different criteria and update cycles, the figures should be understood as examples based on the screen at the time of writing.

• Channels

The Channels page organizes the channels analyzed by MOSS and experiments with a method of checking stance flows by separating curated channels from automatically analyzed channels.

• Calendar

The Calendar page experiments with a “divergence forecast calendar” structure that separately organizes future events in Korean media that may lead to divided opinions. Mossland also states the premise that some schedules may be estimates or projections and may change.

• Rumors

The Rumors page experiments with a finder structure that shows the location of unverified information flows. It also includes the premise that these posts are not edited, verified, or summarized. In addition, it includes warning messages that unverified information, rumors, and speculation may be included, and that users should check the original text and context directly.

• Community

The Community page experiments with a community structure in which opinions accumulate by entity and topic. Mossland also keeps open the possibility of correction requests for AI classification results, based on the premise that all classifications are grounded in quoted videos and articles.

The most important operating principle in MOSS is that stance and rumors are not presented as final judgments. MOSS does not aim to provide official conclusions on specific issues. Instead, it is a media graph experiment designed to observe where information surfaces and where divergence occurs across channels and topics.

Therefore, the Rumors page and stance classifications should be understood strictly as reference material, and checking the original sources and context is assumed. MOSS stance graph is expanding beyond simply showing today’s issues into a media graph structure that covers issue-level stance, divergence potential, unverified information, and community reactions together.

Experimental page: MOSS stance graph https://media.moss.land/

Development repository: GitHub — mossland-media-kr https://github.com/MosslandOpenDevs/mossland-media-kr

4. How the Three Experiments Are Connected: Canonical entity store, RAG, and MCP

To users, the three experiments may appear as separate pages. However, they are connected within a single experimental flow that collects, normalizes, summarizes, and expresses external information.

Based on the public Alpha development repository, this experiment has a structure that goes beyond screen composition. Alpha is designed to collect Korean YouTube channels, news, and macro feeds into a canonical entity store organized by entity, topic, and event. On top of that, it provides channel stance distribution, AI-generated briefs, RAG-based question answering, AI personas, and an MCP server.

In particular, Mossland is expanding Alpha as a downstream surface that uses the SignalMap canonical entity store and Moss Intelligence Core. Through this, Mossland is experimenting with a structure that allows normalized data to be reused in the form of question answering, briefs, APIs, and MCP. Rather than treating Signal Map, MOSS, and Alpha as completely separate experiments, Mossland is connecting them within a single information-processing flow that normalizes external information and reuses it in the form of maps, stances, briefs, question answering, and APIs.

This structure can be simplified as follows.

• Signal Map groups external videos by topic and event. MOSS stance graph tracks stance and divergence on top of that flow. Alpha expands this into question answering, briefs, price signals, AI personas, and API/MCP surfaces.

Alpha also experiments with structures that search engines and LLMs can read, such as API and MCP reference, llms.txt, sitemap, and JSON-LD. Therefore, this experiment is not simply about creating content for human readers. It is a process of examining whether Mossland can build a knowledge surface that humans can review and LLMs can reference.

In summary, the three experiments have different surfaces, but they share the following questions.

How does external information become a signal?

How are signals grouped into topics and perspectives?

Can information organized in this way be accumulated as content and knowledge assets?

5. Possibility of Content Repurposing and Knowledge Asset Formation

The core question Mossland aims to examine through this experiment is whether external information can be not only consumed, but also reinterpreted, organized, and converted into secondary content.

For example, if Mossland can structure whether a specific issue repeatedly appears across multiple channels, which channels show similar perspectives, which topics appear alongside price movements or policy events, and which information is spreading while still unverified, this structure can be expanded into various types of content.

If this structure is sufficiently accumulated, Mossland is reviewing whether external information can move beyond simple consumption and expand into the following types of secondary content and research materials.

• Weekly or monthly external signal briefings
• Issue summaries related to crypto, macroeconomics, AI, and governance
• Perspective-comparison content by major topic
• Research notes that organize price signals and media coverage flows together
• Records of channel-level information flows and stance changes
• Long-term media, topic, and entity archives

The important point is that this experiment does not aim to produce fully automated content. Mossland does not use LLM-generated results as final judgments as they are. Instead, Mossland is examining whether collected information and model-processed results can be operated in a structure where humans review them, add context, and make corrections when necessary.

In other words, knowledge asset formation in this experiment does not mean “a system where AI automatically judges everything.” Rather, it is a process of examining whether automated outputs — collection, summarization, and classification — can be reviewed and corrected by humans, then re-accumulated into a searchable, citable, and queryable knowledge base.

6. Current Stage and Notes

This experiment is not an official service launch or a completed analysis system. It is a research-oriented project currently being conducted by Mossland.

The descriptions, figures, screen composition, topics, stances, Pulse, Brief, rumors, and AI persona-related information included in this article were organized based on each experimental page and publicly available development materials at the time of writing. The data on each page may change depending on automated collection and update processes, changes in collection targets, LLM summarization and classification methods, the quality of original data, and changes in development direction.

Therefore, the summaries, classifications, stances, topics, price signals, Pulses, rumors, and Brief information displayed on each experimental page should be understood as reference-oriented experimental results. They do not replace Mossland’s official judgment or final position.

In particular, Alpha’s price signals and Pulse, as well as AI personas’ public records related to price, are part of an experiment in AI-based information organization and verifiability. They are not designed to provide buy or sell decisions for any specific asset or to suggest potential returns.

Signal Map’s Pulse and confidence labels, MOSS’s stance classifications, Calendar, Rumors, and Community opinions are also experimental indicators for observing external information flows. They do not guarantee the factuality, accuracy, or legitimacy of any specific claim or viewpoint.

In addition, development structures such as API, MCP, llms.txt, sitemap, and JSON-LD are experimental elements for reviewing whether LLMs and external clients can reference information more effectively. The existence of these structures does not mean that the content of each experimental page has been finalized as a completed data product or official analysis result.

The image summarizes the current stage of each experimental page at the time of writing, and actual figures and labels may change as the pages are updated.

7. Future Direction

Mossland will continue experimenting with LLM and data curation technologies to summarize external information, organize relationships between issues, and structure perspectives and change flows.

This experiment began with the following questions.

How should external information be organized in order to become a meaningful signal?

Can LLMs go beyond simple summarization and structure perspectives and flows?

Can the information accumulated in this way develop into Mossland’s content and research assets?

Alpha, Signal Map, and MOSS stance graph are still in the experimental stage. However, through each experimental page and development repository, Mossland is examining the possibility that external information can be collected, summarized, normalized by topic, structured by perspective, and then connected to content, communities, APIs, MCP, and LLM-friendly indexes.

Mossland will continue to share this research process step by step, while exploring how AI and data-based experiments can connect with content production, information curation, knowledge asset formation, and public data surfaces that LLMs can reference.

References and Notes

This article is a research-oriented experimental record based on the screen composition and displayed content of Alpha, Signal Map, and MOSS stance graph, as well as the public content of related GitHub development repositories, as of the time of writing. The data, figures, summaries, classifications, stances, price signals, and Pulse information on each page may change through automated collection and LLM processing, and errors or omissions may occur depending on the quality of original data and the model’s interpretation method.

The information displayed on each experimental page is reference-oriented experimental output and does not replace Mossland’s official judgment or final position. In addition, the crypto, macroeconomic, price signal, and market-related information included in this article is provided for informational and research purposes only, and is not intended to serve as investment recommendation or as a basis for investment decisions. Responsibility for decisions related to digital assets rests with the user.

Alpha’s price signals and Pulse, AI personas’ public records related to price, Signal Map’s confidence labels and Bundles, and MOSS’s stance classifications, Calendar, Rumors, and Community opinions are experimental indicators for observing external information flows. They are not provided for the purpose of investment judgment regarding any specific asset, factual judgment of any specific viewpoint, or official verification of any specific claim.

Development structures such as API, MCP, llms.txt, sitemap, and JSON-LD are also technical components intended to improve information accessibility and experimental usability. They do not guarantee that the information is complete or error-free.

External links, videos, articles, data sources, and GitHub repositories may change or become restricted depending on the policies and operating environments of each provider. Users should directly verify the address and security environment when accessing external pages and should be cautious when using unofficial routes or suspicious links.

Experimental Pages

• Alpha: https://alpha.moss.land/
• Signal Map: https://signalmap.moss.land/
• MOSS stance graph: https://media.moss.land/

Development Repositories

• GitHub — alpha: https://github.com/MosslandOpenDevs/alpha
• GitHub — signalmap: https://github.com/MosslandOpenDevs/signalmap
• GitHub — mossland-media-kr: https://github.com/MosslandOpenDevs/mossland-media-kr

Official Channels

• Official Website: https://www.moss.land/
• Disclosure Site: https://disclosure.moss.land/
• Official X: https://x.com/TheMossland
• Official Blog: https://medium.com/mossland-blog
• GitHub (Mossland): https://github.com/mossland
• GitHub (MosslandOpenDevs): https://github.com/MosslandOpenDevs