HomeArticle

What intelligent agents lack is not knowledge, but the "causality ledger": the causal anchor engine builds neutral infrastructure based on the encoding of the subject's multi-modal social causality.

雅公子2026-11-20 17:35
Causal Anchor functions as a neutral social causal base to tackle the challenge of responsibility attribution for intelligent agents.

Over the past year, most of the industry's efforts have been focused on improving Agents' memory capabilities. Office collaboration platforms have successively connected Agents as independent collaborative entities into organizations, and a group of startups have followed up to develop shared memory, handover documents, and experience pools. The logic is all reasonable: an Agent with operational and cognitive capabilities cannot function without memory.

However, when an Agent actually takes over a task, the first question the organization asks is not "Did you remember this?" but "Why wasn't this task completed?"

Currently, no one can give a valid answer to this question.

1. Intelligent Agents Can Perform Tasks, But Cannot Present a Valid Set of "Reasons"

First, we have to acknowledge the progress made. Today's intelligent agents are already able to call tools, run loops, remember how they performed tasks in previous sessions across different conversations, and even continue to push tasks forward after human employees clock off. The capability ceiling in this regard is shared across the entire industry.

The problem lies on the other side: they can retell the process, but cannot produce a valid judgment.

The bottleneck is specifically reflected in four aspects.

The first aspect: they only output text, not the basis for judgment. What enterprises need is never a well-interpreted statement, but "for the same input, the same conclusion can be recalculated when the operator, deployment environment changes and the same rules are followed". What large language models provide is natural language, not "state + evidence pointer + recalculatable path". As a result, the review reports they write look smooth, but business teams dare not use them for decision-making, legal teams dare not endorse them, and audit teams dare not sign off on them.

The second aspect: they mistake correlation for causation, and the more fluent the output, the more confident they are. This is a structural problem of large language models: their answer to "why" is essentially a post-hoc narrative, which weaves events that appear with the same probability into a seemingly reasonable story. However, in social events, judging causation requires strict thresholds: there must be at least two pieces of evidence, covering at least two observation time points, and the subjective state can only be observed by the party concerned. Models do not have the concept of "threshold", they only judge by "plausibility", and the more plausible the narrative is, the more dangerous it becomes.

The third, and most critical aspect: they do not understand "obligation". The core of social causation is not "what happened", but "which link should do what, whether it has been completed, and who acknowledges it". Models understand facts, but do not understand obligations. Therefore, they can answer "this document has been circulated six times", but cannot answer "this document should have been completed at the third step, but the rights and responsibilities at the second step were not defined, so all subsequent operations are invalid". Most disputes in organizations do not arise at the fact level, but at the obligation level — exactly the area that models cannot access.

The fourth aspect: there is no verifiable carrier. All data is cleared once the session ends; memory products only store abstracts and pointers, which can reuse experience but cannot be used to define responsibilities. When a problem actually occurs, the Agent says "I thought...", but the organization cannot present evidence, and dare not assign responsibility to a system that cannot explain its own logic clearly.

The consequence of these four overlapping problems is straightforward: when intelligent agents evolve from "answering questions" to "taking over tasks", they must be able to account for "why" once they take over the tasks. If they cannot do that, they can only stay in the demonstration stage. This is not a capability problem, but a responsibility judgment problem. What responsibility judgment requires is not a more powerful model, but a underlying framework that can output "judgment + evidence + recalculatability".

Li Long, founder and product expert, put it very directly: intelligent agents are still tools to this day, because they have not truly been integrated into organizations. The sign of being integrated into an organization is not being able to answer questions, but saving time for people — instead of making people spend extra time sorting out right and wrong in the process. The difference between a "tool" and an "assistant" is that a tool only responds when you ask it questions, while an assistant takes over a section of the process and takes responsibility for that section.

To cross this threshold, we do not need to stack more technologies, methods, or causal data generated in laboratories. What we need first is a set of universally shared national standard underlying structure.

Without this structure, the industry will end up in a situation of "a thousand intelligent agents have a thousand Hamlets": for the same event, different vendors give their own judgments, users do not know who to trust, and institutions cannot mutually recognize each other's results. This fragmentation is not caused by model capability, but by the fact that every vendor builds its own causal standard. Only when the structure is unified, can we talk about making users dare to hand over a section of process to the agent.

This is even more critical for organizational processes: what people really need to record is not data, but every decision made during the execution process. Models can produce a well-organized analysis that wins unanimous praise after reading, but people will still hold meetings to check data as usual, and no one dares to make decisions based on the model's output. It answers "what the data relationship looks like", but does not answer "which set of judgments should be followed". These judgments are scattered in different places today, and there is the unwritten rule of "we have always calculated it this way". When N Agents read the database tables respectively, they can give N different answers for the same indicator, and no one can arbitrate when an error occurs.

These types of judgments are not managed by CausalAnchor, which does not judge standards, convert amounts, or verify authenticity. It manages another matter: in any scenario, when these judgments are used, where the basis is, and who has acknowledged them. The model can be replaced, but the process cannot be taken away; when a problem occurs, we should be able to trace it to the specific code position of the relevant rule, instead of a certain paragraph of prompt words.

Therefore, the entry point of CausalAnchor is exactly this layer: we do not build models, do not build memory libraries, do not draw conclusions, and only do one thing: build a "subject-native multimodal social causal foundation", which restores the social multimodal behavior traces of subjects into verifiable national standard causal codes in chronological order, forming a neutral public infrastructure for all subjects and applications.

2. Break "Why" Down Into Verifiable and Mutually Recognizable "Codes"

The prerequisite for this decomposition is to acknowledge that social causation has a fixed structure: whose demand it is, whose obligation it is, at which layer the process breaks, and who holds the evidence. These elements are in fixed positions in any organization, but no one has ever written them into the same set of fields before. Coding does not mean creating a new set of terms out of thin air, but attaching a standard number, item number and effective date to each code, and then upgrading to association standards, local standards and industry standards step by step. With standards to follow, cross-institutional mutual recognition becomes possible; if every party creates its own terms, they can never align with each other.

The project founder introduced that this initiative is supported by multidisciplinary integration, not a single technology. Economic science governs supply-demand relationship and cost, social science governs subjects and obligations, and data science governs structurization and recalculation. The three lines must be valid in the same set of codes at the same time. The team has carried out more than ten years of product R&D and market-oriented pilots in these directions, and has now completed 660 items in total: 296 for demands, 162 for industries and subjects, 62 for obligations, 59 for predicates, 53 for supply destinations, 15 for subjective states, and 13 for metadata. Each item must be attached with a standard number, item number and effective date, and retain three statuses: enabled, deprecated, and replaced.

The decomposition method has five levels. L1 is the subject, which refers to whose demand this event is; L2 is the binary judgment of demand and obligation, which is recorded through ten parallel dimensions, and the result falls into six mutually exclusive statuses; L3 asks "whether this layer is valid" layer by layer according to four levels: strategy, method, behavior and tool; L4 defines supply and destination; L5 completes verification.

Strict rules must be applied at L2. When the judgment is invalid, "comprehensive reason" is not allowed to be written, and the reason must be classified into one of the six categories: insufficient capability, shortage of resources, undefined rights and responsibilities, external interruption, insufficient evidence, and unmet time point. A breakpoint must also be defined, which is selected from seven categories: absence of awareness, missing link in path, missing link in supply, missing link in fulfillment, missing link in evidence, missing link in authorization, and missing link in attribution, and only one main breakpoint is recorded according to priority.

This rule determines whether the whole event is useful. For the same statement of "the process was stuck for 11 days", if it falls into "missing link in supply", supply needs to be supplemented; if it falls into "missing link in evidence", materials need to be supplemented; if it falls into "unmet time point", no action is needed. A wrong judgment will lead to all wrong rectification actions, wasting both money and time.

The product is clearly positioned at the core: it does not parse content, does not generate conclusions, does not build retrieval-based memory, does not replace runtime operations, does not perform accounting, does not judge authenticity, does not convert amounts, and does not exceed authorization. The project team's reasoning is that once this layer starts to "conveniently give conclusions", it will no longer be a public layer, but become another product that requires users to take sides. Therefore, the team deliberately avoids competing with the ecosystem of large manufacturers, and only focuses on the depth of causal relationships, building a neutral public infrastructure that all parties need. The reasoning is very practical: no neutrality, no mutual recognition. Memory can be flexible, but causation must be strict, and these two disciplines are opposite.

3. The Product Serves More Than Just Intelligent Agents

The three-piece product suite has a clear division of labor: CausalAnchor performs calculation, Yinluotong transmits data (only transmits codes instead of content), and Yincangxu stores data (only appends ledgers). Users can purchase the whole suite, or purchase individual components separately; the fields of individual deployment are completely consistent with those of the whole suite — mutual recognition depends on the code library and ledger format, not on purchasing all three components. The storage part is the least "product-like" but the most valuable part of the trinity framework: the ledger only supports appending. Each record uses "code, subject, observation time point" as the composite primary key. Correction does not overwrite historical data, but adds a new correction event in parallel. When a ledger is presented to two departments, all authorized personnel can query it, and no one can modify it secretly.

The reason why this project needs to be launched now is that to further improve the capability ceiling of foundational models, higher quality data is required. The corpus accumulated in the Internet era is almost exhausted, the Scaling Law is approaching its boundary, and the remaining increment is not stored in data warehouses, but in the causal networks of real scenarios — who put forward what demand, who should fulfill what obligation, at which step the process breaks, and where the evidence is. This part of data has hardly been structured so far.

Intelligent agents are the first scenario to hit this wall, but not the only one. All organizations that "must clarify the process" need the same set of standard coding system.

The second type of natural adaptation scenario is government affairs and public services: demand acceptance, dispatch and disposal, and completion archiving come with three built-in observation time points, which exactly match the structure required for causal judgment. The project team provided a reference baseline: the average handling time of people's livelihood demands in a certain sub-district in 2025 is 21.1 hours per item. What they want to do is not to reduce this number, but to break down the "average duration" into "which type of breakpoint each case is stuck at".

The third type of scenario is enterprise service procurement and performance management. After the contract is signed and the payment is made, the actual delivery volume of services is currently judged by acceptance documents and personal relationships; after being decomposed into the binary system of demand and obligation, every transaction can be traced to "which obligation is not fulfilled, and where the evidence is missing".

The fourth type of scenario is platforms and data operators. When multiple suppliers settle accounts, every supplier thinks they have done more work. With the append-only ledger and unified standards, disputes change from "each side argues its own case" to "check against the same line of record".

The project has clearly planned three types of ecosystems that share the same set of codes and the same ledger: the new-build type grows together with the new system; the plug-in type does not modify the existing system, and adds a parallel layer in bypass; the mount type is installed by the subject on demand. The same event remains the same event across different institutions. According to the project's 1.0-4.0 three-dimensional ecosystem plan, the fifth category will launch a plug-in warehouse: subjects install this layer in their own applications, hold the standard data by themselves, and still get the same set of codes and the same ledger; the project only provides the code library and ledger format, does not operate transactions, and does not engage in plug-in trading.

At the same time, the project founder stated that data governance is essentially a paradoxical post-hoc remedy. Based on standardized and structured algorithm mapping, data is not aggregated first and then governed, but coded according to the pre-defined structure from the moment it is generated, which is a prior agreement. Therefore, this project does not do "data governance" from the very beginning. The effect is reflected in two types of costs: transmission only moves codes instead of content, so when the same event is transmitted on the chain for ten times, the volume will not increase tenfold; storage is recorded according to codes, the ledger only supports appending, and repeated corpus will not be stored repeatedly. The larger the scale, the lower the unit transmission and storage cost. The same applies to the computing power side: the data obtained is structured causal data, not the original text that needs to be moved over and over again. The same data center can move forward by expanding dimensions, instead of stacking graphics cards and capacity, to meet the expansion and training requirements of large-scale, deep and massive dimensions.

4. Business Model

Advanced technology does not mean that users have to adopt it. Customers do not care how new this layer of technology is, they only care what problems it solves for them. In terms of business model: capabilities are not divided into different tiers, billing is based on the number of subjects, the volume of valid events, and the number of calls, not based on the quotation of the whole suite. The price is set at a small proportion (in the range of 5%-15%) of the total contract value of adjacent projects, to avoid becoming another project-based custom vendor. Delivery is mainly based on private deployment, with open interfaces for upper-layer applications.

In terms of industry positioning, there is currently no independent third-party supplier for this layer, and most of the related functions are split into memory products, data middle platforms or project customization: memory products manage "what happened", middle platforms manage "where the data is", and custom vendors manage "how to deliver this time". None of the three can produce facts that can be mutually recognized across institutions, and what is missing is a cross-domain mutually recognizable multimodal causal foundation.

The value of this product does not lie in the first judgment, but in the Nth judgment. The first judgment only clarifies one event; when the same type of breakpoint is repeatedly recorded as the same code across multiple institutions and multiple quarters, rectification can be based on evidence, procurement can be compared for price, and policy evaluation can have a valid denominator. Any conclusion can be traced back to the original event along the evidence pointer; any correction can only be newly added, not overwritten.

5. Progress and Plan

The MVP has been completed, and the current status is "the skeleton is complete, not open to the public". The internal test plan for 50 users will be launched in January 2027, stratified according to 5/15/30; three POC scenarios have been defined, and there is no public trial entrance yet. Multiple rounds of recalculation drills have been carried out internally and in small-scale markets, and the conclusion is "conditionally passed".

In terms of intellectual property rights, one invention patent has been authorized, which focuses on the digital intelligence of "human consciousness and behavior capability"; a full set of 12 invention patents have been completed and are pending submission; Level 3 Cybersecurity Protection Certification and 7 core algorithm filings have been planned (the first batch is planned to be submitted in January 2027). The project is also synchronously planning the landing areas for joint projects in 2027, to build a series of national key laboratories named "Causal Algorithm Mapping Innovation Experiment and Engineering R&D Center".

At present, the project team has completed the MVP and is looking for seed round investors, and has started contact with some urban governance scenarios in Guangzhou and Shenzhen; it is also recruiting "CTO VP" and "PR VP".