It takes up three of the top 10 spots on the FutureX list.
In July, DigClaw’s forecasting framework Rhizome v1 secured three spots at #1, #3, #7 on the FutureX evaluation platform.
The three spots come from three different base models, including Kimi-K3, DeepSeek-V4-Pro, etc. The same set of frameworks enabled all of them to enter the Top 7 simultaneously, and DigClaw is the only participant that has achieved this.
There are 59 real-event forecasting questions covering the political, economic, and technological fields, with no possibility of training data leakage.
This set of results supports the judgment that DigClaw is verifying:
Forecasting capabilities can be precipitated outside the base model.
As the base model advances, the system can gain capability dividends; forecasting trajectories, settlement feedback, calibration experience, and the continuously evolving workflow are continuously precipitated inside the system.
This is not an accidental competition result, it is the first external verification that DigClaw has given to the question of “how forecasting should be done”.
Forecasting is the Most Underestimated Capability of AI
As forecasting tasks gradually move towards standardization and engineering, AI is also expected to handle complex forecasting problems at scale.
But there is a fundamental problem: large language models are inherently not good at forecasting.
What LLMs learn is correlation, not causation. They extract patterns from past corpora, but “learning the past” and “forecasting the future” are two essentially different things.
This brings three fatal problems:
Blind Spot in Causal Direction. The model knows that A and B often appear together, but does not know whether A causes B, B causes A, or there is a common cause C.
Failure of Intervention Reasoning. You cannot ask the model “what impact will the Fed’s interest rate cut have on Southeast Asian tech stocks” — because it has only learned historical co-occurrence, and there is no causal graph to support intervention calculation.
Lack of Calibration. The “70% probability” output by the model has not been calibrated in any probabilistic sense, and is essentially just a by-product of the token distribution.
So far, existing solutions have their own limitations: human collective intelligence (prediction markets) requires liquidity, and prices are not credible for niche problems;
LLM pattern matching has no causal structure; end-to-end training has outcome-oriented bias — a reasoning process with strict logic but “wrong answer” will be punished, while a rough judgment that happens to “guess correctly” will be reinforced instead.
This is exactly why DigClaw exists.
DigClaw builds a forecasting infrastructure with causal structure as the skeleton, probability calculation as the engine, and search intelligence as the data pipeline.
The core assumption is: forecasting should not be completed by a single model end-to-end.
Search, causal reasoning, and probabilistic inference are three orthogonal capabilities that should be solved by three dedicated systems separately and then combined structurally.
Three Top 10 Spots Verify the Cross-Base-Model Migration of Forecasting Capabilities
FutureX is currently the most challenging real-time forecasting leaderboard: real-world event forecasting questions are released every week, the standard answers have not yet been generated when forecasts are submitted, and settlement is conducted afterwards.
The dataset is hosted on HuggingFace, the evaluation framework is open
Rhizome adopts the same set of forecasting frameworks and operating conditions on three base models, independently generates and submits answers respectively, and no cross-model aggregation is performed.
Therefore, these three rankings are the separate operation of the same systematic method on the three base models, that is, a clear cross-base-model control experiment.
But this does not mean that the base model is unimportant.
The base model still provides general capabilities such as language understanding, reasoning, and tool use.
The result of this leaderboard illustrates that: how to organize retrieval, process time, express probabilities, maintain evidence, and control long-running operations can form system capabilities independent of model weights.
The following is the technical report of the Rhizome framework of DigClaw on FutureX:
Rhizome Technical Report
The design of Rhizome revolves around three engineering judgments: search and reasoning must be decoupled, the forecasting trajectory must be completely retained and form calibration assets, and probability updates must be aware of the causal structure:
Design Decision 1: Decoupling Search and Reasoning with Multiple Models
The core design insight of Rhizome: search quality and reasoning quality are two orthogonal problems that should not be optimized simultaneously by the same model.
Current mainstream DeepResearch agents (Perplexity, Gemini Deep Research) bind search and reasoning in the same model — search quality is dragged down by reasoning burden, and reasoning quality is polluted by search noise.
Rhizome’s approach is complete decoupling: the search agent is only responsible for finding relevant information, and the reasoning layer is only responsible for structured reasoning on existing evidence.
The information demand of forecasting tasks is not “to accurately answer user questions”, but “to discover all related signals as much as possible”.
The optimization goal of the search agent is information relevance, not answer accuracy. If the search agent is trained to “find answers”, it will tend to find content that looks like a conclusion, which is exactly the most dangerous scenario.
The training adopts a reinforcement learning framework (SearchRL), with two paths iterating in parallel:
Path A: RL fine-tuning of open-0B parameters, taking search relevance as reward (Reference: Search-R1, COLM 2025; ReSeek, ICML 2026)
Path B: Harness of closed-losed-
Different base models show stable differences in forecasting tasks: some are suitable for long-time retrieval and complex reasoning, some are better at quantitative modeling, and some have advantages in cost and response speed.
Rhizome places forecasting protocols, Agent orchestration, tool calls, and result evaluation outside the base model. Institutions can choose the base model according to task value and operation scale, and make trade-offs between reasoning capability, cost, and response speed.
Design Decision 2: Trajectory Recording and Probability Calibration to Precipitate Data Assets
Rhizome retains a complete trajectory with version information for each forecast: the time conditions and settlement standards of the question, the evidence available at the time of forecasting, the Agent orchestration and tool call process, the final answer and probability, as well as the corresponding model and system version.
These records are formed before the results are revealed, preserving the real judgments made by Rhizome when it does not know the answers yet.
After the event is settled, Rhizome puts the forecasting trajectory and the actual result back into the same record, looking back at what information was mastered at that time, which reverse evidence was missed, and whether the error occurred in retrieval, time judgment, reasoning, answer expression, or probability calibration.
A high-confidence error and an error close to 50-50, although both are recorded as “wrong answers”, expose different problems.
Rhizome will make multiple independent forecasts for the same question. The system does not simply average the results, but first aggregates them in the logit space, and then adjusts the extremization intensity combined with the Brier Score of the settled questions.
The more independent the information provided by different trajectories, the clearer the aggregation result can be; the more overlapping the evidence, the more restrained the adjustment will be.
On this basis, Rhizome uses Platt scaling to identify persistent overconfidence or over-caution through settled questions, and corrects subsequent probabilities.
The calibration standard is intuitive: 60% of the events given by the system should occur about 60% of the time in the long run; 80% of the events given by the system should occur about 80% of the time in the long run.
Each operation corresponds to the system version and key configuration at that time, so that probability changes and abnormal results can be traced back to specific causes.
The base model can be upgraded or replaced, while problem definitions, evidence records, belief changes, settlement results, and calibration experience remain continuous.
This kind of feedback data cannot be batch-created after the results are revealed — each sample must leave a judgment before the future happens, and then wait for reality to give the answer.
Code can be replicated, but data assets precipitated over time cannot be achieved quickly.
Design Decision 3: Continuous Update with Causal Chain Awareness
When the event has not been settled, new data, policies, and market information will continue to emerge, and the system needs to judge whether the original probability should be updated.
Rhizome saves the belief state for unsettled questions. Each piece of evidence separately records the occurrence time of the event, the release time of the content, and the system reading time.
When the new information is completely duplicated with existing records, the system skips the update; when the new evidence conflicts with old records, the old evidence will not be deleted, and the system retains the process of judgment correction.
When a single probability change exceeds 0.15, it must point to the specific new evidence that triggers the change.
Here is a more difficult problem: whether the new evidence represents multiple independent forces, or signals left by the same causal chain at different positions.
Interest rate hike announcements, interest spread changes, and capital flows may appear one after another, but they are not necessarily three pieces of independent information.
If the system includes them in the probability update separately, the same cause may be double-counted, pushing the probability to the extreme of overconfidence.
Rhizome is developing a Bayesian update framework with causal chain awareness. Before the evidence enters the probability update, it first identifies the causal transmission chain it belongs to, and then adjusts the evidence weight according to the intra-chain relationship.
This framework consists of four layers:
Causal Knowledge Base: stores verified causal chains, and records the transmission time lag, influence attenuation, and historical credibility of key relationships.
Same-Chain Signal Deduplication: subsequent signals from the same causal chain will no longer be repeatedly included with full weight, avoiding multiple calculations of the same force.
Global Posterior Cap: limits the cumulative impact brought by multiple evidences in the same direction, and adjusts the constraint intensity according to the forecast period.
Transmission Time Lag Awareness: after the event at the head of the chain occurs, the system updates gradually according to the transmission progress of the causal relationship, and does not immediately include all the impacts into the result at the tail of the chain.
On DigClaw’s internal forecasting system, this framework has completed prototype implementation and preliminary verification, and has guided multiple investment practices.
Experiments show that compared with direct Bayesian aggregation, same-chain signal deduplication and global posterior cap reduce the overconfidence rate (the proportion of forecasts with probability higher than 85% but finally judged wrong) from about 25% to 12%.
Why an Investment Institution Develops Forecasting Models
Newborn Ventures is initiated by DigClaw, and it is the world’s first investment and incubation institution driven by AI to mine Beta trends.
DigClaw’s trend forecasting base model is the underlying technical framework that supports this AI-Native VC.
The essence of investment is forecasting.
Judging whether a track will explode, whether a team can stand out, and whether a technology will become mainstream are all forecasting problems.
Traditional investment relies on the experience, intuition and information asymmetry of partners, but DigClaw believes that these judgments can be systematized and modeled.
From a broader perspective: investment return = Beta (event/trend driven) + Alpha (asset specificity).
Alpha research is relatively mature, but Beta — the forecasting of macro events and trends — there is no truly effective AI solution at present, which is exactly the problem DigClaw wants to solve.
When forecasting changes from an intuitive judgment to a callable, integrable, and calibratable parameter, the decision scenarios it can embed are far more than what we see now.
For listed companies, it means completing strategic pre-judgment and risk warning before industrial chain changes and policy trends take shape;
For investment institutions, it means discovering value before consensus is formed;
For government guidance funds, it means using a systematic method to study and judge industrial trends and policy effects.
The same set of forecasting capabilities accepts public evaluation on FutureX externally, drives investment decisions internally, and also opens services to industrial parties, financial institutions, and government guidance funds.
For DigClaw and Newborn Ventures, the first place on FutureX is the starting point, not the end.
About DigClaw
DigClaw is an AI technology company focusing on forecasting intelligence, whose core mission is to build a calibratable, auditable, and integrable forecasting infrastructure.
Its flagship forecasting framework Rhizome realizes system capabilities independent of the base model through a three-layer architecture of causal reasoning, probability calibration, and search intelligence — the base model is replaceable, and forecasting assets continue to accumulate.
DigClaw’s forecasting system has obtained external verification on public evaluation platforms such as FutureX, and applies the same set of capabilities to practical scenarios such as investment decision-making, industrial trend research and judgment, and strategic risk assessment, opening cooperation to industrial parties, financial institutions, and government guidance funds.
About Newborn Ventures
Newborn Ventures is an AI-native early-stage venture capital and incubation institution.
Its core belief is: the essence of investment is forecasting, and judging whether a track will explode, whether a team can stand out, and whether a technology will become mainstream are all forecasting problems that can be systematized and modeled.
Through self-developed Deep Research Agent and causal forecasting system, Newborn Ventures tracks innovative signals in the direction of AI reasoning worldwide, and discovers structural opportunities that have not been priced by the market — the real Beta.
From the first contact to investment decision, it uses AI to reconstruct every link, and promises to give clear feedback with technical depth within 48 hours.
Research-Driven. Non-Consensus. AI-Native.
*This article is authorized to be and the views belong only to
