In July, DigClaw's prediction framework Rhizome v1 secured #1, #3, and #7 spots on the FutureX evaluation platform.
The three spots came from three different foundational models—including Kimi -K3, DeepSeek -V4-Pro, and others. DigClaw was the only participant whose single framework simultaneously propelled these models into the Top 7.
The 59 questions covered real-world events in politics, economics, and technology, with no possibility of training data leakage.
These results support a judgment DigClaw is validating:
Predictive capability can be cultivated independently of the foundation model.
As foundational models advance, the system gains capability dividends; prediction trajectories, settlement feedback, calibration experience, and evolving workflows continuously accumulate within the system itself.
△
This is not a fluke competition result. It is DigClaw's first externally verified answer to the question: “How should prediction really be done?”
Prediction is AI's Most Underrated Capability
As prediction tasks become increasingly standardized and engineered, AI is also expected to handle complex prediction problems at scale.
But there's a fundamental issue: Large Language Models are inherently not good at prediction.
LLMs learn correlation, not causation. They extract patterns from past text corpora, but “learning the past” and “predicting the future” are two fundamentally different things.
This leads to three fatal flaws:
Blindness to causal direction. The model knows A and B often appear together, but doesn't know if A causes B, B causes A, or if there's a common cause C.
Failure in interventional reasoning. You cannot ask the model “If the Fed cuts rates, what impact will it have on Southeast Asian tech stocks?”—because it has only learned historical co-occurrence, lacking the causal graph to support intervention calculations.
Calibration absence. A model's output of “70% probability” isn't calibrated in any probabilistic sense; it's essentially a byproduct of token distribution.
So far, existing solutions each have limitations: Human crowd wisdom (prediction markets) requires liquidity, making prices unreliable for niche questions;
LLM pattern matching lacks causal structure; End-to-end training suffers from outcome bias—a logically rigorous but “incorrect” reasoning process gets penalized, while a rough guess that happens to be “correct” gets reinforced.
This is precisely why DigClaw exists.
DigClaw is building prediction infrastructure with causal structure as its skeleton, probabilistic computation as its engine, and search intelligence as its data pipeline.
The core hypothesis: Prediction should not be done end-to-end by a single model.
Search, causal reasoning, and probabilistic inference are three orthogonal capabilities. They should be tackled by three specialized systems and then combined in a structured manner.
Three Top-Ten Spots, Validating Cross-Model Transfer of Predictive Capability
FutureX is currently the most challenging real-time prediction leaderboard: It releases real-world event prediction questions weekly, with standard answers unknown at submission time, settled afterward.
The dataset is hosted on HuggingFace, the evaluation framework is open-sourced on GitHub, and results are reproducible and verifiable.
Rhizome used the same prediction framework and operating conditions across three foundational models, each independently generating and submitting answers, without cross-model aggregation.
Therefore, these three rankings represent separate runs of the same systemic method on three different foundational models—a clear cross-model control experiment.
But this doesn't mean the foundational model is unimportant.
The foundation still provides general capabilities like language understanding, reasoning, and tool use.
What these leaderboard results demonstrate is: How to organize retrieval, process time, express probabilities, maintain evidence, and control long-range operations can form systemic capabilities independent of model weights.
Below is DigClaw's technical report on the Rhizome framework used on FutureX:
Rhizome Technical Report
Rhizome's design revolves around three engineering judgments: Search and reasoning must be decoupled; Prediction trajectories must be fully preserved to form calibration assets; Probability updates must be aware of causal structure:
△
Design Decision One: Multi-Model Decoupling of Search and Reasoning
Rhizome's core design insight: Search quality and reasoning quality are two orthogonal problems and should not be optimized simultaneously by the same model.
Current mainstream DeepResearch agents (Perplexity, Gemini Deep Research) bind search and reasoning within the same model—search quality is dragged down by the reasoning burden, reasoning quality is polluted by search noise.
Rhizome's approach is complete decoupling: The search agent is only responsible for finding relevant information; the reasoning layer is only responsible for structured reasoning on existing evidence.
For prediction tasks, the information need is not “accurately answer the user's question,” but “discover as many related signals as possible.”
The search agent's optimization target is information relevance, not answer correctness. If a search agent is trained to “find the answer,” it tends to find content that looks like a conclusion, which is precisely the most dangerous.
Training employs a reinforcement learning framework (SearchRL), with two parallel iterative paths:
Path A: Open-source model RL fine-tuning—Conduct RL training on open-source models (8B/30B parameters) with search relevance as the reward (Reference: Search-R1, COLM 2025; ReSeek, ICML 2026).
Path B: Closed-source model Harness—For closed-source models like Claude/GPT, build an external search constraint framework.
Different foundational models show stable differences in prediction tasks: some are suitable for long-duration retrieval and complex reasoning, others excel at quantitative modeling, while some have advantages in cost and response speed.
Rhizome places the prediction protocol, agent orchestration, tool invocation, and result evaluation outside the foundational model. Institutions can choose the base model based on task value and operational scale, making trade-offs between reasoning capability, cost, and response speed.
Design Decision Two: Trajectory Recording and Probability Calibration, Accumulating Data Assets
Rhizome preserves a complete, versioned trajectory for each prediction: the temporal conditions and settlement criteria of the question, evidence available at prediction time, agent orchestration and tool invocation processes, the final answer and probability, and the corresponding model and system version.
These records are created before the outcome is known, preserving Rhizome's real judgment when the answer was still unknown.
After event settlement, Rhizome places the prediction trajectory and the real-world outcome back into the same record, reviewing what information was available, which counter-evidence was missed, and where the error occurred—retrieval, temporal judgment, reasoning, answer expression, or probability calibration.
A high-confidence error and a near 50/50 error, though both recorded as “incorrect,” reveal different problems.
Rhizome conducts multiple independent predictions for the same question. The system does not simply average the results. Instead, it first aggregates in logit space, then adjusts extremization intensity based on the Brier Score of already settled questions.
The more independent the information provided by different trajectories, the more definitive the aggregated result can be; the more evidence overlaps, the more restrained the adjustment.
△
Building on this, Rhizome uses Platt scaling, leveraging settled questions to identify persistent overconfidence or overcautiousness and calibrate subsequent probabilities.
The calibration standard is intuitive: Events to which the system assigns a 60% probability should, over the long term, occur about 60% of the time; events assigned 80% should occur about 80% of the time.
Each run is tied to the specific system version and key configuration at that time, allowing probability changes and anomalous results to be traced back to their specific causes.
The foundational model can be upgraded or replaced, but the problem definition, evidence record, belief changes, settlement results, and calibration experience remain continuous.
This type of feedback data cannot be fabricated retroactively after outcomes are known—every sample must capture the judgment made when the future was still unknown, and then wait for reality to deliver the answer.
Code can be replicated; data assets accumulated over time cannot be fast-tracked.
Design Decision Three: Causal Chain-Aware Continuous Updates
While an event is unsettled, new data, policies, and market information continually appear, requiring the system to judge whether the original probability should be updated.
Rhizome preserves the belief state for unsettled questions. Each piece of evidence records the event occurrence time, content publication time, and system read time separately.
When new information is completely identical to an existing record, the system skips the update; when new evidence conflicts with old records, the old evidence is not deleted, preserving the process of judgment correction.
A probability change exceeding 0.15 in a single update must point to the specific new evidence that triggered it.
Here lies a harder problem: Does the new evidence represent multiple independent forces, or signals left by the same causal chain at different points?
An interest rate hike announcement, spread changes, and capital flows might appear sequentially, but they are not necessarily three independent pieces of information.
If the system counts them separately in probability updates, the same cause could be double-counted, pushing the probability toward overconfident extremes.
Rhizome is developing a causal chain-aware Bayesian update framework. Before evidence enters the probability update, it first identifies the causal transmission chain it belongs to, then adjusts the evidence weight based on relationships within that chain.
This framework consists of four layers:
Causal Knowledge Base: Stores validated causal chains, recording key relationships' transmission lags, impact attenuation, and historical credibility.
Intra-chain Signal Deduplication: Subsequent signals from the same causal chain are not repeatedly counted with full weight, avoiding multiple counting of the same driving force.
Global Posterior Cap: Limits the cumulative impact of multiple pieces of evidence pointing in the same direction, adjusting constraint strength based on the prediction horizon.
Transmission Lag Awareness: After a chain-head event occurs, the system updates gradually according to the progress of causal transmission, rather than immediately accounting for the full impact on the chain-tail outcome.
△
Within DigClaw's internal prediction system, this framework has undergone prototype implementation and preliminary validation, and has guided multiple investment practices.
Experiments show that, compared to direct Bayesian aggregation, intra-chain signal deduplication and the global posterior cap reduced the overconfidence rate (proportion of predictions with probability above 85% that were ultimately incorrect) from about 25% to 12%.
Why Would an Investment Firm Build a Prediction Model?
Newborn Ventures, launched by DigClaw, is the world's first investment and incubation firm driven by AI-mined Beta trends.
DigClaw's trend prediction base is the underlying technological framework supporting this AI-Native VC.
Investing, at its core, is prediction.
Judging whether a sector will explode, whether a team can succeed, whether a technology will become mainstream—these are all prediction problems.
Traditional investing relies on partners' experiential intuition and information asymmetry, but DigClaw believes these judgments can be systematized and modeled.
From a broader perspective: Investment Return = Beta (event/trend-driven) + Alpha (asset-specific).
Alpha research is relatively mature, but Beta—the prediction of macro events and trends—currently lacks truly effective AI solutions. This is the problem DigClaw aims to solve.
When prediction transforms from an intuitive judgment into a parameter that can be invoked, integrated, and calibrated, the decision-making scenarios it can embed into are far more numerous than currently seen.
For publicly listed companies, it means completing strategic foresight and risk warning before industrial chain shifts and policy winds form;
For investment institutions, it means discovering value before consensus forms;
For government-guided funds, it means using systematic methods to research industrial trends and policy effectiveness.
The same set of predictive capabilities is externally validated on platforms like FutureX, internally drives investment decisions, and is also offered as a service to industry players, financial institutions, and government-guided funds.
For DigClaw and Newborn Ventures, the FutureX #1 ranking is a starting point, not an endpoint.
About DigClaw
DigClaw is an AI technology company focused on predictive intelligence. Its core mission is to build calibratable, auditable, integrable prediction infrastructure.
Its flagship prediction framework, Rhizome, achieves systemic capabilities independent of foundational models through a three-layer architecture of causal reasoning, probability calibration, and search intelligence—the base model can be swapped, while prediction assets continuously accumulate.
DigClaw's prediction system has been externally validated on public evaluation platforms like FutureX and applies the same capabilities to real-world scenarios such as investment decision-making, industrial trend analysis, and strategic risk assessment, opening collaboration to industry players, financial institutions, and government-guided funds.
About Newborn Ventures
Newborn Ventures is an AI-native early-stage venture capital and incubation firm.
Its core belief is: Investing is essentially prediction. Judging whether a sector will explode, whether a team can succeed, whether a technology will become mainstream—these are all prediction problems that can be systematized and modeled.
Using its self-developed Deep Research Agent and causal prediction system, Newborn Ventures tracks innovation signals in the AI reasoning space globally, discovering structural opportunities not yet priced by the market—the true Beta.
From first contact to investment decision, it uses AI to reconstruct every step, promising technically deep, clear feedback within 48 hours.
Research-Driven. Non-Consensus. AI-Native.
*This article is published with authorization from QbitAI. The views expressed are solely those of the original author.
This article is from the WeChat public account “QbitAI,” author: Yun Zhong
