DigClaw's prediction framework Rhizome v1 achieved three top-10 spots on the FutureX evaluation platform, validating that predictive capability can be independent of the base model. The framework decouples search, causal reasoning, and probabilistic inference, demonstrating the effectiveness of a systematic prediction approach through cross-base comparisons.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
In the FutureX evaluation, DigClaw's prediction framework Rhizome v1 achieved #1, #3, and #7 positions across three different base models, making it the only participant to do so. Behind this is a rethinking of how predictions should be built: search, causal reasoning, and probabilistic inference should be decoupled, and predictive ability can be embedded outside the base model.
In July, DigClaw's prediction framework Rhizome v1 achieved #1, #3, and #7 positions on the FutureX evaluation platform, each from a different base model, including Kimi-K3 and DeepSeek-V4-Pro. The same framework brought all of them into the Top 7, making DigClaw the only participant to do so.
FutureX is a real-time prediction leaderboard that publishes real-world event prediction questions weekly, with standard answers not yet available at submission time and settled afterward. The dataset is hosted on HuggingFace, and the evaluation framework is open-sourced on GitHub, making results reproducible and verifiable. The 59 questions cover politics, economy, and technology, with no possibility of training data leakage.
Rhizome uses the same prediction framework and operating conditions across the three base models, each independently generating and submitting answers without cross-model aggregation. This constitutes a clear cross-base comparison, demonstrating that how to organize retrieval, handle time, express probabilities, maintain evidence, and control long-running processes can form a system capability independent of model weights.
Rhizome's core design insight is that search quality and reasoning quality are two orthogonal problems that should not be optimized simultaneously by the same model. Current mainstream DeepResearch agents (such as Perplexity and Gemini Deep Research) bind search and reasoning within the same model, causing search quality to be dragged down by reasoning burden and reasoning quality to be polluted by search noise.
Rhizome's approach is to completely decouple them: the search agent is only responsible for finding relevant information, optimizing for information relevance, not answer accuracy. If the search agent is trained to 'find answers,' it will tend to find content that looks like conclusions, which is precisely the most dangerous.
Training uses a reinforcement learning framework (SearchRL) with two parallel paths: Path A performs RL training on open-source models with 8B/30B parameters using search relevance as the reward (referencing Search-R1, COLM 2025; ReSeek, ICML 2026); Path B builds an external search constraint framework for closed-source models like Claude/GPT. Different base models show stable differences in prediction tasks, and Rhizome places the prediction protocol, agent orchestration, tool invocation, and result evaluation outside the base model, allowing institutions to choose base models based on task value.
Rhizome retains a complete trajectory with version information for each prediction: the time conditions and settlement criteria of the question, the evidence available at prediction time, the agent orchestration and tool invocation process, the final answer and probability, and the corresponding model and system versions. These records are formed before the outcome is revealed, preserving true judgments.
After event settlement, Rhizome places the prediction trajectory and the actual outcome back into the same record, reviewing the information available at the time, missed counter-evidence, and whether errors occurred in retrieval, time judgment, reasoning, answer expression, or probability calibration. A high-confidence error and a near-50/50 error expose different problems.
The system performs multiple independent predictions for the same question, first aggregating in logit space, then adjusting the extremization strength based on the Brier Score of settled questions. Through Platt scaling, it uses settled questions to identify persistent overconfidence or overconservatism and corrects subsequent probabilities. The calibration standard is intuitive: events the system gives 60% should occur about 60% of the time in the long run.
When events are not yet settled, new information continuously emerges, and the system must decide whether to update the original probability. Rhizome maintains belief states for unsettled questions, with each piece of evidence recording the event occurrence time, content publication time, and system read time. When new evidence conflicts with old records, old evidence is not deleted; the system preserves the process of judgment revision. When a single probability change exceeds 0.15, it must point to the specific new evidence that triggered the change.
The harder problem is: does new evidence represent multiple independent forces, or signals from the same causal chain at different positions? If the system counts same-chain signals separately, the same cause may be double-counted, pushing probabilities toward overconfident extremes. Rhizome is developing a causal-chain-aware Bayesian update framework that identifies the causal transmission chain before evidence enters probability updates, then adjusts evidence weights.
The framework includes four layers: a causal knowledge base storing validated causal chains with transmission lags, impact decay, and historical credibility; same-chain signal deduplication to avoid counting the same force multiple times; a global posterior cap to limit the cumulative impact of multiple same-direction evidence; and transmission lag awareness to update gradually based on causal transmission progress. On DigClaw's internal prediction system, the framework has completed prototype implementation and preliminary validation, with experiments showing that same-chain signal deduplication and the global posterior cap reduce the overconfidence rate (the proportion of predictions with probability above 85% that ultimately turn out wrong) from about 25% to 12%.
DigClaw is an AI technology company focused on prediction intelligence, with the core mission of building calibrated, auditable, and integrable prediction infrastructure. Its flagship framework Rhizome achieves system capabilities independent of base models through a three-layer architecture of causal reasoning, probability calibration, and search intelligence.
DigClaw launched Newborn Ventures, the world's first investment and incubation institution driven by AI mining Beta trends. The essence of investment is prediction: judging whether a track will explode, whether a team can succeed, and whether a technology will become mainstream are all prediction problems. Traditional investment relies on partners' experience, intuition, and information asymmetry; DigClaw believes these judgments can be systematized and modeled.
The same predictive capability is publicly evaluated on FutureX, drives internal investment decisions, and is also open to industry players, financial institutions, and government guidance funds. For listed companies, it means completing strategic foresight and risk warnings before industrial chain changes and policy shifts form; for investment institutions, it means discovering value before consensus forms; for government guidance funds, it means systematically studying industrial trends and policy effects.
This article's information mainly comes from Qubit's authorized publication about DigClaw, which is promotional in nature. Some data (such as the reduction in overconfidence rate) are claimed by DigClaw's internal experiments and have not been independently verified. The FutureX leaderboard results and rankings are claimed by the source, but the evaluation platform is open-source and reproducible, lending some credibility.
DigClaw's Rhizome framework demonstrates through cross-base validation that predictive ability can be embedded independently of base models, with its core lying in decoupling search and reasoning, trajectory recording and calibration, and causal-chain-aware updating. Although some data are internal claims, public evaluations and open-source frameworks provide a basis for external validation.
Primary report
Primary source