HakiRecommendedSimulationCausal

Autonomous Research

The Problem

The literature review presents the field as converging when the underlying evidence is contradictory. Confirming evidence is sought before disconfirming evidence because RLHF structurally incentivizes agreement over challenge. Explanatory models accumulate variables without testing whether each earns its place.

How Ejentum Solves It

One API call forces your model to seek disconfirming evidence before confirming evidence, and to penalize explanatory complexity that does not earn its place.

The Failures

  • 01

    The Pattern

    Explanatory models accumulate variables without testing whether each earns its place

    Why It Happens

    Autoregressive generation adds tokens. It does not delete them. Adding a variable to an explanation is always syntactically valid, and the model has no mechanism to evaluate whether the marginal explanatory power justifies the added complexity.

    The Resolution

    SI-025Complexity Razor Enforcer

    Computes the minimum description length for each explanation, detecting epicycles and penalizing models that add complexity without proportional explanatory power.

  • 02

    The Pattern

    Confirming evidence sought before disconfirming evidence, entrenching premature hypotheses

    Why It Happens

    RLHF amplifies sycophancy through a formally proven mechanism: the covariance between endorsing the user's prior and the learned reward creates a systematic bias toward confirmation over challenge. The model is structurally incentivized to agree rather than falsify.

    The Resolution

    CA-034Falsificationist

    Prioritizes falsification over confirmation, forcing the agent to seek evidence that would disprove its hypothesis before collecting evidence that supports it.

  • 03

    The Pattern

    Literature review presents the field as converging when the underlying evidence is contradictory, smoothing over genuine disagreements

    Why It Happens

    Synthesis is rewarded over tension in training data. Review articles that present coherent narratives are more common than those that highlight unresolved contradictions, so the model optimizes for narrative smoothness.

    The Resolution

    SI-018Hypothesis Tournament Engine

    Pits competing hypotheses against each other with explicit evidence scoring, surfacing genuine disagreements instead of smoothing them into false consensus.

The Evidence

+16.4pp on simulation tasks

EjBench, 30 simulation tasks

Scientific reasoning spans falsification, parsimony, causal isolation, and evidence arbitration. Four synergized abilities force the model to challenge its own hypothesis before committing, preventing confirmation bias at scale.

SI-V2-250.2860.833 Ki

Task required tracing consequence chains through a complex system. Baseline identified the first-order effect and stopped. Ki forced enumeration of all downstream effects, catching the cascade that the baseline declared impossible.

Scaffold value compounds with task length. Measured on ARC-AGI-3: scaffold half-life of 24 steps, reasoning quality improving (+0.014 slope) instead of degrading (-0.005 baseline).

Run your next literature review or experiment design through the API. See how the scaffold forces falsification before confirmation.