LLMs can’t fix their own epistemic blind spots.
Not with better prompting.
Not with stricter verification.
Not with more sophisticated agents.
The MarCognity benchmark reveals a persistent 8–15% failure rate across 8 domains,
even when the system is equipped with a full metacognitive layer.
This residual error is not noise.
It may be an emergent property of autoregressive generation itself —
a structural boundary where the epistemic space accessible to a probabilistic model
is narrower than the space of claims requiring justification.
I call this the Epistemic Boundary.
Resources:
• Zenodo : https://doi.org/10.5281/zenodo.19020581