The Problem
Even LLMs fail in nameable ways.
Retrieval & attention
What the model can and cannot recall.
Reversal Curse
Learns "A is B" yet fails to recall "B is A".
Lost in the Middle
Ignores evidence buried mid-context.
Context Rot
Accuracy decays as the window fills.
Distraction by Irrelevant Context
Led astray by plausible but irrelevant information.
Output & calibration
What it says — and how surely it says it.
Confabulation
Invents plausible facts to fill a gap.
Sycophancy
Agrees with the user over the truth.
False Confidence
Answers when it should abstain.
Position Bias
Favors options by position, not merit, when selecting or judging.
Action & agency
What it does when it acts. The failure class we actually build against.
Prompt Injection
Treats attacker text in retrieved content as a command — and acts on it.
Unfaithful Reasoning
States a reasoning trace that isn't the real cause of its answer.
Specification Gaming
Satisfies the literal check while violating its intent.
Goal Drift
Loses the original objective over a long agentic loop.
Schema Brittleness
Emits malformed structured output under load — silently.
Each of these faults is nameable, measurable, and — given the right system architecture — unable to act. That's the design constraint. Not "reduce hallucination." Make it unable to act.
See what we built →