LLM Knowledge-Work Utility — Economically Valuable Task Performance

Hallucination under low supervision. LLMs produce confident, fluent, and factually incorrect output at non-trivial rates. In supervised settings, human review catches and corrects errors before they produce harm. As supervision decreases, uncorrected errors propagate. The mechanism creates a hard constraint on the supervision level at which the claim holds: the claim is true under adequate supervision; the supervision threshold required varies by task stakes. This is not a reason to contest the claim but a reason to specify its domain of validity precisely.

Agentic error compounding. In single-turn assisted tasks, errors are local and recoverable. In multi-step agentic workflows, errors compound: a mistaken intermediate step propagates through subsequent steps, producing outputs that are difficult to audit and expensive to correct. The mechanism is evidenced by INST-006. It is structurally distinct from hallucination — it is not primarily about factual accuracy but about error propagation in sequential decision-making. As the frontier moves toward agentic deployment, this mechanism becomes the primary constraint on the claim's "limited supervision" condition.

Knowledge-work task scope is undefined. The claim spans a range from low-stakes, single-turn drafting assistance to high-stakes, multi-step autonomous professional work. The evidence is positive at the low end and negative or qualified at the high end. The bottleneck is not evidential — more evidence will not resolve it — it is definitional. A scoped version of the claim (e.g. restricted to single-turn assisted tasks in low-stakes domains) would already be confirmable. A broader version (including autonomous high-stakes professional work) would not. The Observatory records this without resolving it: the claim as stated spans both, and both are worth tracking.

Agentic capability threshold as resolution point. The primary open question for this record is whether LLMs can perform valuable knowledge-work autonomously over extended task sequences — the agentic frontier. Current evidence shows strong performance in supervised single-turn settings and weaker performance in autonomous multi-step settings. If agentic systems demonstrate reliable multi-step task completion in commercially deployed settings, the claim's upper scope boundary moves significantly. This is AT-001: the agentic capability frontier is drawing the evidence trail toward a potential resolution of BN-001. It is not yet reached; it is the attractor the record is currently moving toward.