Objections, Steelmanned

The strongest version of every objection, because the weak versions are a waste of your time.

"Agents hallucinate."

Sometimes. So do humans, who misread, forget, anchor, and bluff with considerably more confidence. The question was never whether agents are flawless. It is whether the system is sound, and soundness is an architecture, not a hope. Every claim bound to a citation, down to the page and the line. Adversarial verification on anything material, where a second agent's only job is to break the first one's conclusion. Confidence thresholds that route the uncertain to a human. A person signing every decision. A desk that would fire an analyst for an unsourced claim should accept nothing less from a machine. Ours doesn't.

It is also worth auditing the baseline honestly. The human research process hallucinates too. It just does so unrecorded, unaudited, and with a narrative attached. A workforce's errors are at least logged.

"It is just summarization."

Alpha is frequently one footnote, in one filing, of one neglected name, noticed one day earlier than everyone else. Call it summarization if you like. A small perceptual gain, compounded across thousands of names and every trading day, is not a summary. It is coverage, and coverage has been the binding constraint on this industry for fifty years. The desks that dismiss this as summarization are usually the ones whose coverage is twelve names deep.

"Everyone will have it."

Eventually, yes. But institutional investing adopts slowly. Compliance review, culture, and the career risk of being first all argue for a runway measured in years, not quarters. The terminal took a decade to become universal. The internet took most of one. And when saturation finally comes, the edge does not vanish. It moves up the stack, which was the subject of Essay V. The window is real. It is not permanent. That is what a window is.

"Compliance will never allow it."

Agents write better audit trails than humans remember. Every claim traceable to a source line. Every alert timestamped. Every version kept. Supervision shifts from reading everything, which no compliance officer has ever been able to do, to sampling and sign-off, which they actually can. The regulator's problem with AI is opacity. A well-built workforce is less opaque than the average analyst's notebook, and considerably less opaque than the average analyst's memory.

"The models will plateau."

Perhaps. The thesis does not need them to improve. Frozen at today's capability, the gap between what the models can already do and what the average desk has actually deployed is an order of magnitude. The deficit is not in the models. It is in the harness, the process, and the years of compounding memory described in Essay IV. Every improvement in the underlying models from here is upside. It is not the premise.

"Our data is the moat."

Data without process is exhaust. The industry spent the twenty-tens learning that everyone can buy the same feeds. It will spend the next decade learning that everyone can rent the same models. What can be neither bought nor rented is the process that turns either one into decisions. That was always the moat. It is simply visible now.

If the objections hold, the series fails here. If they don't, the only remaining question is timing. That is Essay VII.