Emergent novelty in venture capital

The objection that should win

Start with the strongest version of the case against this thesis, because venture capital is where it should land hardest.

A partner's job is pattern recognition applied to companies that do not yet exist as categories. Every fund's edge is supposed to be exactly the kind of recognition emergent novelty says cannot be listed in advance. So the objection runs: if novelty in venture is real, no stream of filings, hiring data or telemetry will save a partner from missing it, because the miss happens in judgement, not in intake. Sequoia did not miss the shift from on-premise to cloud because a filing was late. Kleiner did not miss mobile because a sensor was pointed the wrong way. The failure mode in this industry is conviction, not coverage. A partner who has built a career on a thesis defends it past the point of falsification because reputation, carry and portfolio construction all lean the same direction. No amount of open streaming fixes a human who will not update. The metaphysics of emergent novelty, on this account, is doing no work; the actual failure is psychological and institutional, and dressing it up as an intake problem flatters the wrong culprit.

This is close to right, and it deserves to be taken seriously rather than waved off with a reframing.

What survives the concession

Concede the psychology entirely. A partner who has raised a fund on a thesis — say, that vertical marketplaces win by owning transaction data that incumbents cannot replicate — has every incentive to keep believing it. Carry vests on conviction sustained long enough to exit. LPs were sold a thesis, not a stream. Firms build entire brand identities around a bet: "we saw the shift to usage-based pricing before anyone." Walking that back in month eleven, in front of the same LPs, costs something real. None of that is solved by better data. A partner with perfect telemetry can still choose not to look at it, or look and rationalise.

But notice what this objection actually concedes: that the failure is a defended belief outliving the evidence against it, which presupposes the evidence existed and was ignorable, not that it was structurally unavailable. That is a different claim from the one about intake architecture. The psychological failure and the intake failure can coexist and often do, but they are not the same failure, and only one of them is fixable by design rather than by character. The thesis about intake is narrower than "partners will make good decisions." It is: whatever a partner eventually needs to see, was there a channel through which it could have arrived, tagged with its source, before the thesis calcified? That is a question about the fund's information architecture, not about anyone's nerve.

Filings, hiring signals, telemetry, and what each one enumerates

Look at what a fund actually streams and where each stream's catalogue runs out.

Delaware incorporations and SEC Form D filings tell you a company exists and roughly how it is capitalised. They enumerate the legal shell. They say nothing about a business model that has not yet been named in a filing's boilerplate, and they lag reality by weeks. Hiring signals — LinkedIn postings, AngelList listings, comp data scraped from sites like Levels.fyi — enumerate roles against an existing taxonomy of job titles. A company inventing a genuinely new function, the way "growth engineer" or "prompt engineer" had no slot a few years earlier, shows up as noise in the hiring feed until someone decides the noise is a category. Product telemetry — App Store rank, Sensor Tower download estimates, API call volume where it is observable — enumerates whatever the platform chooses to expose. A market structure shift that happens inside a platform's black box, such as a distribution algorithm change that quietly kills a wedge, never appears in the telemetry at all; it appears later, as an unexplained collapse in a metric nobody had instrumented for cause.

This is the domain-specific version of the general pattern. A memo written from a fixed corpus of case studies is the Large Language Model failure: it interpolates the next vertical marketplace from the shape of the last one, missing that the underlying unit economics no longer transfer. A due-diligence process built around a fixed dashboard of metrics chosen at fund formation is the Large World Model failure: it senses the present deal beautifully, through channels decided years earlier, and cannot register a signal that arrives outside those channels — a regulatory change abroad, a platform policy update buried in a developer changelog, a competitor's stealth pivot that never touches a tracked metric. Neither failure is about willpower. Both are about which catalogue was fixed before the novelty arrived.

Where closure fails, and why it does not sink the claim

The second objection worth taking seriously is that the third row of that table is itself an enumeration — of data vendors, of scrapers, of the fields a fund's research team decided to log. A regulatory change published only in a foreign-language gazette, or a platform's internal algorithm shift that leaves no public trace until users notice organically, radiates into no stream the fund has built. "Everything, continuously" describes an intention, not a coverage guarantee. A fund that believes its own marketing about total visibility is exposed in exactly the way a fund with a static dashboard is exposed, just later and with more confidence.

Accepted, and the claim has to be narrowed to survive it. The commitment being defended is not that any given fund's streams are complete. They never will be. The commitment is that closing a gap never requires abandoning the intake architecture, only extending it. When a fund discovers, in year three of a thesis, that a competitor's pivot happened inside a Discord server no scraper covered, the fix is to add Discord monitoring as a new stream with its own provenance tag — not to rebuild the research function around a different kind of intake. Contrast this with a fund whose diligence process is built on fixed comparables: discovering the same gap forces a redo of the entire memo from scratch, because the comparable set itself was the intake mechanism. Coverage grows within the open-stream position without changing its kind. It does not grow at all within a fixed corpus or a fixed dashboard; it requires migrating to a different regime.

The thesis that outlives its market

Put this back in the partner's chair. A thesis is written in year one: freight brokerage will be won by whoever owns the load-matching data, because incumbents' load boards are fragmented and slow. The portfolio company is funded on that basis. In year two, one of the largest incumbent load boards ships a free, real-time matching API as a defensive move, absorbing the exact wedge the thesis depended on — a market structure change visible in a changelog and a pricing page, not hidden. The partner keeps defending the company for another year anyway, because the fund's board memos, LP updates and internal culture were built around the original thesis and reversing it costs standing.

The intake failure and the conviction failure are visible separately here. The market structure stream existed and could have surfaced the change in weeks. Whether it did surface it, and whether anyone acted on it, are questions about incentives and courage that no amount of streaming settles. What streaming settles is narrower and still real: it determines whether the contradicting evidence was retrievable with its source and its date, so that when someone finally does look — a new partner, an LP on a diligence call, the founder themselves — the record shows exactly when the assumption broke, rather than forcing a reconstruction from memory and defensiveness.

The narrower claim

Venture capital will keep producing categories no prior fund could have listed — new business models, new distribution mechanics, new failure modes for old models under new regulation. That is unbounded and no architecture changes it. What can change is the interval between a market's dissolution and a fund's recognition of it, and whether the evidence that the dissolution happened is sitting somewhere with a timestamp and a source, waiting to be read, rather than lost inside a dashboard nobody widened. That is the whole claim. It does not fix conviction. It only guarantees that when conviction finally breaks, it breaks against a record rather than against silence.