Immune memory and affinity maturation in venture capital

The objection that should win

Here is the strongest version of the case against this page. A memory B cell recognises an epitope: a specific molecular shape, refined over weeks of somatic mutation until its receptor binds one thing very tightly. It does not recognise "a respiratory threat." It cannot generalise from influenza to a coronavirus it has never seen, because its affinity was purchased at the cost of breadth. Calling this a working model of anything is a category error. It is a lookup table with exquisite resolution and zero range.

Apply that to venture capital and the analogy looks worse, not better. An investing partner's job is not epitope matching. It is judging whether a market structure will hold, whether a hiring pattern signals real demand or a founder's optimism, whether a filing buried in a competitor's S-1 changes the shape of the opportunity. That is causal reasoning about situations, not shape recognition. A system that revises confidence on contact with new evidence — filings, hiring signals, product telemetry — sounds like discipline. But if the underlying representation is as narrow as an antibody's paratope, the discipline is cosmetic. You can update fast and still be wrong about the thing that matters, because you were never modelling the thing that matters.

That objection is correct on its own terms. It should be stated in full before any defence is attempted, because the defence does not deny it.

What a term sheet has in common with a B cell

The reply starts by giving up the ground the objection claims. Immunity has no theory of why a pathogen appeared. It has no model of epidemiology, no sense of seasons or ports of entry or the mutation rate of a virus family. It has affinity: how tightly a receptor binds, tuned by selection pressure inside a germinal centre, nothing more. If the analogy to venture capital rested on representational richness — on claiming that continuous intake makes a partner's model of a market smarter — it would fail immediately, and it should fail.

The analogy is not about richness. It is about intake discipline: how a system that cannot afford to pause acquires evidence, dates it, weights it, and retires it under threat of a costly error. On that narrower axis, adaptive immunity is genuinely advanced, and venture investing is one of the few commercial domains that runs a comparable regime whether or not anyone names it. A fund's beliefs about a portfolio company are not fixed at the term sheet. They are continuously revised by four streams that never stop: regulatory and financial filings, hiring signals scraped from job postings and LinkedIn movement, product telemetry when it is shared or inferred, and market structure — who else is raising, who just folded, what a competitor's pricing page now says. Each of these is dated. Each carries provenance: this belief about the total addressable market came from a Q3 filing, that belief about traction came from a headcount spike in a specific function. A partner who is doing this well is not holding a static thesis. They are holding a population of competing hypotheses about the company, mutating confidence up or down as evidence lands, exactly as a germinal centre holds competing B cell clones and starves the ones that stop binding.

The narrowness objection lands on what the system knows. It does not land on how the system stays current. Those are different questions, and venture capital's characteristic failure comes from confusing them.

The failure this domain knows by name

Original antigenic sin is the immunological name for a specific pathology: a first exposure to influenza biases the response to every subsequent variant for decades, because the immune system's memory cells out-compete naive ones even when a newer, better-matched response is available. The first encounter wins by incumbency, not by fit.

Venture capital's version of this failure has no textbook name but every partner has seen it. A thesis is formed on strong early evidence — a category is heating up, a wedge product is landing, a specific hiring pattern (say, a company pulling senior infrastructure engineers from a particular incumbent) reliably preceded breakout growth in three prior portfolio companies. The thesis gets funded, gets defended in partner meetings, gets restated in the annual letter. Then the market it assumed dissolves — a platform shift removes the wedge, a larger player commoditises the category, the hiring signal stops meaning what it meant because everyone now knows to fake it — and the thesis survives anyway, for a year, sometimes two, because the partner's memory clones of the original, successful pattern out-compete the fresher, worse-fitting evidence coming in from filings and telemetry. The prior conviction has affinity to a market that no longer exists. It keeps winning selection inside the partner's head long after it has stopped winning outside it.

This is not a failure of intake. The streams were still running: the filings kept arriving, the hiring signal kept being logged, the competitor's pricing page kept updating. The failure was in selection — in which hypotheses got starved and which got protected. That is exactly the immunological failure mode, and naming it precisely is more useful than treating it as generic stubbornness, because it points at the fix immunology already found.

A model that revises on contact with evidence should never be able to defend a position for a year after the evidence turned against it. If it can, the revision mechanism is decorative.

That objection is fair, and it is the second one worth answering directly, because it is the one that matters most in this domain.

Governance, not narrower intake

Adaptive immunity did not solve its own version of this problem by shrinking its intake or slowing its mutation rate. It solved it by adding governance on top of open intake. Thymic negative selection destroys the majority of T cells whose receptors bind self-tissue too strongly, using the AIRE transcription factor to force a deliberately constructed reference set of self-antigens into the selection process. Regulatory T cells continue that policing after cells leave the thymus. Checkpoint receptors throttle responses that are winning too hard. None of this narrows what the system observes. All of it narrows which internally-generated hypotheses are allowed to keep out-competing better-fitting rivals.

Venture capital's structural equivalent already exists in weak form and is worth stating as what it actually is: a governance layer, not an intake fix. An investment committee that requires a partner to restate the thesis against this quarter's filings, rather than against the memo that got the deal approved, is running something like negative selection — forcing the belief to be re-justified against a reference set it did not choose. A portfolio review that explicitly asks "what evidence would change our position" and checks whether that evidence has, in fact, arrived, is a checkpoint receptor. A fund that tracks the provenance of its thesis — this confidence level, dated to this filing, superseded by this hiring signal — can see the moment a prior conviction started winning on incumbency rather than fit. Most funds do not build this. They keep the intake — the streams are monitored, the data rooms are current — but they never police which beliefs inside a partner's head are allowed to keep beating fresher ones. The intake axis is often fine. The selection axis is where a fund's original antigenic sin actually lives.

The claim, narrowed

None of this makes an investing partner's judgement into a Large Universe Model, and no such claim is being made. It makes the intake regime the same shape: filings, hiring signals, telemetry and market structure as four permanently open streams, each observation dated and provenance-tagged, confidence in any given thesis raised or lowered on contact rather than fixed at the term sheet and revisited once a year. A Large Language Model's equivalent would be a fund that trained its judgement once on historical outcomes and never updated it — a repertoire frozen at a cutoff, enormous and naive. A Large World Model's equivalent is the partner who reacts sharply to whatever is in the current data room and forgets it once the deal closes — fast, local, largely without memory. The position this domain is straining towards, whether or not it is ever built as software, is the third one: every stream still running, every belief tagged with the filing or signal that produced it, decaying on schedule unless renewed by contact.

What remains open after that is exactly what remained open for vertebrates after adaptive immunity arrived: not a new category of evidence, but how fast to mutate a thesis, how strictly to select against it, how long to let a conviction persist without renewal, and how hard to punish a partner's own attachment to their best prior call. Those are calibration questions. They are also, in this domain, the ones that separate a fund that retires a stale thesis in a quarter from one that defends it for a year after the market underneath it is gone.