The strongest objection first
Take a partner who backed a vertical-SaaS company on the thesis that procurement teams in mid-market logistics would pay for a workflow layer. Eighteen months in, growth stalls. Bookings miss forecast. The partner's answer: not the thesis, the market — a customs rule changed, freight rates collapsed, the buyer's budget cycle moved. The company is fine; the world moved. A sceptic is entitled to ask: how would you ever know whether that's true, or whether it's the oldest move in venture, the founder-friendly excuse dressed as macro?
This is the Duhem-Quine thesis, and it is not a debating trick imported from philosophy of science to flatter venture capital with borrowed rigour. Pierre Duhem observed, writing about physics in 1906, that no experiment tests a single hypothesis. A predicted result follows from the hypothesis plus a set of auxiliary assumptions — instrument calibration, background theory, the absence of interference. When the prediction fails, logic tells you only that the conjunction is false. It does not tell you which member of the conjunction to blame. Willard Van Orman Quine, in 1951, generalised this into a claim about the whole of belief: any statement can be saved from refutation if you're willing to adjust something else. A venture thesis is exactly such a conjunction — market size, buyer psychology, competitive response, timing — and a missed quarter is exactly such a failed prediction. The objection writes itself: if blame is genuinely undetermined by logic, then a partner defending a thesis for a year after the market dissolved isn't committing an error. They're doing what the philosophy says is unavoidable. Any of us, faced with the same ambiguity, would have to make the same kind of call, and no amount of additional data changes the fact that the call is a judgement, not a deduction.
That is a serious objection and it deserves to be taken at full strength before any rebuttal.
Working through it honestly
The logical core is untouchable. Given "revenue missed forecast," you cannot deduce whether the fault lies in the market-size assumption, the sales-cycle assumption, the churn assumption, or the founder's execution. All four, and more, jointly entailed the forecast. The falsified prediction falsifies their conjunction, not any one of them. A partner who says "the market moved, not the thesis" has not committed a logical error merely by saying it. Quine's radicalisation makes this worse, not better: even the belief that the customs rule changed is itself a statement embedded in a web of other beliefs — about the reporting agency, about the translation of the ruling, about its actual effect on freight buyers — any of which could instead be revised to save a different story.
Concede further: Duhem himself limited the claim to mature physical theory, and explicitly carved out physiology, where he thought bundles were shakier and judgement correspondingly less reliable. Venture theses are closer to physiology than to physics — messy, low-N, contaminated by reflexivity, since the act of funding a thesis changes the market the thesis describes. If Duhem thought the argument needed narrowing even inside physics, it should not be smuggled into venture capital as though ambiguity there is somehow easier to close.
And the naive answer — "just watch more data" — does not, by itself, refute the objection. It relocates it. Filings, hiring signals, product telemetry, market-structure indicators: these are themselves reports, produced by instruments (a scraper, a survey panel, a headcount database) with their own auxiliary assumptions. A dashboard showing churn ticking up is not raw truth arriving to settle the dispute. It is another conjunction — churn metric plus definition of active user plus billing-system behaviour — that could itself be the thing that's wrong. Continuous intake does not step outside holism. It adds more nodes to the web.
What survives
Here the objection overreaches, and Quine's own text is the correction. Quine did not conclude that any revision is as good as any other. He insisted on a pull toward minimal mutilation of the web — you save the belief that costs the least elsewhere. Duhem called the analogous instinct in physics bon sens: trained judgement about which part of the apparatus was likely disturbed. Neither philosopher licensed "keep the thesis, blame the world" as a free move. They licensed it as a move that has to earn its keep against everything else you know.
What earns it in venture capital is exactly the kind of streaming, dated evidence the partner in the example didn't have. Filings tell you whether the customs rule actually changed and when. Hiring signals at the buyer's organisation tell you whether procurement headcount contracted before or after the deal cycle in question — sequence that a lagging churn number can't give you. Product telemetry tells you whether the drop-off is concentrated at a specific step in onboarding, which points at execution, not macro. Market-structure data — competitor pricing, category funding volume, substitute product adoption — tells you whether the whole category re-rated or just this one account. None of these settle the question by logic. They settle it the way Le Verrier settled Uranus's orbital residuals in the 1840s: the bundle was Newtonian gravitation plus an assumed planetary inventory, and revising the inventory produced Neptune, found within a degree. The identical manoeuvre applied to Mercury produced Vulcan, which doesn't exist — same logical structure, opposite answer, decided only by decades of further observation, not by choosing the more comfortable revision.
The venture case for OPERA's faster-than-light neutrinos has its analogue too. In 2011 the anomaly sat unresolved for months; it was not further theorising but a logged timing record that found the loose fibre-optic connector and the miscalibrated oscillator. A partner without a logged record of the buyer's budget cycle, the customs filing date, and the churn timeline is in the position of a physicist with the neutrino anomaly and no instrument log: entitled to a guess, not entitled to certainty, and unable to tell a real regime change from a badly connected cable.
The Bayesian will say this is already solved: put priors on the auxiliaries, condition on the miss, and the posterior redistributes blame without any need for new observation. Dorling showed as much in 1979.
That's a real result and it's underrated in this argument. But a prior on "how often does a mid-market logistics buyer's budget shrink for reasons unrelated to our product" is itself an empirical quantity. Where does the partner's prior come from, if not from a record of how often that has happened before, across enough portfolio companies and enough cycles to be more than a hunch? Without that record the Bayesian machinery is real but its inputs are guesses, and the posterior only launders the guess into something that looks computed. Continuous intake — the streams of filings, hiring data, telemetry, structure — is what makes the prior an estimate rather than a stipulation. The formalism is correct. It presupposes exactly the intake in question.
The failure mode named
The characteristic failure in this domain is not an intellectual error so much as an incentive one wearing philosophical clothing. An investing partner defends a thesis for a year after the market it assumed has dissolved, because Duhem-Quine gives them genuine logical cover to do so, and because the partner's own carry, reputation and board seat all sit on the "market moved, not me" branch of the ambiguity. The thesis is holistically underdetermined, so a partner motivated to protect it can always find something else in the web to blame. The tell is when the blame assignment stops updating even as new streams keep arriving — when the filings show the customs rule was reversed eight months ago, hiring at the target accounts has recovered, and the telemetry still shows the same drop-off at the same onboarding step. At that point the ambiguity has been resolved by the data and the belief is being held by something other than evidence.
Objection on regress
The narrower claim
Duhem-Quine is right that no missed quarter refutes a thesis by itself, and right that the choice among competing repairs is never forced by logic alone. What it does not license is indifference between a repair backed by dated, cross-checked streams and one backed by a partner's preference for the story that keeps their fund's stake intact. The gap between hypothesis and evidence is real and permanent. What continuous intake buys is not closure of that gap but a record — filings, hiring, telemetry, structure — against which the partner's chosen repair can be checked, and eventually, if it's wrong, caught. A frozen quarterly deck cannot do this. Neither can a single board meeting's worth of live discussion, once the meeting ends. Only a belief dated to its last confirming observation, held against streams that keep running, gives bon sens something to work with.