Two ways to read a term sheet in trouble
An investing partner holds a thesis: direct-to-consumer brands, sold through paid social, would beat the retail incumbents on unit economics because acquisition was cheap and repeat purchase would compound. The fund put money behind six companies on that thesis between 2018 and 2020. By late 2021, filings from public comparables showed customer acquisition cost on Facebook and Instagram up roughly 35–45% year on year. Hiring signals showed brand-marketing heads leaving two of the six portfolio companies within the same two quarters. Product telemetry — repeat purchase rate, the number the thesis actually depended on — was falling at four of six. Market structure had consolidated: Amazon's share of the categories in question had grown, not shrunk, exactly the outcome the thesis said would not happen.
The partner kept the thesis for another year. Two more cheques went out under it. This is the concrete version of the abstract problem: a thesis defended long after the market it assumed has dissolved. The interesting question is not whether this happened — it happens routinely — but what kind of failure it is, and whether the opposite instinct, holding a position through a wave of bad-looking data, is not sometimes the correct one.
The conviction defence
Venture returns are famously non-normal: a small number of positions produce nearly all the fund's outcome, and the modal experience of a good decision is a long stretch that looks like a bad one. A partner who abandons a thesis at the first quarter of rising CAC would have sold every company that survived a hard patch before working. Airbnb's unit economics looked wrong through most of 2009. The correct response to a noisy signal against a genuinely strong prior is often to hold, not to revise. This is not stubbornness dressed up; it is a defensible epistemic stance with a name — conservatism under holism, the position that no single observation refutes a theory because any observation can be absorbed by adjusting something else in the surrounding structure. Quine and Duhem made the general case; venture investing supplies a domain where it is not academic. A single quarter of rising ad costs can be a platform pricing change, not a verdict on the category. A partner who revises the thesis every time one stream ticks the wrong way will be whipsawed out of the position that would have paid for the fund.
The intake case against it
Set against that: none of the four streams in question — filings, hiring, telemetry, market structure — was noisy in isolation. They moved together, for four quarters, across four of six portfolio companies, in the same direction, and the direction matched a structural explanation the partner had access to the whole time: paid social acquisition costs were rising because platform inventory was being repriced by an auction with more entrants bidding for the same attention, a mechanism that does not revert on its own. This was not one anomaly to be absorbed by holism. It was convergent evidence from independent sources pointing at a load-bearing assumption of the thesis — cheap acquisition — and the assumption was the thing that broke.
The partner saw the data. Every memo from the following four quarters cites it. What happened to it on arrival is the diagnostic fact: each stream was routed into a defence of the thesis — "CAC is rising for everyone, so relative position improves," "hiring churn reflects the labour market, not the company" — rather than into a test of whether the thesis's central mechanism still held. That routing is the signature of belief perseverance, and it is worth being precise about what that term does and does not mean here.
What a 1960 questionnaire and a 1975 debriefing room have to do with a partner's memo
Milton Rokeach built the dogmatism scale in 1960 to measure closed belief systems independently of what the belief actually was: the diagnostic feature was structural, a system that admits new information only when it confirms what is already held, and rejects the source rather than revise the conclusion. Leon Festinger's dissonance research in the same decade supplied the motivational engine — revision is costly, and people protect the cost by discrediting the messenger. The sharpest demonstration came in 1975, when Ross, Lepper and Hubbard gave subjects fabricated performance feedback, then told them plainly that it had been invented. Subjects' self-assessments, built on the fake feedback, did not return to baseline. The evidence had been fully withdrawn and the belief it produced remained.
That result matters here because it isolates a residue that has nothing to do with wanting to be right. Subjects were not defending status against the experimenter; they had already been told they were wrong. What persisted was the explanatory structure the fabricated evidence had let them build — the story of why they were the kind of person who performed that way — which stayed available long after its evidentiary support was gone. A partner's thesis memo is exactly that kind of structure: a causal story about why cheap acquisition would compound into a moat, built once, then defended by fitting each new stream to the existing story rather than testing the story against the streams.
Two objections worth taking seriously
Belief perseverance is a motivational phenomenon. A partner defends a thesis because reversing it in month eighteen means writing down a position in front of the investment committee, marking a loss earlier than necessary, and admitting a partner-level misjudgement that colleagues will remember at the next fund close. Calling this "dogmatism" imports a psychological pathology onto what is really a career incentive. Fix the incentive — blind memos, mandatory quarterly re-underwriting — and the perseverance disappears without anyone needing a theory of belief.
This is largely right, and it is not a small concession. Carry structures, IC politics and the asymmetric visibility of write-downs versus write-ups do enormous work in explaining why a thesis outlives its evidence in this industry specifically. But the Ross-Lepper-Hubbard result is useful precisely because it shows a non-motivational residue underneath the motivational one: even with the incentive to abandon a discredited belief removed entirely — subjects had nothing left to protect once debriefed — the belief persisted because the explanatory story remained available while the evidence had been withdrawn. A partner facing no committee, no fund-close optics, still keeps a thesis for the same structural reason: the causal story about the market was built first, and each new stream gets fitted to it because that is cheaper than rebuilding the story from the current data. The incentive amplifies the failure. It does not manufacture it from nothing.
No anomaly should collapse a strong prior. A thesis built on genuine diligence, real founder quality and a large market should not be abandoned on one bad quarter of ad pricing, and a partner who re-underwrites at every wobble in the streams will sell early and often, capturing none of the tail the fund exists to catch.
Also granted, and it is the stronger of the two objections because it is correct as stated. The failure at the six DTC companies was not that the partner showed insufficient conservatism in month one. It was that by month twelve, four independent streams agreed on a structural cause, and the thesis had by then had a full year to be tested against that cause rather than defended from it. Rational conservatism presupposes exactly the kind of continuous, provenance-tagged intake this argument is about — a partner needs to know not just that CAC rose, but whose numbers, over what period, corroborated by how many independent sources, before deciding whether an anomaly is noise or verdict. Conservatism without provenance is not caution. It is just delay.
Where the streams sit in the lineage
| generation | what the partner would have | what breaks it |
|---|---|---|
| Large Language Model | a market map fixed at the corpus cutoff — the DTC thesis as understood circa 2019 | post-2019 CAC inflation cannot be seen; it happened after intake closed |
| Large World Model | a live dashboard during diligence — this quarter's filings and telemetry, in the room | the correction is real but ends with the diligence sprint; month fourteen is invisible again |
| Large Universe Model | filings, hiring, telemetry and market structure streamed continuously, each claim tagged to its source and its date | nothing — the position is that revision becomes structurally possible, not that it becomes automatic |
The narrow resolution
Neither position collapses into the other. The conviction defence is right that non-normal returns require holding through noise, and a fund that re-underwrites at every quarter's wobble will sell its own outliers before they pay off. The intake case is right that four convergent, independently sourced streams pointing at the same structural mechanism for four consecutive quarters is not noise, and that the partner's failure was not conviction but the routing of disconfirming evidence into defence rather than test. The resolution narrows the original thesis rather than vindicating either side: hold through disagreement between streams, revise when streams that were previously independent start agreeing against you, and require the memo itself to name, in advance, what pattern of convergence would count as refutation. That is not a cure for dogmatism. It is a specification for the one case — sustained, cross-source convergence against a load-bearing assumption — where perseverance stops being conviction and starts being the thing Rokeach had a name for.