7.1What Knight's grouping needs
Read for its conditions rather than its limit, Knight's Chapter VIII asks three things of a pool before the average can be relied on. The cases must be numerous enough for their errors to cancel. They must not fail together, or their number counts for less than it seems. And the person relying on the average must last long enough for the cancelling to appear. Knight states the first condition at length and with confidence, extending it even to cases that share no visible kinship. He states the second and the third in passing, the second in a sentence about fires and the third in a sentence about working lives.
The two records each turn on one of the passing conditions. Tail loss behind average refuses an average because the account must survive to collect it, which is the third. Strategy count diversification refuses a count of names because the named strategies fail together under stress, which is the second. Neither refuses the first. Both records accept that pooling works when its conditions hold; both decline to assume the conditions hold because the average looks good or the count is large.
This is the sense in which the records say what grouping leaves out. They do not say that averaging unlike cases is wrong. They say that averaging unlike cases is safe only if the cases stay unlike when it matters, and only if the decider is still there when the average forms.
7.2The debate, as the records leave it
On the question that divides the later literature, where the rare loss comes from, the records are silent in a specific way. The first record describes its rare losses as arising in sharp-change intervals, without saying what causes those intervals or whether the strategy contributes to them. The second describes its common losses as arising on days of sharp moves, through shared liquidity, direction and volatility, without saying why the strategies share those factors.
Each record's evidence list leans one way, though. The first asks for the lower tail of the profit and loss distribution, which is a measurement of the past, and then, in its revisit line, warns against trusting the past's worst case. The second asks for common factor exposure and concurrent losses, which are measurements of the pool's internal dependence. Measured against the two accounts, the first record leans toward the account that places the tail outside the class, and the second toward the account that places it inside the pool. The lean is not strong enough to count as evidence for either account. Two synthetic records written for study cannot weigh in a debate about how markets actually behave.
What the records can carry is narrower. They show that a decision written before its outcome can name the conditions of grouping without naming their cause. That is a finding about records, not about markets. It suggests that the question of where the tail comes from may not need to be settled for a decision to refuse an average on sound grounds, since both records refuse on grounds that hold on either account.
7.3What the records cannot carry
Both records are synthetic. They describe no real strategy, portfolio, firm or market event. They are modern summaries written for a decision platform, translated from Korean, and dated by their own text as 2026. Nothing in them is a historical case, and the readings of Knight here are not tested against how insurers, investors or speculators actually behaved in his time or since.
The records end at the moment of decision. They give an example judgment, marked by WSTiger as its own dated reading, and a condition for revisiting it. They do not report what happened next. So they cannot show whether refusing the average was wise, whether the separated correlation table captured the stress dependence, or whether the pre-set budget held. Fischhoff (1975) and Baron and Hershey (1988) describe how knowing an outcome changes the judgment of a decision; a record that ends before the outcome is free of that effect, and also free of any test.
The records name no decider, no bearer of the cost and no deadline. The parent study's ledger confirms this across all 53 records: none names a decider or a deadline, and none lists what is missing. The two records here are typical in this respect. Who would bear the rare loss in the first record, or the forced liquidation in the second, is not stated. Luhmann (1993) separates risk, a possible loss attributed to one's own decision, from danger, a possible loss attributed to something outside one's control; the same loss is a risk to the one who decides and a danger to those affected who did not decide. The records do not say who stands on which side of that line.
7.4The study's counts beside the two records
The parent study's ledger offers one count that bears on the records' options. Across 159 options in 53 records, both coders agreed that 44 reduce the immediate loss. The study's hypothesis that options cutting the immediate loss most often give up evolvability and resilience was read as supported, with 24 cases for and 13 against, a share of 65%, on both strict and lenient readings. For Tail loss behind average, the coders agreed that the first option, adoption by average indicators, reduces the immediate loss and gives up long-term stability and resilience. That fits the case reading: adopting on averages collects the steady gains now and gives up survival later. For Strategy count diversification, the coders did not agree on whether the first option reduces the immediate loss, and the hypothesis was not applicable on either reading.
The revisit conditions of both records were read by code as not observable: neither contains a number, a threshold or a time. Across the study, 5 of 53 (9%) conditions do. The records' conditions name a pre-set budget and a stress-period test without stating either. This is consistent with the study's finding that conditions for revisiting are almost always present and rarely stated in a form that could be checked by an outsider, and with Knight's remark that the factors decisive for a business judgment lie largely inside the person who makes it.
7.5Knight's footnote and the fairness of classes
One passage in Knight's chapter has not been used above and deserves mention because it bears on the question the records leave open, who bears the tail. In a footnote on life insurance, Knight steps outside his main argument to say that, from a social point of view, the classification of risks may be undesirable. Apart from occupational hazards whose cost can be passed to consumers, he can find no sound ground for charging people more because misfortune has handicapped them, and suggests that all might be insured at one rate, or even at rates graded inversely to the risk. He adds at once that only a compulsory state scheme could work this way, since competition forces private insurers to classify as finely as they can.
The footnote turns classification from a technical question into a question of who pays. A fine classification puts the cost of each risk on those who carry it; a coarse one spreads it. The second record's choice between counting names and aggregating shared risk is, among other things, a choice about how finely to classify, and therefore about where within the portfolio the cost of the shared loss falls. The record does not raise this. Beck (1992) describes modernisation as producing risks of its own making, distributed across people who did not choose them, and the portfolio-level collateral demand in the second record is a small instance of a loss passed to lines that did not cause it. Neither the record nor the notes draw that parallel.