Comparing the first draft of a feasibility analysis circulated ahead of an investment committee with the version ultimately approved is frequently more instructive than examining the underlying data itself. Between the two versions no new measurement has been taken, no site survey revised, no supplier quotation refreshed; what has changed is how the same dataset has been cut. The comparison window has been extended from three years to five, the pandemic period excluded as unrepresentative, the unit cost measure shifted from a simple arithmetic mean to a weighted average, and the comparable project set redefined by capacity band rather than geographic proximity. Each of these choices, taken in isolation and challenged directly, is defended with a technically persuasive rationale. Taken together, they account for the greater part of the movement that carries the IRR across the internal hurdle.
The same pattern surfaces in pre-acquisition commercial diligence, where the analyst measuring a target's customer concentration must decide whether to compute it on revenue or on gross profit, whether the threshold is the top five accounts or the top three, and whether affiliated entities within a group are counted as a single relationship or treated separately. These decisions look procedural, and they typically survive only as a footnote in the methodology note, yet they directly determine the measured magnitude of the single most consequential structural exposure the target carries — a revenue base tethered to founder relationships. For one target, four different concentration definitions, all of them defensible, can generate four tables whose implied risk profiles differ by something close to an order of magnitude.
The mechanism at work here is researcher degrees of freedom: the latitude retained by whoever converts data into conclusion, composed of choices each of which is legitimate on its own terms. The source of that latitude is not bad faith but the nature of measurement, since no real dataset announces how it ought to be cut. Which observation qualifies as an outlier, which period counts as representative, which subgroup is analytically meaningful, which normalising adjustment applies — all of these remain matters of professional judgement. The difficulty lies not in the existence of such judgements but in when they are exercised; made before the result is visible, they constitute method, and made after the result is visible, they become the instrument that produced the result.
The width of this latitude tends to exceed what most decision-makers intuitively assume. A modest analytical protocol containing five independent binary choices — period, outlier rule, averaging convention, comparator set, normalising adjustment — carries the capacity to produce thirty-two distinct outcomes, every one of which remains methodologically defensible. Once the analysis appears in the final report as a single figure, the existence of that distribution becomes invisible, because the reported confidence interval captures only the sampling uncertainty within the one path actually taken and carries nothing of the uncertainty generated by the choice of path. The uncertainty measure placed in front of the committee is therefore, and systematically, a narrow subset of the uncertainty that genuinely surrounds the decision.
There are conditions under which this same latitude is functional, and ignoring them produces a badly designed intervention. In early-stage exploratory work, cutting the data from multiple angles is precisely the required behaviour; confining an analyst to a single pre-specified slice while attempting to locate where a cost variance concentrates will suppress the very pattern the exercise exists to find. An analyst carrying genuine institutional memory may likewise be applying valuable knowledge when excluding a particular period. The distinction runs between exploration and confirmation: flexibility is productive while a hypothesis is being formed, and corrosive once the object is to test one, because it blurs what has actually been tested. The characteristic institutional failure is the collapse of both phases into one document, one meeting and one presentation.
The first surface on which the institutional cost appears is not the investment decision that turns out to be wrong; that is the outcome rather than the cost. The cost is the defensibility armour that keeps the error invisible for an extended period. A figure produced through accumulated analytical flexibility clears the audit committee, the credit committee and independent review without friction, because a ready and reasonable justification exists for every individual choice embedded in it. The error leaves no marker pointing back at itself. Correction consequently tends to arrive far downstream — after first drawdown, or during the first operating year — at the point where reversal has become expensive and where the available remedies are contractual rather than analytical.
The second surface becomes visible in working capital behaviour and in covenant performance. A credit package's DSCR threshold is calibrated against the model maintained by the lender's independent adviser, and the inputs to that model rest on the sponsor's own analytical choices. Which scenario band the revenue forecast is built upon, over which period seasonal variation has been smoothed, and which averaging convention has been applied to maintenance downtime — each of these produces not a few basis points of difference at the covenant heading but a categorical difference in trigger frequency. The invoice arrives in the first weak quarter: a breach technically originating in an optimistically cut modelling assumption is recorded in the credit file as an operating failure, and it durably narrows the sponsor's negotiating position in subsequent transactions.
The third surface sits directly on the valuation desk. When a buyer's commercial diligence team observes the pattern of definitional choices in the target's own analyses — particularly where the same metric has been computed under different definitions across different reports — it reads this not as an arithmetic error but as an indicator of management reporting maturity. The consequence for transaction structure is predictable: the representations and warranties package widens, a separate specific indemnity attaches to the financial information heading, the escrow ratio rises, and the earn-out trigger is shifted from a negotiated metric onto an audited line item. Analytical flexibility is therefore priced not as a balance sheet item but as an asymmetry within the risk-allocation architecture of the transaction agreement.
The mechanism that neutralises this tendency is not individual rigour, which is generally present already; the difficulty is that the direction of that rigour is being set by the desired result. The effective intervention is a written analytical protocol, fixed before the data is opened, and it comprises four components. First, the definition of the metric to be measured — numerator, denominator, scope and time window — is settled before analysis begins and recorded in a dated note. Second, outlier and exclusion rules are written in advance, with any exclusion applied subsequently listed in a separate annex alongside its justification. Third, the decision threshold — which result constitutes approval and which constitutes rejection — is set before the result is known rather than after. Fourth, abandoned analytical paths are not deleted but retained in a sensitivity table appended to the principal report.
The fourth component attracts the most resistance in practice and delivers the highest return. A record of the abandoned scenarios presents the decision-maker with a distribution rather than a single figure, and the width of that distribution is the most honest available indicator of a project's real uncertainty. Where defensible methods applied to the same project yield an IRR range contained within two hundred basis points, the decision rests on solid ground and the base case carries genuine information. Where the same exercise spreads the range across seven hundred basis points, the informational content of the single reported figure is low, and the item warranting committee time is not the number at all but the structural uncertainty responsible for the width of the band.
In the project and transaction processes BEIREK runs, this intervention is operated by recording the analytical protocol before the data collection phase begins. As a feasibility study, model review or commercial diligence opens, the definitions of the metrics to be measured, the exclusion rules and the decision thresholds are fixed in a dated protocol note; every subsequent departure is entered into the revision record of that same note together with its rationale and its quantified effect on the result. When the model is delivered, what sits in front of the decision-maker is not only the base case but the outcome band generated by the alternative combinations of choices that the protocol itself identified as defensible at the outset.
The second mechanism is the separation of exploration from confirmation at the document level. Exploratory work, in which the data is cut freely and repeatedly, is carried as a distinct working note that does not travel to the committee; the confirmatory analysis that does travel tests the hypothesis generated in that note against a protocol fixed in advance. This separation preserves the productivity of exploration while making visible the constraints under which the committee's number was produced. In practice it does not narrow the analyst's freedom in any meaningful sense; it simply detaches the moment at which that freedom is exercised from the moment at which the answer becomes known.
What determines the quality of an investment decision is, more often than not, less a matter of how deep the analysis went than of which question was fixed, and when, by the party conducting it. The question an institution might reasonably put to itself is narrow and answerable from its own files: across the analyses that reached the committee over the last twelve months, in how many was the measurement definition and the decision threshold available in writing before the result was seen?