Midway through an investment committee presentation, a three-column table appears on screen: downside, base, upside. All three columns move the same two or three variables in the same direction, generally by symmetrical percentages — sales price down ten, volume down ten, operating expense up five. The table traces a narrow, well-behaved band around the base case, shows a return above the hurdle even in the downside column, and the room relaxes at precisely that moment. What will in fact determine the outcome twelve months later is none of those three variables, but the interconnection permit date mentioned in a single line on the fourth slide.
When the table was built says more than what it displays. The sensitivity tab is typically appended after the model has been assembled, the base case calibrated and the answer already known; it functions less as an instrument of interrogation than as a comfort band drawn around a conclusion already reached. Nor is the selection of tested variables random. The variables chosen are those that occupy a discrete input cell, that do not break the model when nudged by a single percentage, and that are intuited in advance not to drive the result below the threshold. The difficult assumptions go untested precisely because testing them would require rebuilding the model.
The pattern has a name — sensitivity-analysis failure, the condition in which the influence of critical assumptions on the output is never examined — and its mechanism arises not from carelessness but from the intersection of three structural conditions. First, the cost of testing falls on whoever builds the model while the benefit accrues to the institution accountable for the decision. Second, the model's architecture determines in advance which assumptions are movable at all. Third, to the degree that the sensitivity table becomes an institutional approval artefact, its function drifts from generating information toward documenting that the process was completed. With all three operating together, the table is produced reliably and teaches nothing reliably.
It is worth recognising that the shortcut is rational under identifiable conditions. Where input volatility is low, counterparty behaviour predictable and the project belongs to a typology executed many times before, examining every assumption independently is an inefficient use of analytical capacity; institutional memory already knows which line item tends to break and where. The difficulty lies not in the shortcut itself but in its persistence once conditions change. Introduce a new jurisdiction, an unfamiliar technology, an untested financing structure or a contractor engaged for the first time, and the sensitivity template inherited from the prior typology no longer carries the decision's actual fragility.
Layered onto this is the fact that a significant proportion of critical assumptions never inhabits a numeric cell in the first place. Which basket the escalation index is tied to in the contract, whether the offtaker's credit quality holds through the drawdown period, whether the LD cap is sufficient to absorb realised delay, how many months the schedule shifts if the objection window in the permitting process opens — all of these reside in a contract annex or a legal opinion rather than on the model's input sheet, and consequently never appear in the table. To this must be added the structural blindness of single-variable testing: in an actual stress pattern, delay, cost inflation and rate movement do not arrive independently but as three faces of the same macro wave, and a table that moves each by ten percent in isolation cannot by construction observe that combined movement.
The balance-sheet consequence surfaces most sharply in debt sizing. Because project debt capacity is solved backward from free cash flow against a targeted DSCR, an untested revenue or schedule assumption translates directly into the scale of the capital structure; when the assumption fails, what emerges is not a lower return but a ratio sitting below the covenant threshold. Given that credit committees conventionally require debt service coverage to remain within a defined band, even a modest deviation from the base case converts into a restructuring discussion, an equity cure call on the sponsor, or a distribution lock-up. The absence of genuine sensitivity work is therefore not an analytical gap but a miscalibrated capital structure.
The second cost is one of timing. An untested assumption surfaces not when it is cheapest to correct — before FID, while the structure remains adjustable — but when correction is most expensive: after first drawdown, with the EPC contractor mobilised and long-lead equipment ordered. At that stage the only remaining buffers are the reserve accounts and the contingency line, and because the calibration of both is typically derived from the same untested assumption set, both prove systematically inadequate at the moment of breach. Setting contingency as a percentage of total installed cost is common practice; that percentage, however, is generally inherited from the previous project's template rather than from this project's actual breaking points.
On the M&A side the same tendency is translated into valuation language. The moment a buyer's due diligence team identifies which assumption in the target's projections was never examined, it converts the finding into a structural discussion rather than a price discussion: the uncertainty carried by the untested assumption is written into the earn-out trigger, the escrow percentage, a condition precedent, or an expanded scope of representations and warranties. The result is a transaction architecture in which headline price is preserved while cash receipt is distributed across time and contingency; for the seller, that means paying for a single unexamined assumption not at closing but over the three years that follow. What determines valuation at this point is not performance itself but the ability to demonstrate which assumptions that performance rests upon.
The mechanism that neutralises this tendency is not individual rigour but a design composed of four separable components. The first is maintaining the assumption register independently of the model and in advance of it, so that the owner, source, last verification date and verification method of every critical assumption live in the decision record rather than inside a spreadsheet. The second is ranking criticality before the model is built — the question of which variable, moving across which range, reverses the direction of the decision is answered before any number is produced, and testing capacity is allocated according to that ranking. The third is reverse stress testing, in which the question is reframed from what happens to the output if the input moves ten percent into how far the input must move before the output breaches its threshold, and whether that movement is plausible within the observed range. The fourth is defining scenarios with correlation, since a single combined case in which schedule, cost and financing conditions move together carries more information than three separate univariate tests.
The significance of these components shifts by party. For the sponsor the question is the probability of an equity cure call and the point at which the distribution lock-up engages; for the senior lender it is which input movement breaches the debt service threshold and how ordinary such a movement is within the observed band; for the EPC contractor it is whether the schedule assumption is consistent with the LD structure; for the offtaker it is whether the index underlying the price formula continues to carry the same risk over the long term. The same sensitivity table needs to answer four different questions for four different readers, and a single set of columns cannot address all four simultaneously.
BEIREK's concern at this point lies not in rebuilding the model but in establishing the record and cadence discipline surrounding it. Across the project development and financing mandates we manage, the assumption register is a distinct document opened on the first day of the decision file, carried by a single owner through to closing, holding each assumption alongside its source, verification method and breach threshold, and travelling independently of any particular model version. It is accompanied by a breach threshold table in which the distance each critical input must travel before the DSCR or return hurdle is violated appears on a single page, so that the discussion proceeds against the credit committee's real limits rather than against percentages selected by whoever built the model.
The second intervention is separating testing from advocacy. When the team constructing the model is also the team charged with producing the counter-argument, sensitivity analysis becomes structurally an exercise in confirming its own conclusion; the examination of critical assumptions is therefore assigned to a role distinct from the line advancing the transaction, and operated on a defined cadence — before FID, before first drawdown, and upon any material contract amendment. The same discipline extends to the contractual side: assumptions that do not live in cells — permitting schedule, escalation index, counterparty credit quality, LD cap — enter the register with a numeric range and become visible on the decision's breach map even where they remain absent from the model tab.
The quality of an investment decision is measured not by the number of scenarios produced but by whether the institution has written down and debated which assumption, once it moves, reverses the direction of the decision; where that has not been written down, the project is still financed and still built, and the breaking point is simply learned at a later and more expensive date.