Confirmation Bias: When the Thesis Writes Its Own Evidence

The vocabulary of the questions asked in the second round of an investment committee file differs structurally from the vocabulary of the first. Early questions tend to test whether the project is sound; once the file acquires an internal owner, the questions quietly change direction and begin measuring how sound it is. Nothing in the minutes records this shift, and no procedure prescribes it — it becomes visible only in the ordering of agenda items, in the composition of the supplementary request list sent to the data room, and in which cells of the model are opened to sensitivity analysis. Placing the first and third versions of the same file side by side, one observes that the overwhelming majority of added pages constitute further corroboration of the core thesis, while the items that threaten it, rather than expanding, thin into a footnote.

A comparable pattern repeats in reference calls conducted during EPC contractor selection. The reference list is furnished by the contractor, the interview questions are constructed around capability headings the technical team already regards favourably, and the output of the conversation is — predictably — positive. The counterparty who is never called, meaning a prior employer who terminated the contract or pressed the LD cap, is absent from the list and therefore never enters the assessment at all. The same mechanics operate in supplier pre-qualification, in stakeholder consultations during land development, and in customer references supplied by an acquisition target; in each instance the evidence set is determined as much by whoever defines the pool as by the evidence itself.

The mechanism beneath this pattern is confirmation bias — the tendency, when testing a proposition, to search for instances that would support it while systematically neglecting instances that would refute it. It operates on two distinct layers. The first is search asymmetry, whereby the mind, testing a hypothesis, naturally generates examples consistent with it. The second, and institutionally the more costly, is evaluative threshold asymmetry: a supporting finding passes through a light methodological review, whereas a contradicting finding is examined heavily for sample size, source reliability, and definitional boundaries. The second layer is particularly hazardous because it presents itself as rigour; the team dismissing the adverse data believes, at that moment, that it is raising the quality standard.

This tendency is not a defect but a shortcut that lowers cost under specific conditions. Given that the space of testable hypotheses is practically unbounded while the decision calendar is measured in weeks, testing every hypothesis symmetrically imposes a burden well beyond available analytical capacity; gathering evidence that supports the standing thesis keeps search cost within a tolerable band and produces adequate outcomes in most low-stakes decisions. The problem lies not in the shortcut but in its persistence once conditions change: where an irreversible commitment, a capital-intensive structure, and a long-tailed contract are involved, the ratio between the cost of a false positive and that of a false negative shifts by several orders of magnitude, yet search behaviour does not track that shift on its own.

Within an institution this tendency does not remain confined to a single mind; to the extent that hierarchy, incentives, and reporting lines align, it hardens into an organisational filter. From the moment a file acquires an internal owner, every interim finding produced at analyst level is framed in the light of the senior-level commentary already offered on the file; an adverse finding, even where it is not suppressed outright, settles into a later slide, a smaller font, and a more heavily qualified formulation. Where the incentive structure is tied to transaction closing, the filter tightens further, since any finding that delays closing imposes a measurable cost on the person who surfaces it, while responsibility for a problem emerging after closing attaches to no identifiable individual.

The first point at which this structure reaches the balance sheet is typically the contingency line. The contingency percentage is calibrated against a scenario set, but where that scenario set is itself derived from the question list the thesis produced, the reserve prices the thesis's own margin of error rather than the actual distribution of risk; a percentage that appeared reasonable before FID is consumed by a single permitting delay in the second year of construction. The second point is the schedule: float placed in the works programme according to the set of assumptions that made the project feasible, rather than according to the historical distribution of delays, means the first slippage displaces the entire chain and distorts the first-year revenue profile on which the DSCR calculation rests. The third is covenant calibration; where the base case modelled by the lender and the case internalised by the sponsor draw on the same source, the breach threshold may well sit inside the genuine band of volatility.

On the M&A side the same mechanism closes into a tighter loop. The seller's data room is already a curated set of documents; the buyer's due diligence question list is drafted from the investment thesis; the thesis therefore produces the question list, the question list produces the findings, and the findings reproduce the thesis. The trace this loop leaves in the transaction structure is generally found not in the price but in the protective layers built around it: which headings the representations and warranties actually cover, over what period the escrow proportion is held, and to which metric the earn-out triggers are indexed. When those layers re-cover the regions of the thesis that were already tested, rather than closing the regions that were never tested, the problem that surfaces after closing arrives precisely along the line where no question was asked.

One structural feature makes this cost particularly difficult to recognise institutionally: rejected alternatives leave no trace in the record. An organisation can track, line by line, the cost of projects it undertook that went badly, yet it can never learn how many early-eliminated projects would in fact have created value; by the same logic, the success of a file approved for reasons that merely looked sound rather than reasons that were sound is entered into the record as a validation of the process rather than as evidence of bias. Measurement design works in the same direction. Where performance indicators are selected from variables consistent with the thesis, the dashboard becomes over time a reflection of the thesis, and signals that would contradict it — customer concentration, single-source supplier dependence, a lengthening sales cycle — are treated as non-existent because they are never measured.

What neutralises this tendency is decision architecture rather than individual awareness, and that architecture has four separable components. The first is the assumption register: critical assumptions are written down at the moment of proposal rather than the moment of approval, together with the identity of whoever advanced them and the evidence relied upon. The second is the falsification criterion: alongside each critical assumption, the observation that would invalidate it is defined in advance, so that when an adverse finding arrives the threshold debate cannot be reopened. The third is the counter-argument role, which becomes effective only where it carries independent access to the data room, its own budget, and a reporting line separated from transaction closing. The fourth is review cadence: the register is reopened at signing, at closing, at first drawdown, and at commercial operation, and the direction in which each assumption has moved is recorded.

BEIREK embeds this architecture into the process itself on the projects it manages. Early in the development phase the critical assumptions underlying the financial model are separated one by one, and two things are recorded in the same document for each: the evidence supporting the assumption and the observation that would invalidate it. Because this second column is maintained in very few organisations, the emergence of an adverse finding ordinarily produces a debate about the validity of the finding; where the column has been written in advance, the discussion moves directly to consequence. The same discipline applies in term sheet negotiation — the reason a counterparty presses a particular clause is recorded alongside the gap in our own thesis that the insistence may be pointing towards, since a counterparty's persistence is frequently the cheapest available indicator of the weakest point in one's own analysis.

The second layer of intervention is cadence. Ahead of FID, a stakeholder pre-mortem is conducted, working backwards from the premise that the project has failed; participants are asked not why the project will succeed but along which line it collapsed eighteen months later, and the headings produced are entered not into the risk matrix but directly into the assumption register. Through construction and operation the register does not close; ahead of each drawdown request its relevant lines are reopened, and the divergence between each assumption and realised data is written down numerically before any commentary is attached. The principal output of this cadence is not a document but institutional memory: which categories of assumption prove systematically optimistic is known on the next project from the record itself rather than from what individuals happen to recall.

The quality of an investment decision is measured not by how internally coherent the file is, but by whether a channel through which incoherence could surface has been deliberately left open. A file that persuades without friction is more often the product of a narrow question list than of a strong thesis; and the question that warrants asking is not what the file proves, but which question the file never asked.