There is a recurring behavior around the table whenever the commissioning programme of a new production line is discussed: the operations team acknowledges the low yield of the early months in realistic terms, often stating it explicitly in its own presentation, and yet, several pages into the same deck, unit cost is assumed to have settled at the target level from the sixth month onward. No contradiction is felt between the two statements, since both are true in their own terms — yield is low at the outset and improves with time. The contradiction lies in the question that was never asked, namely at what pace and with what degree of smoothness that improvement will actually occur. The same pattern shows up in a supplier transition, an ERP implementation, an assembly shop opened in a new geography, or a construction method an EPC contractor is executing for the first time; what is being estimated is not the existence of the gain but its calendar, and the error runs consistently in one direction.
The second observation is subtler. When an organization performs the same work a second time, it tends to compress the first instance's learning curve in its own memory; looking back, a six-month settling period on the first project is recalled as a three-month stretch of difficulty. This compression is not a matter of bad faith, and it is in fact documented: project closure reports narrate success and recovery rather than delay. The plan for the second project therefore draws on the narrative of the first rather than on its record, and optimism becomes incrementally more institutionalized with each cycle.
The behavior has a name — learning-curve underestimation, the systematic over-optimism regarding both the duration and the magnitude of efficiency gains — and its mechanics are fed by three distinct layers. The first is the mathematical form of the curve itself: unit cost declines against cumulative output, not against elapsed calendar months. Plans, however, are constructed on a calendar, because budgets and financing operate on one. When production volume falls short of forecast, the learning curve slides back in exactly the same proportion, which means weak demand corrupts the cost side simultaneously with the revenue side — two effects that most models continue to treat as independent variables.
The second layer concerns the shape of the progression, which is stepped rather than smooth. In live production environments, yield plateaus until a bottleneck is resolved, jumps when it is, and flattens again at the next constraint. Changeover duration, the throughput of a single inspection station, a supplier's minimum batch size, the placement of the maintenance window — each generates a plateau of its own. The divergence between the gentle slope assumed in the model and the staircase produced on the floor accumulates in timing rather than in averages; even where the year-end figure approaches target, the question of which month produced at which cost yields an entirely different cash profile.
The third layer is the one most often overlooked: learning accrues in people, not in the institution. The intuition earned on a line — which sound precedes which failure, which raw material lot demands which adjustment — is carried by the shift supervisor and a handful of key operators. In a facility with high turnover, the curve partially resets in proportion to that turnover, and the organization enters its second year holding only a portion of what the first year taught it. This tendency is nonetheless functional and, under particular conditions, entirely rational: assuming the gain optimistically at the outset opens the way to investments that would otherwise never be initiated, and protects the organization from abandoning a position prematurely. The difficulty lies not in the shortcut itself but in its persistence after it has entered a commitment-generating document — a price quotation, a delivery programme, a credit model.
The first surface on which the institutional cost appears is generally not the unit cost table, since variance there can be legitimately carried for several months under the heading of commissioning effects. The cost surfaces first in price. A sales organization signs a twelve-month supply contract on the basis of a unit cost that becomes valid from month six, and when that cost is actually achieved in month nine, the first three quarters of the contract operate at a structurally negative margin. Written into a delivery schedule, the same assumption produces a harsher outcome; where the liquidated damages ceiling has been tied to a ramp calculated on the most optimistic production assumption in the contract, the LD cap absorbs not the deviation of the learning curve but only its first half.
The second surface is working capital, and the effect there is typically larger. A learning period means elevated scrap, elevated rework, and elevated safety stock, all three of which tie up cash simultaneously. When the yield assumption is optimistic, the inventory turnover target is set optimistically alongside it, and the financing structure is accordingly sized below the working capital that will genuinely be required. The result is a company experiencing a cash squeeze precisely while its product is improving — a facility moving in the right direction operationally while approaching a covenant threshold financially. On the credit side this reads as a DSCR calculation coming in tighter than expected in the first measurement periods, and a tight first measurement permanently alters the tone of the lending relationship even where subsequent periods normalize completely.
The third surface emerges in valuation. During diligence on a facility or a business unit, the buy side examines not the target unit cost but the track record of reaching target unit cost. That a company has been through three comparable ramps and settled later than planned in all three is not, by itself, a red flag; what generates the signal of a systematic estimation bias is the absence of a record of those ramps and the attribution of each shortfall to a discrete external cause. In such circumstances, rather than negotiating on the multiple, the counterparty typically intervenes in the structure instead — an earn-out tied to a yield threshold, a production test as a condition precedent to closing, or an escrow tranche carved out for the ramp period. Each of these defers the seller's cash receipt and proves considerably more expensive than a price concession.
The mechanism that neutralizes this tendency is not individual caution; an instruction to estimate more conservatively dissolves within a few quarters, because it leaves the incentive structure of whoever owns the estimate untouched. The intervention that holds consists of three components. The first is keeping the learning assumption as a separate and visible item in the model: not a slope buried inside the unit cost row, but an explicit proposition anchored to volume rather than to the calendar, stated as unit cost reaching a defined level once cumulative output reaches a defined quantity. The second is comparing that proposition against realized production data on a fixed cadence — monthly rather than quarterly, since a quarterly rhythm reveals the first deviation only in the second quarter. The third is binding the deviation to a decision: where the curve trails the target, the thresholds at which pricing policy, capacity commitment, and the financing plan are respectively triggered are written down in advance.
The mechanism BEIREK establishes in capital-intensive projects operates precisely along this axis. For the commissioning and ramp period, the yield assumption embedded in the business plan is separated from the contract and the model and converted into an assumption record in its own right; that record carries who made the assumption, which reference facility or which prior ramp it rests on, and which cumulative volume threshold it is anchored to. Deviation arising during the ramp period thereby becomes the revision of a recorded proposition rather than an argument over responsibility, and the organization orients toward calibrating the deviation instead of defending it.
The ramp period is likewise operated not as a single pass-or-fail moment but as a sequence of predefined thresholds: first stable production, first full shift, first full month, attainment of a specified proportion of design capacity. At each threshold the realized unit cost is placed alongside the assumed curve, and the deviation is separated into three distinct lines — bottleneck-driven, supply-driven, and people-driven. That separation matters, because the correction time and correction cost of the three differ entirely; people-driven deviation closes through retention and documentation policy, supply-driven deviation through batch sizing and inventory policy, and technical deviation through engineering intervention. The same architecture operates on the commercial side as well: which threshold must have been passed before a price commitment is given is settled before contract negotiation, not during it.
The difference between believing in the existence of a learning curve and believing in its calendar is, in most capital-intensive projects, the difference between profit and loss. Where an organization does not maintain a record of the realized curves of its prior ramps, the estimate it makes on the next project is not a forecast but an aspiration; and an aspiration, at the moment it converts into a price commitment, becomes an option written in the counterparty's favor. The operative question is not whether the efficiency gain will materialize, but whether it is known today which commitment breaks first when the timing of that gain shifts.