11
Pools with a stated basis
3
Estimating methods
12.4%
Provision the mix implies
12%
Reserve authorized
Contents
1. An Estimate Without a Method Is a Number Somebody Made Up
This program was authorized at $243,040,000. That figure has been quoted in every artifact in this suite, and none of them says where it came from. This document does, pool by pool, because a number without a stated basis cannot be challenged, cannot be improved, and cannot be defended when it turns out to be wrong.
| A basis of estimate records | So that |
|---|---|
| The method | A reviewer can tell whether the approach suits the thing being estimated. Bottom-up on an undefined scope is precision theater; analogous on a fully specified study list is laziness. |
| The cost driver | Everyone knows what would move the number. If the driver is participants dosed, a two-month schedule slip that changes no participant count barely moves the cost — and a protocol amendment adding one visit moves it a great deal. |
| The source | The estimate can be re-derived by somebody else. “Vendor quotation” and “experience” are different kinds of claim. |
| The accuracy class | The reader knows how much of the number to believe. A figure quoted to the dollar from a Class 4 estimate is asserting a precision nobody has. |
The most common estimating failure is not being wrong. It is being unable to say why.
Every estimate in a development program is wrong to some degree, and that is tolerable and expected. What is not tolerable is a variance nobody can explain, because an unexplained variance cannot be distinguished from a scope change, a control failure, or a bad estimate — and those three have completely different responses.
The basis of estimate is what lets a program tell them apart eighteen months later, when the person who built the number has moved on and all that survives is the number.
Every estimate in a development program is wrong to some degree, and that is tolerable and expected. What is not tolerable is a variance nobody can explain, because an unexplained variance cannot be distinguished from a scope change, a control failure, or a bad estimate — and those three have completely different responses.
The basis of estimate is what lets a program tell them apart eighteen months later, when the person who built the number has moved on and all that survives is the number.
2. Three Methods, Chosen by What Is Knowable
| Method | Pools | Value | When it is the right choice |
|---|---|---|---|
| Parametric | 6 | $172,300,000 | A unit rate multiplied by a countable driver. Right where the driver is fixed by protocol and the rate is contracted — which describes most clinical cost in this program. |
| Bottom-up | 2 | $17,700,000 | Built from a defined list of things. Right where the scope is genuinely enumerable: a study list, a batch count. Wrong wherever the list is the thing you are uncertain about. |
| Analogous | 3 | $27,000,000 | Scaled from comparable programs. The honest choice where nothing is fixed yet — and it should be declared rather than dressed up as something firmer. |
The method is dictated by the thing, not by the effort available.
Phase 3 is half the program and it is estimated parametrically, not bottom-up. That looks like a shortcut on the largest pool in the budget and it is the opposite: the cost of a pivotal trial is participants multiplied by a per-participant rate multiplied by duration, and every one of those three is either protocol-fixed or contracted. Decomposing it further would produce more lines and no more accuracy.
Conversely, early CMC development is only $6,200,000 and carries the weakest basis in the program, because process development cost scales with the problems encountered — and problems encountered are not knowable in advance by any method.
Phase 3 is half the program and it is estimated parametrically, not bottom-up. That looks like a shortcut on the largest pool in the budget and it is the opposite: the cost of a pivotal trial is participants multiplied by a per-participant rate multiplied by duration, and every one of those three is either protocol-fixed or contracted. Decomposing it further would produce more lines and no more accuracy.
Conversely, early CMC development is only $6,200,000 and carries the weakest basis in the program, because process development cost scales with the problems encountered — and problems encountered are not knowable in advance by any method.
3. The Basis, Pool by Pool
| Pool | Amount | Method | Cost driver | Source | Class | What the estimate rests on |
|---|---|---|---|---|---|---|
| Nonclinical pharmacology & GLP toxicology | $8,500,000 | Bottom-up | Study count and species | CRO quotations against a defined study list | 2 | Fully scoped before Gate 1. Study-by-study pricing from three bidders; the estimate is close to a contract. |
| CMC — early development | $6,200,000 | Analogous | Molecule class | Two prior peptide programs, adjusted for step count | 4 | ⚠ The weakest early estimate in the program. Process development cost scales with problems encountered, and problems encountered are not knowable in advance. |
| IND-enabling package & submission | $3,800,000 | Parametric | Document count | Internal hours per module × blended rate | 3 | Assembly work with a well-understood shape and low historical variance. The module list is fixed by regulation, so the only real uncertainty is review cycles. |
| Phase 1 clinical | $8,500,000 | Parametric | Participants × per-participant cost | Unit pricing from the Phase 1 unit | 2 | Single site, fixed protocol, 64 participants. One of the two most reliable estimates here. |
| Phase 2 clinical | $38,000,000 | Parametric | Participants × per-participant cost | CRO unit pricing, blended across regions | 3 | 480 participants at a known per-participant rate. The rate is contracted; the participant count is protocol-fixed. |
| Phase 3 clinical | $109,100,000 | Parametric | Participants × per-participant cost × duration | Meridian master services agreement unit rates | 3 | ⚠ 50% of the program in one pool. Reliable per participant and sensitive to anything that changes the participant count or the visit schedule — a single added assessment moves this pool by more than most pools cost. |
| NDA preparation & review support | $12,400,000 | Analogous | Submission complexity | Two prior NME submissions, escalated | 4 | The Integrated Summaries are new analyses rather than compilations, and their effort is hard to predict before the data exists. |
| CMC scale-up & process validation | $9,200,000 | Bottom-up | Batches and validation runs | Aldergate batch pricing plus internal oversight hours | 3 | Batch pricing is firm. The number of batches is not, if a validation run fails. |
| Regulatory affairs | $7,600,000 | Parametric | Internal hours | Roster hours × blended rate | 2 | Internal labor with a known rate and a known headcount. Effectively a control estimate. |
| Program management | $5,300,000 | Parametric | Internal hours | Roster hours × blended rate | 2 | As above — a small pool with a known headcount and a known rate, so the confidence is high and the amount is too small to matter if it is not. |
| External fees, advisors & other non-labor | $8,400,000 | Analogous | Program size | Historical percentage of comparable programs | 4 | A residual category by construction, and residual categories are always the least well estimated. |
| Base program | $217,000,000 |
Two rows deserve attention. Phase 3 clinical is 50% of the base program in a single pool — reliable per participant, and sensitive to anything that changes the participant count or the visit schedule. A single assessment added to the protocol moves this pool by more than most other pools cost in total.
And external fees and advisors is a residual category. Everything that did not fit elsewhere ended up in it, which is exactly why it carries the weakest method. Residual categories are always the least well estimated, and naming one as residual is more useful than pretending it was built up.
4. Accuracy Class, and Why Effort Cannot Improve It
Accuracy class describes how much of a number to believe. It is limited by how much of the thing is knowable, not by how much work goes into estimating it.
| Name | Expected accuracy | Provision | This program | Share | When it applies | |
|---|---|---|---|---|---|---|
| Class 5 | Order of magnitude | -50% / +100% | 35% | — | — | Analogous only. Appropriate at Gate 0 for anything past the next stage. |
| Class 4 | Study estimate | -30% / +50% | 20% | $27,000,000 | 12% | Parametric from comparable programs, few parameters fixed. |
| Class 3 | Budget estimate | -20% / +30% | 12% | $160,100,000 | 74% | Parametric with the driving parameters fixed. Most of this program at Gate 4. |
| Class 2 | Control estimate | -15% / +20% | 8% | $29,900,000 | 14% | Bottom-up on a defined scope, vendor pricing in hand. |
| Class 1 | Definitive | -5% / +10% | 5% | — | — | Executed contracts and a fixed protocol. Rare before the final stage. |
No amount of estimating effort moves a Class 4 pool to Class 2 while its scope is undefined.
This is the point most often missed when a program is asked to “tighten up” its numbers. A finance function under pressure will ask for a more detailed estimate of NDA review support, and it can be given one — broken into modules, hours and rates, to two decimal places. It will still be Class 4, because the Integrated Summaries are new analyses whose effort depends on data that does not exist yet.
The detailed version is not more accurate. It is more confident, which is worse, because it invites the reader to plan against a precision that is not there.
This is the point most often missed when a program is asked to “tighten up” its numbers. A finance function under pressure will ask for a more detailed estimate of NDA review support, and it can be given one — broken into modules, hours and rates, to two decimal places. It will still be Class 4, because the Integrated Summaries are new analyses whose effort depends on data that does not exist yet.
The detailed version is not more accurate. It is more confident, which is worse, because it invites the reader to plan against a precision that is not there.
14% of this program sits at Class 2 and 12% at Class 4. The Class 2 portion is largely internal labor and contracted unit pricing; the Class 4 portion is the work whose shape depends on results not yet obtained. That distribution is a property of where the program stands, and it will improve on its own as the trials run.
5. Contingency Is an Output, Not an Input
The charter authorizes contingency of $26,040,000, which is 12% of base. Twelve percent is a suspiciously round number, and the obvious question is whether it was calculated or chosen.
| Value | |
|---|---|
| Provision implied by the estimate mix | $27,004,000 |
| As a percentage of base | 12.4% |
| Authorized in the charter | $26,040,000 — 12.0% |
| Difference | 964,000 (0.4% of base) |
Each accuracy class carries a provision, and the weighted result is what the reserve should
be.
A Class 2 pool carries 8%, a Class 3 pool 12%, a Class 4 pool 20%. Applied pool by pool across this program's mix, the implied provision is 12.4% — which the authorized 12% rounds down to.
That is the right order of operations, and it is the one that gets reversed most often. A reserve set as a round percentage first and justified afterward tells you nothing about the estimate underneath it. A program with the same budget and a fully contracted scope would need about half this reserve; one still at Class 5 would need three times it. The percentage is a property of what is known, not a convention.
A Class 2 pool carries 8%, a Class 3 pool 12%, a Class 4 pool 20%. Applied pool by pool across this program's mix, the implied provision is 12.4% — which the authorized 12% rounds down to.
That is the right order of operations, and it is the one that gets reversed most often. A reserve set as a round percentage first and justified afterward tells you nothing about the estimate underneath it. A program with the same budget and a fully contracted scope would need about half this reserve; one still at Class 5 would need three times it. The percentage is a property of what is known, not a convention.
This relationship is asserted in the fact base. If a future edit changes the estimate mix — by re-classing a pool, or by adding one — and the derived provision drifts more than four points from the authorized reserve, nothing in this suite builds. The reserve and the estimate quality are not allowed to disagree silently.
6. How the Estimate Matures
The same program, estimated at five different points, gets better without anybody estimating harder.
| Gate | Best class achievable | Why |
|---|---|---|
| Gate 0 | Class 5 | Order of magnitude for everything past Stage 1. The program was authorized on a number that could have been wrong by a factor of two in either direction. |
| Gate 1 | Class 4 | Nonclinical actualized; clinical still analogous. |
| Gate 3 | Class 3 | Phase 2 contracted; Phase 3 parametric on agreed unit rates. |
| Gate 4 | Class 3 | Phase 3 protocol fixed and the CRO rate card executed. This is as good as it gets before the trial runs. |
| Gate 5 | Class 2 | Everything but review support is actual or contracted. |
At Gate 0 this program was authorized on a number that could have been wrong by a factor of two
in either direction.
That is not a failure of diligence. At candidate selection there was no protocol, no site count, no CRO, no manufacturing route at scale, and no dose — a Class 5 estimate is the only honest thing available, and the response to it is not more analysis but a smaller commitment.
Which is exactly what stage-gate funding does. The first tranche is $17,400,000, and it is authorized against the worst estimate the program will ever have. By the time $125,800,000 is committed at Gate 4, the protocol is fixed and the rate card is executed. The size of the commitment tracks the quality of the estimate, and that is the whole design.
That is not a failure of diligence. At candidate selection there was no protocol, no site count, no CRO, no manufacturing route at scale, and no dose — a Class 5 estimate is the only honest thing available, and the response to it is not more analysis but a smaller commitment.
Which is exactly what stage-gate funding does. The first tranche is $17,400,000, and it is authorized against the worst estimate the program will ever have. By the time $125,800,000 is committed at Gate 4, the protocol is fixed and the rate card is executed. The size of the commitment tracks the quality of the estimate, and that is the whole design.
7. What This Program Estimated Badly
Recorded here rather than in the closure report, because a basis of estimate that only documents successes is not a basis of estimate.
| Estimate | What was assumed | Why it was wrong |
|---|---|---|
| Site activation duration | 18 weeks from selection to green light | ⚠ Not a cost estimate at all — a duration estimate that drives cost. It was built from an average across prior programs without adjusting for the fact that this trial activates 260 sites simultaneously rather than sequentially, which contends for the same contracting and ethics-review capacity. |
| Early CMC development | Analogous from two prior peptide programs | The comparators had six- and seven-step syntheses that transferred without incident. Analogous estimating carries the assumption that the comparator was typical, and a method transfer that fails is not typical — it is the tail. |
| External fees and advisors | A historical percentage of program size | A residual category estimated as a percentage is a placeholder with a decimal point. It was adequate, and it was adequate by luck rather than by method. |
The most consequential estimating error in this program was not about money.
Site activation was estimated in weeks, not dollars, and it appears in no cost pool. It is nevertheless the estimate that has done the most damage, because activation duration sits upstream of enrolment, enrolment sits upstream of the last participant, and the last participant sits upstream of the filing date.
A basis of estimate that covers only the cost baseline has documented the less dangerous half of its assumptions. Duration estimates deserve the same treatment — method, driver, source, class — and in this program they did not get it.
Site activation was estimated in weeks, not dollars, and it appears in no cost pool. It is nevertheless the estimate that has done the most damage, because activation duration sits upstream of enrolment, enrolment sits upstream of the last participant, and the last participant sits upstream of the filing date.
A basis of estimate that covers only the cost baseline has documented the less dangerous half of its assumptions. Duration estimates deserve the same treatment — method, driver, source, class — and in this program they did not get it.
The correction is recorded rather than acted on retrospectively: the integrated schedule now carries an explicit lead-time contract for Phase 2, asserted at build time. That mechanism exists because of this gap, and it covers durations rather than costs.