A fixed-scope line item for a research phase is a strange object. To the buyer it arrives as one commitment on one line, and the whole apparatus that produced it, the staffing, the seniority mix, the review allocation, the contingency reserve, the compute that was deliberately moved off the labour line, is compressed out of view. What is left is a figure that either looks reasonable or does not, with no obvious way to interrogate the difference. For a phase we delivered to a mid-sized carbonate client, that figure was 250 person-days, and this paper is about the machinery behind it.
We are writing as the people who built the technical work and sat close to how it was scoped. Most accounts of an applied-AI engagement, ours included, describe the science and the software. What they rarely cover is how a single phase was engineered as a commitment: the person-day ladder, the review line, the carve-out and the gates that turn a stack of roles into something a procurement team can sign. That is the gap this whitepaper fills. If you are the person on the buy side deciding whether a scoped phase is disciplined or hopeful, this is the anatomy you need.
The thesis is simple and we will hold it through every section: a 250-person-day fixed-scope phase is engineered, not estimated. A ladder fixed role by role before anything else, senior review on its own line, infrastructure carved out as a separate package, and a named reserve are what turn a role stack into a defensible commitment. Read the phase as an assembled artefact and its logic becomes legible. Read it as a guess and you cannot tell a good one from a bad one.
The ladder: five roles, 250 person-days, one fixed shape
The phase was staffed by five roles, and the effort behind each was fixed in person-days (mandays, in the workbook's term) before anything else was settled. This ordering matters. The team decided how much of each seniority the work required, then built the commitment around that effort, rather than starting from a total and reverse-fitting the people. The ladder was 20, 60, 30, 70 and 70, summing to exactly 250 person-days.
Reading top to bottom: a principal, carrying quality assurance and technical oversight, at 20 person-days. A senior AI researcher and domain expert at 60. A project manager who doubled as a senior researcher at 30. A PhD-candidate researcher at 70. A junior researcher, also at 70. The shape of that ladder is itself a claim about the work. The bulk of the effort, 140 of the 250 person-days, sits with the two most junior roles, which is what you expect from a build phase where a great deal of the labour is data preparation, experiment running and iteration. The principal's 20 person-days are deliberately scarce: senior oversight is the scarcest resource on any research team, and it is spent on review and direction, not on production.
There is a second signal in the ladder that a buyer should learn to read: the presence of a dedicated quality-assurance role at the top. The principal's 20 person-days are not general management overhead. They are a specific, ring-fenced allocation for review of the research output, and scoping them as a distinct line rather than burying them in a blended team is a statement that the work will be checked by someone senior and that the buyer is contracting for that check explicitly. In a research phase, where the deliverable is a model and a set of results rather than a fixed artefact, that oversight line is often the difference between a result you can defend and one you cannot. A ladder with no senior QA time looks leaner on paper and is riskier in practice.
Look also at what the ladder does not contain. The two infrastructure and software-development roles that appear elsewhere in the workbook, the MLOps infrastructure developer and the software developer, stand at zero person-days in this particular sheet. That is not an oversight. It is the labour-side expression of the carve-out we will meet again below: the infrastructure work is scoped separately, as a package, rather than as days on the labour ladder. Seeing a role at zero on the labour sheet and a matching entry on the infrastructure line is exactly the kind of internal consistency a buyer wants to find, because it shows the two lines were reconciled rather than double-counted.
Every one of those person-day counts is a sourced figure from the phase effort-calculation workbook, not an estimate we are reconstructing after the fact. What set the shape of the ladder, though, was a decision taken earlier still: who would deliver the work.
The lever that shapes the ladder: who delivers the work
The phase was delivered through a partnership between two parties: the client-facing AI firm and a university. That is the single largest structural lever on the shape of the phase, and it is worth being precise about why it exists. It is not an accident of who happened to be available and it is not a by-product of the phase's size. It is a feature of who delivers the work, decided at the shape of the engagement, long before the commitment is signed.
A university partnership changes what the ladder can look like. It gives the phase access to doctoral and early-career researchers who can carry sustained experimental work, dataset preparation and iteration, which is exactly where a build phase spends most of its effort. It also changes what the senior rungs are for: when the volume of research labour sits lower on the ladder, the principal's and the senior researchers' time can be spent on direction and review rather than on production. The result is a ladder in which 140 of the 250 person-days, more than half, sit with the two early-career roles, and the senior rungs are deliberately thin.
The ladder above reads the same roles by seniority, one bar per rung, with the infrastructure and software roles drawn at zero because their work sits on its own line. The reading lever switches between person-days and each rung's share of the phase, and the orange bracket marks the two early-career rungs together: 140 person-days, 56 percent of the phase. A buyer who sees only the total sees a number; a buyer who sees the ladder understands that its shape is engineered into the delivery structure, and can ask the right question, which is whether that structure will hold for the life of the work.
Capability building sits directly on top of this. A university partnership is not only a staffing mechanism; it is a training mechanism. Across the wider programme, young professionals were trained through the partner institution, which means the partnership is producing two things at once: research effort for the phase, and a pipeline of applied-AI talent that a purely commercial delivery would not produce. For a client building its own capability, that second product is not a side effect. It is often a reason the engagement is structured with a university in the first place.
From role stack to fixed scope: the waterfall
With the ladder set at 250 person-days, the commitment is assembled in a few deliberate moves, and the order of those moves is the argument of this whole paper.
First, the senior quality-assurance allocation is fixed as its own line, so that review cannot be quietly traded away when the work gets busy. Second, the research and delivery roles are stacked on top of it, rung by rung, to the 250-person-day total. Third, and separately, the infrastructure for the build is set beside the labour ladder as its own package: zero person-days on the labour sheet, its own scope on its own line. Finally, a contingency reserve is named and sized against the scope actually contracted, not against a notional larger one.
The waterfall above is the spine of the argument, so read it deliberately. Each step adds one role's person-days to the running total. The orange step is the principal's quality-assurance allocation, the one line that is easiest to cut and most damaging to lose, and it is orange because it is the element that most changes whether the buyer can defend the result. The larger teal steps that follow are the research and delivery roles. The flat step labelled as the infrastructure package is the carve-out: it adds nothing to the labour ladder because the compute and tooling are scoped on their own line. The final bar is the phase total. Toggle the grouping lever and the same ladder regroups by tier, senior oversight, senior research and delivery, early-career research, and the thesis becomes visible in one gesture: senior time is a thin, deliberate layer on top of a broad base of research effort.
Two design choices in that build deserve to be named, because they are the difference between a disciplined commitment and a padded one.
The first is that infrastructure is carved out. The compute and tooling for the build are not folded into the labour ladder and are not hidden inside anyone's days. They sit on their own line, scoped as a package. This matters for honesty and for control. Compute is a real, separable item with its own drivers, and a buyer who can see it on its own line can question it on its own terms, can ask what hardware, what duration, what utilisation. Blend it into the labour and it becomes invisible and unaccountable. The run-side logic points the same way: serving a trained model is a property of the infrastructure, not of the research team, which is exactly why compute belongs on its own line rather than smeared across labour.
The second is that the contingency is named and held against the scope that will actually be delivered. A reserve folded silently into every role is a reserve nobody can see, question or release. A reserve sized against a notional scope larger than the one being contracted is padding. Holding it as a single named line, sized on the contracted ladder, is the sequence that keeps each component scoped on its own logic.
A sceptical buyer might ask whether the QA line is real or decorative, a senior name added to the cover sheet. The ladder answers that question. The principal's allocation has its own person-days, fixed before the rest of the ladder, and it sits above a research base large enough to need checking. A vendor adding decorative oversight would have done the opposite: blend a senior title into the team, with no days attached and nothing a buyer could ask to see. Here the review line is visible, sized and separate, which is the signature of an allocation that will be spent rather than a staffing optic.
Why the sequence is the substance
It is tempting to treat the review line, the contingency and the carve-out as administrative details that wash out in the total. They do not. Change the order and you change both the commitment and what it means.
Consider the contingency. A reserve is meant to cover the risk of the work that will be done, at the scope that will be contracted, so it belongs on that scope. Sized against an inflated scope, or spread invisibly across every role, it stops being a reserve and becomes a margin. The workbook holds it as a named line against the contracted ladder. That is a small decision with a clear rationale, and it is the kind of decision that distinguishes a scoped commitment from a padded one.
Or consider what the carve-out protects. If infrastructure were inside the labour ladder, any pressure to trim days would quietly erode the compute the work actually needs, and the first casualty of an under-provisioned compute line is the science. By scoping infrastructure separately and in full, the structure guarantees that a change to the labour ladder is a change to labour, and not a stealth cut to the hardware the models have to run on. The buyer keeps control of the scope; the delivery keeps its compute.
The expression above states the labour side of the build in one line. The ladder total is the sum over the five roles of person-days, and each role's share of the phase is its person-days divided by that total: 8 percent for the principal, 24 for the senior researcher, 12 for the project manager and senior researcher, and 28 each for the PhD and junior researchers. Infrastructure is not a term in the sum, because it is scoped as its own package with its own drivers, and the contingency is not a term either, because it is held as a named reserve rather than spread across the roles. There is nothing hidden in the total; every term is nameable and every term traces to the workbook.
Behind the fixed scope: a two-party delivery structure
The 250-person-day phase is what the buyer contracts. Sitting behind it is a delivery structure that describes how the two parties, the client-facing AI firm and the university partner, share the wider programme across its stages: a runway of two early phases, then the build phase this paper anatomises.
Read by party, the firm is the client's counterpart throughout, while the university brings research capacity. Read by stage, the weight of the work shifts: the university's work concentrates in the runway, consistent with its role in dataset development and training, while the firm's concentrates in the build, consistent with the engineering-heavy, infrastructure-heavy character of the phase.
The bridge above shows why this matters for a buyer even though the buyer contracts only the phase. The delivery split tells you where the work, and the risk, actually sits. In the runway, the university's dataset development and training is the main stage of its contribution; in the build, the orange card, the firm's engineering and infrastructure work is the main stage of its own. A two-party structure is not an administrative curiosity; it is the reason the ladder has the shape it has, because one of the two parties is the academic institution that opens the door to its broad base of research effort. The delivery split and the person-day ladder are the same fact seen from two sides.
There is a governance point folded into this too. When two parties deliver a programme jointly, the intellectual property, the publication rights and the effort-versus-result obligations have to be agreed at the structural level, not the line-item level. We settle those terms in the founding agreements, and we frame research services as an effort obligation rather than a guarantee of a specific result, which is the correct framing for a research phase and one a buyer should expect to see. The delivery structure and the legal structure are built together, and the fixed scope is the visible tip of both.
What the fixed scope does and does not commit
A fixed scope on a research phase is a specific kind of promise, and it is worth being exact about which risks it transfers and which it leaves with the buyer. This is where the effort-versus-result framing does real work rather than sitting in the contract as boilerplate.
The person-day ladder fixes effort. The vendor is committing to deliver 250 person-days of the specified seniority mix, and the fixed scope converts that commitment into a single line the buyer can plan against. What the fixed scope does not do, and honestly cannot do, is guarantee a particular research outcome. The nature of a research phase is that the result is discovered, not manufactured to specification, and a vendor who fixed the scope and guaranteed a specific accuracy figure would be either misunderstanding the work or misjudging the risk. Framing the services as an effort obligation is the correct and honest position, and it is why the person-day ladder, not a deliverable specification, is the object being contracted.
This has a direct consequence for how contingency should be read. The reserve is not a hedge against the research failing to reach a target, because no such target is being underwritten. It is a hedge against the ordinary friction of the work: an experiment that has to be rerun, a data intake that arrives dirty and needs a cleaning pass that was not scoped, a week where the compute misbehaves. Held as a modest, named line against the contracted ladder, it is a proportionate reserve for those frictions, not a margin buffer dressed up as risk cover. A reserve that is large relative to the phase, or that cannot be seen at all, is a signal that the vendor is either uncertain about the scope or padding the commitment.
The multi-phase structure changes the risk calculus again. Because the programme is staged, the buyer is not committing to the entire body of work at the point of signing one phase. Each phase ends at a decision point, and the delivery split we just traced, with its runway distinct from the build, is the structural expression of that staging. The infrastructure carve-out fits the same logic: a compute package sized for the build phase is committed when the build phase is, not spread blindly across the whole programme. Staging is a risk-management feature for the buyer as much as a delivery convenience for the vendor, and a fixed-scope phase inside a staged programme is a much safer object to sign than a fixed commitment for an entire multi-year research effort agreed up front.
How a buyer should read a fixed-scope phase
The reason to reconstruct one phase in this much detail is that the method generalises. A buyer handed a fixed-scope research phase cannot see the workbook, but can ask the questions that would expose its shape, and the answers separate a disciplined vendor from a hopeful one.
- Is the scope built from a person-day ladder, or is it a total in search of a justification? A total that is the exact sum of per-role allocations, like the 250 person-days here, with effort assigned role by role first, is evidence the work was scoped before it was committed. A total that arrives first, with no ladder behind it, is evidence of the reverse.
- Who delivers each rung, and why? If the ladder leans on early-career researchers, the buyer should understand the mechanism, a university partnership, a capability programme, an offshore team, rather than assuming the mix will hold. A staffing shape with a structural reason is durable; one with no stated reason is a future change order.
- Is senior review on its own line? A quality-assurance allocation with its own person-days can be asked for and checked. Oversight blended into a team is oversight nobody can see.
- Is infrastructure on its own line? Compute is separable, has its own drivers, and should be questionable on its own terms. If it is invisible inside the labour, the buyer has lost the ability to control the single most volatile item in an AI programme.
- Is contingency named, and sized against the scope actually contracted? A reserve that is hidden, or sized against a larger notional scope, is padding.
- Is the phase gated? A phase with a defined decision point before the next one lets the buyer stop, adjust or continue on evidence rather than momentum.
None of these questions require access to the vendor's spreadsheet. They require only the assumption that a fixed scope is an assembled artefact, and the willingness to ask how it was assembled. The 250-person-day phase in this paper answers all six, which is the whole reason it reads as a disciplined commitment rather than a hopeful one.
What the anatomy is for
We opened with the observation that a fixed-scope line item compresses its own machinery out of view. The purpose of this paper has been to decompress one such commitment completely: five roles on a 20, 60, 30, 70 and 70 ladder summing to 250 person-days, senior quality assurance on its own line, the infrastructure and software roles at zero on the labour sheet with compute scoped as its own package, and a named contingency against the contracted scope, all sitting inside a staged two-party programme with a runway ahead of the build and a gate between them.
The ladder is specific to this engagement. The method is not. Every applied-AI research phase can be read the same way, as an effort ladder by seniority, a review line, a compute package that stands on its own, a named reserve and a gate before the next phase. A vendor who can show that structure is telling you the scope was engineered. A vendor who cannot is asking you to trust a total. The thesis holds: a fixed-scope phase is built, not guessed, and the buyer who knows how it was built is the buyer who can tell the difference.
Limitations
This paper reconstructs the shape of a single research phase for a single client, and its figures should be read with that scope in mind. The person-day ladder and the zero entries for the infrastructure and software roles are the values recorded in the phase effort-calculation workbook for this engagement, and the separate infrastructure package is recorded in the infrastructure-package memo; they are not a market benchmark, and other engagements will differ in every parameter. Commercial terms are outside the scope of this paper. The seniority mix reflects delivery with a particular university partner and the needs of this phase, and it should not be assumed to generalise to other institutions, regions or role ladders. The two-party delivery split is described qualitatively, by stage and contribution, not as a line-item mapping to the phase. The instruments in this paper plot only sourced effort figures and carry no illustrative inputs, but they are a reading aid, not an audited record. Finally, this is an analysis of how the phase was scoped, not a claim about whether the scope was right; judging that would require the delivered outcomes, which this paper does not assess.
References
Effort-calculation workbook (engagement proposal artefact), Phase 3. The primary source for the person-day ladder (20/60/30/70/70), the zero entries for the MLOps infrastructure and software-development roles, the contingency line and the staged two-party delivery structure. Internal, held by the delivery team.
Infrastructure-package memo (engagement proposal artefact), Phase 3. The source for the Phase-3 infrastructure package scoped separately from the labour ladder. Internal, held by the delivery team.
Founding service and collaboration agreements (engagement governance artefacts). The source for the two-party delivery structure and the stage-by-stage contributions described in the delivery-structure discussion. Internal, held by the delivery team.




