System architecture and requirements review
The cheapest review to buy and the one that saves the most money, because architecture mistakes surface late and cost a great deal more by then.
The decisions that get expensive later
Run this before schematic capture, when you have a spec, a block diagram, interface definitions and a parts shortlist. Everything is still cheap to change at this point, and it is the last moment that will be true.
What goes wrong here is rarely dramatic. A requirement that can’t be verified and nobody noticed. A power budget with no margin allocated, or margin allocated twice. An interface where both sides assumed the other one terminated it. A part chosen because it was in the drawer, now load-bearing. A whole class of requirement missing, usually environmental, EMC or maintainability, because nobody owns it.
The other thing I look for is what got decided by nobody. Inertia is the most common architect in engineering and it never signs its work, so a large share of load-bearing choices in any programme arrived by copying the last one rather than by anybody choosing.
This one is interview-heavy
More of this review comes from talking to your people than from reading your documents, which is deliberate. Expect a longer scoping call and a proper question list. The answers are the findings.
Best value in the catalogue
Changing a part choice costs nothing at spec, hours at schematic, days once layout is underway, and a respin after fab release. The gradient is roughly an order of magnitude per stage, and this review sits at the cheap end of it.
What gets worked through
Requirement quality
Each requirement singular, verifiable, unambiguous, free of implementation, with a stated tolerance and a verification method. The ones that fail this test are the ones that get argued about at acceptance.
The requirement classes that go missing
Environmental, EMC, power quality, thermal, mechanical, safety, security, maintainability, transport and storage, end of life. Whole categories get forgotten because no individual owns them.
Traceability and verification
Whether every requirement has a route to being proven, what test equipment and fixtures that implies, and what needs a dedicated test board that nobody has budgeted for yet.
Interface definitions
Electrical, mechanical, protocol, timing and error behaviour, with a named owner on each side. Interfaces where both parties documented their half and nobody checked the two halves agree.
Block diagram integrity
Every block with an owner, every interface with a defined signalling standard, no magic arrows. Blocks drawn as independent that share a supply, a clock, a ground, a supplier or a person.
Budgets that have to add up
Power with worst-case load and real margin, thermal against worst-case ambient and the airflow you can actually guarantee, size, weight and power with margin allocated somewhere specific, and the performance chain: latency, jitter, noise and error rate budgets that survive being added up.
Margin policy
How much, allocated where, and who is allowed to spend it. A programme where anyone can quietly consume margin has no margin, it just doesn’t know that yet.
Key part defensibility
Availability, lifecycle status, lead time against your schedule, second source, qualification status and export control. Whether each choice can be defended on its merits or arrived from a previous programme.
Risk register quality
Whether the top risks are the real ones, or whether the register lists what was easy to write down. Assumptions and open items registered rather than buried in a slide from four months ago.
Certification route
Which standards apply, what evidence they demand, and when that work has to start. Certification evidence gathered retrospectively costs several times what it costs collected as you go.
Make against buy
Technology choices and where a COTS module is buying you less than it appears to once integration, qualification and obsolescence are counted.
Design rationale
Whether the reasoning behind each architectural choice is written down anywhere. A choice nobody can defend is a choice nobody can safely revisit when the constraint that forced it changes.
The package
- Requirements document or specification, however rough.
- System block diagram.
- Interface control documents, all of them.
- Concept of operations, or just a description of how the thing gets used.
- Environmental and qualification requirements.
- Power and thermal budgets as they stand.
- Size, weight and power constraints, with current allocations.
- Key parts shortlist, with lifecycle and availability if you know it.
- Risk register.
- Verification and test strategy, however early.
- Certification route and applicable standards.
- Heritage: what this derives from, and what changed.
Early-stage material is expected to be incomplete. Send what exists and say what doesn’t.
Depth
| Tier | Depth | From |
|---|---|---|
| Scan 3 days | Architecture integrity, obvious requirement gaps, budget sanity, and a first read of key part defensibility. | £750 +VAT |
| Standard 7 days | Full checklist plus a structured interview. Requirement quality assessed line by line, interfaces checked for two-sided agreement, budgets checked for margin and for double counting, certification route assessed. | £2,750 +VAT |
| Deep 15 days | Adds rebuilding the power and thermal budgets independently, worst-case analysis on a nominated critical chain, and a re-review after the architecture is revised. | £6,500 +VAT |
Starting prices. Where a job lands within a tier depends on size and complexity, and the proposal carries one firm number before anything begins.
Full pricing and what changes it →
Pairs well with a schematic review
Architecture review now, schematic review once capture is complete. Two review types in one engagement takes 10% off the second, because there is real efficiency in already knowing the design.
Reviewing something else?
Four review types, run to the same standard. Two in one engagement takes 10% off the second.
Still at spec and block diagram?
Good. This is the cheapest moment to find out something is wrong. Thirty minute call, no charge.