The cost isn't the bug you fear.
It's the ten features you never attempt

Two things you can buy today. Everything after that is priced before you commit to it.

You never have to work out which category you are in. That is the job of the free triage call, and the answer arrives in writing within 48 hours.

WHAT IT COSTS TO GET TO EACH DECISION · ◆ = YOU CAN STOP HERE

1. · €0 TRIAGE
2. ◆ €750 BLAST RADIUS
3. ◆ €9,500 DISCOVERY SPRINT
4. · €3,500+ SAFETY NET
5. ◆ €12,000+ BUILD MILESTONE
6. · €3,000+ / month RETAINER

IF YOU DECIDE TO BUILD — ALL OF THIS IS PRICED INSIDE YOUR SPRINT, NOT INVENTED AFTERWARDS

• Safety net €3,500 — €6,000 Once, before any build.
• Build milestones €12,000 — €25,000 each Fixed on signature. Underwritten to +20%.
• Retainer €3,000 — €5,500 / month Optional. Keeps the map true as the system moves.

FIXED ON SIGNATURE · 40% UP FRONT · NEVER DISCOUNTED — A SHORT BUDGET MEANS LESS SCOPE, NOT A LOWER RATE

We sign the build price. Overruns inside the map are ours, up to +20%.

Every milestone is fixed on signature and quoted from what the sprint actually measured. If we misread your system, absorbing that is our problem — which is the only reason two weeks of reading is worth paying for.

What it covers

Anything inside the scope the sprint mapped. A misread on our side is our cost, not a change request in your inbox.

What voids it

Verbal scope changes, blocked repository or staging access, and client-side dependency changes after signature. Printed on every quote.

Why we can offer it

Because we read the system for two weeks first, and because no single milestone is ever quoted above €25,000.

An agency quoting from a discovery call cannot do this, because it has not read the system. An agent cannot do it either, because when it is wrong there is nobody to bill.

You are not choosing between us and nothing. You are choosing between five options.

Four of them are cheaper than we are. Here is an honest account of what each one actually gives you, written so you can disqualify us quickly if one of them fits better.

◑ means partially — a capable team or agency will produce diagrams, and they will be roughly right. Any one of these five columns is cheap to claim on its own. The price only holds because all five are true at the same time.

Doing nothing

The most common choice, and never really a decision — a deferral. Not free either: the drag above is already on your payroll, it just has no line item.

Your own team with an AI agent

You should do this, and we use the same tools ourselves. It will draw a decent diagram in an afternoon. It cannot stand your system up and measure it, tell you not to do the thing you asked for, or sign a price it is liable for. And the people who built the system cannot see it — nor say what they see without political cost.

An agency that uses AI

The closest competitor, and the difference is not tooling. They bill for capacity, so their incentive is more code; we bill for a decision at a fixed price, so ours is less. They point models at the system before a spec or a test exists to catch them. We point them at narrow, test-gated tasks after both do. And they skip the safety net, because it looks unbillable.

A traditional software agency

Sells capacity, paid by project size, so the project grows. Most are clustered on greenfield MVP work, which is commoditising. Reading an existing system and changing it safely is a different job.

Every claim on this page is checkable. Here is how you check it.

Process claims are worth nothing on their own — every agency makes them and every buyer discounts them to zero. So we publish the rules, open the delivery log while the work is running, and print our own numbers every quarter, including the ones that make us look bad.

Before you buy

The rules, published in full: three attempts then a human takes over · secrets never enter an agent’s context · production deploys stay manual · nothing merges into auth, payments or data deletion without a human · no agent touches an existing system before the safety net exists.

While we work

Your task log is readable by you, and the weekly report carries the same numbers we use internally. Expect a high needs-human rate on an existing system — that is the designed distribution. Zero would mean we were lying.

After handover

Every gate leaves a receipt and you keep all of them: an ADR, a review record, a release note and a demo recording.

WHAT THE RECEIPTS LOOK LIKE · FORMAT IS REAL, CONTENT IS ILLUSTRATIVE

Context
Pricing rules live in two services and
nobody knew which one wins.

Decision
Keep both. Add a contract test at the
seam and make the winner explicit.

Consequence
No rewrite. Release lock removed.

reviewer   architect (human)
comments   3, one blocking

blocking   touches auth/session.ts —
           needs a second pair of eyes
           per SEC-4

resolved   14:22, merged after fix
deploy     manual, 15:05

date,proj,type,task,att,result
06-11,acme,T2,#218,2,green
06-11,acme,T2,#219,3,needs-human
06-12,acme,T2,#220,1,green
06-12,acme,T2,#221,3,needs-human
06-13,acme,T2,#222,1,green

first-pass green   3 / 5
needs-human        2 / 5

OUR OWN NUMBERS · PUBLISHED EVERY QUARTER · Q3 2026

WHAT WE MEASURE ABOUT OURSELVES THIS QUARTER

Discovery Sprints delivered

—

Sprints that ended in a written “do not proceed”

—

Build milestones closed at the signed price

—

Milestones where we absorbed the overrun, and how much

—

Milestones delivered late

—

Engagements where the safety net was skipped

—

No engagements yet, so the column is empty. It publishes from the first sprint onward and it is not curated — a milestone we ran late on appears here the same quarter it happens. Milestones where you signed a scope change are reported in a separate column, so “closed at the signed price” cannot be inflated by change requests.

What you see: your task log, the weekly report, every gate receipt, CI results. What you don’t: our internal retrospectives, other clients’ data, our margins, raw model transcripts. These figures are operational data, not a service-level guarantee — the only contractual promise on this page is the underwrite.

We publish the method on systems nobody paid us to look at.

One open-source system a month, torn down in public: the drawings, the ranked risks, and one number measured from the system actually running. Including the ones where the honest answer is that nothing needs rewriting.

⚠ PLACEHOLDER — REPLACE EACH CARD WITH A PUBLISHED TEARDOWN BEFORE LAUNCH

• TEARDOWN 01 00% Share of traffic reaching a service the maintainers believed was retired. VERDICT — REPLACE
• TEARDOWN 02 0.0× Gap between the documented request path and the real one. VERDICT — REPLACE
• TEARDOWN 03 0 Rewrites recommended. The architecture was fine; the release process was not. VERDICT — REPLACE

There are no client logos on this page. When there are three we will put them here, and not before. Until then the teardowns are the evidence, and you can check every one of them yourself against the repository it came from.

The five questions everyone asks.

Why not point Claude Code at our repo ourselves?

You should — it will draw a decent diagram in an afternoon, and we use the same tools inside our own process. What it cannot do is stand your system up and measure where the time actually goes, tell you not to do the thing you asked it for, or sign a price it is liable for. If a diagram is all you need, don’t hire us.

What access do you need?

Read access to a repository that stays in your organisation, least privilege. Secrets never enter an agent’s context. Production deploys stay manual, always.

Do we have to move to your stack?

No. On an existing system the stack is yours and we work inside it — that is the entire job.

What if the sprint says don’t proceed?

Then it says that, in writing, and you keep every artefact. Finding out in week two is considerably cheaper than finding out in month six.

Who actually writes the code?

One human architect owns the design and reads every diff before it merges. Models draft inside narrow, test-gated tasks. One accountable person, not a rotating bench.