The pitch, on one card
goldspec turns the requirements you already have into a form a machine can check — so overlapping work, decisions with no outcome, and steps missing their inputs surface before anyone builds.
plumb then measures the built system against each requirement one at a time, and names the ones nothing is holding.
Together they make done a measurement instead of a judgement, so work produced through a keyhole is tested in the context of the whole project. That is what lets AI scale.
Apache-2.0 · runs on a laptop, in CI, or on-premise · a person answers every case · our judgment, conversion and the deep review are $120 a seat a week
Specification and proof, for work a person cannot review line by line
AI stops scaling where proof runs out.
A model sees the task you hand it and almost nothing else. Every piece it writes is correct through that keyhole and unchecked against the system it lands in — which is why adding more pieces stops helping, and why the ceiling was never how fast it can type. Make the whole project's requirements the thing each piece is tested against, and the keyhole stops being the limit.
Measured, not modelled
What it has found, on real systems
Every figure here comes from a run on a working codebase, and every one names where it came from. Nothing on this site is extrapolated from a pilot.
The trading-platform figures are a trusted tester's six-day run against a live Java system of some 374 production classes and nineteen hundred tests, against a written specification of forty-nine requirements that predated us. The full write-up publishes with the repositories; the operator who ran it is available as a reference via contact@plumbspec.org.
The tools are measured the same way, on every release. Each one publishes its specification with what its own suite proved — plumb, goldspec — unproven stories included, and the history of those runs is readable live, on the same page a paying customer reads.
Thirty seconds, in your browser
Try it on requirements you already have
The first thing the method checks is whether a requirement is something a person performs. On a specification written before anybody asked that question, most are not — and a requirement nobody performs cannot be ordered, scheduled, or tested.
Paste sample requirements
Anything will do: a few lines from a document you already have, or the sample below. This names the ones no test could ever prove.
Nothing leaves your browser. These checks are structural — the page works the same with the network disconnected.
waiting for input
- Paste a few lines, or load the sample, and this will fill in.
Support
A method this different needs somebody to ask
Most of what we have built is behind the tools rather than in them: a service that answers questions, takes cases from inside the tool, holds your team's run history and reaches a person when it cannot help. This is what it does, and what each part costs.
Free
Ask, and get the written answer Questions are answered from our manuals and this site. One we cannot answer goes to a person, and the answer is written down: we email it to you if you leave an address, and the next person who asks gets it straight away. Ask a question →Free
One account for the team Sign in with a work address and a mailed link, with no password. Members, roles and seats are managed in the browser, so whoever handles billing never installs anything. Each machine gets its own credential, shown once and revocable on its own. Sign in or sign up →Paid
Your run history, hosted Each run's results are pushed from wherever plumb runs. The whole team then reads one history, and a requirement that was proven and stopped being proven is named, along with the run where it turned. How the history works →Free
Cases filed from the terminalplumb support files a case from inside the tool, along with the run it is about. You and support
take turns on that one case, and you can open, follow and answer cases in the browser too. A person
answers every case, free or paid. If it has waited too long, you are
told so plainly and given the best answer we hold. Nothing goes quiet.
Get the tool's credential →
Paid
Conversion and orchestration We convert the requirements you already have into goldspec's checkable form, and run guided capture so your people write the specification and it stays whole. Pricing →Open to anyone
Whether it is working, right now The support page checks each part of the service when you load it and says which parts are not working. It does not just show a green light. Service status →What you file is sealed under a key held for your account alone, and the key can be destroyed. Your address is stored as a keyed digest for lookup, so a copy of our database does not reveal who our customers are.
Where it applies
Three kinds of programme, three different arguments
The mechanism is the same in all three. What changes is who signs, what the acceptance question is, and what it is worth — so each has its own page rather than one page hedging between them.
Government & regulated
Replacing a mainframe Decades of mandated documentation, original engineers gone, and an acceptance question somebody will be audited on. Includes the contracting and threshold detail, and what to do when the documentation is stale or missing. The legacy modernisation case →Commercial
Rebuilding a system you own A platform replacement or package migration where behaviour has to be preserved but no auditor is watching. Same measurement, same seat rate, and a straight comparison against doing it the way you do now. The rebuild case →Greenfield
Building something new No legacy behaviour to preserve, and the specification is the only thing standing between a model and a system nobody can review. Includes what a formal decision process cost before this — 49 records, 70,000 words, and nothing settled. The greenfield case →