66% “almost right, but not quite” — the #1 developer frustration of 2025 (SO)
Your AI writes the code. You still own the one thing it can’t: saying what correct means.
plumb shows your code does what you said — right where “said” can be made precise. An independent oracle derives intent from your signature and docstring, asks once, then freezes the verdict as reproducible evidence.
Under a reversed burden of proof, “unknown” is your weakest position. This is the same evidence the CLI above produces — distilled for the people who answer for the release.
The oracle reads your intent, never your code. It asks once.
In 2026, the first reader of your code is an agent — and the one thing it can’t do is decide what correct means. That line stays human. plumb is where it’s drawn.
From the signature + docstring — the oracle never sees the body. Your code stays local.
No silent default. Where your intent leaves room, it surfaces the assumption — agreement among models isn’t correctness — and you ratify it once. The one call it can’t make for you.
After ratify, the verdict is deterministic, re-runnable — no model in the loop.
A diligence record, designed to be signed two ways — native on every export.
It lands in .plumb/, next to your code — only a digest ever crosses the wire. Diligence you can show later, not just claim.
“mg dose = weight_kg × mg_per_kg, clamped to [0, max_single_dose]”
The standard was derived without sight of the code — 3 independent models, frozen.
We checked code 9f2c… against this standard before the v2.4 release.
Where the claim begins and ends.
- is
- an independent verification layer — shows code does what you said, on pure logic
- is
- a diligence record — contemporaneous, attributable, reproducible
- is not
- another LLM opinion — “AI checking AI”
- is not
- a bug-finder, or a promise of correctness
- is not
- a judge of whether what you said is right — that stays yours
Join the waitlist
We’ll tell you when the plumb-bob hangs true. Email only.
I have a case
Role · sector · codebase language · regulated · urgency. No free text.
From 9 December 2026 the EU Product Liability Directive treats software as a product — on complex systems the burden of proof can shift to the maker. “We were careful” will need to be shown, not said.