Verification

Deployments & AI Verification

Deployed isn't done. DevSpec turns each item into acceptance criteria written from your own words — then the agent that did the work proves the ones it can actually observe, cites the evidence, and names what still needs your eyes.

Acceptance criteria · 3 of 4 settledstaging

Exponential backoff for Stripe webhook retries

Checked by the agent that did the work. Evidence cited per line.

  • Retries use exponential backoff with jitterSeen in the logs. 6 retries, 2s to 64s.
  • Each event is processed exactly onceSeen in the data. No duplicate rows after a replay.
  • The new column is live on stagingSeen on staging. Schema check passed.
  • The retry banner reads clearly on mobileLeft for you. A judgement call, not an observation.
1 line left for you3 proved with evidence

— criteria settled one line at a time, each with the evidence attached.

The Problem

'Deployed' is where most tracking stops

The merge went green, the deploy went out — but did the feature work? On most teams someone checks manually, or nobody does. So the queue fills with work that is finished but unconfirmed, and 'done' quietly comes to mean 'probably fine'.

Four acceptance criteria. Three are things a machine can genuinely observe — the retry backoff in the logs, no duplicate rows after a replay, the column live on staging. One is whether a banner reads well on a phone. Today all four wait on a person. Tomorrow you only look at the fourth.

The Mechanics

How it works

Verification is evidence, not opinion. The rule is one line: was the behaviour observed, or merely reasoned about? Reading the code is never enough — running it, querying it, or watching it is.

Criteria come from what you said

When work is captured in conversation, the acceptance criteria are written from your own words — not invented afterwards from a diff. Each one is a separate line that can be settled on its own, so a single item is never all-or-nothing.

Fig. i — Acceptance criteriaapp.devspec.ai/action-items

Your intent, in your words, on the record.

Deployments link themselves

A webhook fires when your pipeline deploys. Because Autopilot and the plugin tag commits with [devspec:id], DevSpec connects the deployment to the exact action items it shipped — and marks them with a deployed badge.

Fig. ii — Linked deploysapp.devspec.ai/changelog

The deploy knows what it shipped.

The agent proves what it can observe

Your coding agent — Claude Code, Cursor, or any MCP client on your machine — already has the tools to check its own work: query the logs, confirm the migration applied, run the thing and watch it. It records a verdict per criterion with the evidence it actually saw. Dev, the assistant inside DevSpec, can do the same on request.

Fig. iii — Recorded evidenceapp.devspec.ai/action-items

Observed, and cited — or it does not count.

What's left is named, not hidden

Criteria that need human judgement stay open and say so. When every line is settled the item is verified without ceremony; when some are not, you see exactly which ones and why — instead of a single unexplained "unverified" and a queue nobody wants to touch.

Fig. iv — Honest stateapp.devspec.ai/testing

One line left for you beats four.

An Index of Capabilities

Evidence, not optimism

01

Verified line by line

An item is a set of criteria, each settled on its own evidence. Three proved and one outstanding is a real state — not a rounding error.

02

Observation beats inference

Reading the code and concluding it should work is advisory. Querying the logs, checking the schema, or running the thing is verification. The distinction is enforced, not suggested.

03

Your agents, your machine

The agent that implemented the work is usually best placed to prove it — it has the context and the tools. Reviews can also be handed to another of your connected agents, or to Dev.

04

Intent outranks the letter

When implementation diverges from the original wording for a good reason, the agent has to show the outcome still delivers what you asked for — rather than quietly rewriting the criteria to match what it built.

05

Full traceability

Conversation → action item → commit → deployment → verification. Every link in the chain is recorded and clickable.

06

Human judgement, where it belongs

Taste, wording, and whether it feels right stay yours. DevSpec's job is to make sure that's the only thing waiting on you.

DevSpec

Convinced?

Set up your first project in under five minutes. Connect your repo, start a session, and see what your duck can do.

Start your project

Start with $15of usage, on us · nothing charged until you buy