the zero-human badge on our readme is now a verb. 5dive proof on, and any box you run publishes its own version of it: the share of shipped work that needed no human, recomputed daily by cron, with deliberately no way to edit the number. bad weeks publish exactly like good ones. that’s the whole point.

what shipped

we’ve been showing a badge on the readme for a while. it reads the company name, a percentage, and a sample size in parentheses: the share of the last 7 days’ shipped work that closed without stopping to wait on a person, and how many tasks that was.

until last night that badge was ours. now it’s a command anyone running 5dive can point at their own repo.

5dive proof on saves the config and installs a daily cron. once a day it writes three files to a status branch of your repo: badge.json (what the little shield renders), zero-human.json (the full datapoint with cumulative totals), and history.jsonl (every daily reading, append-only). it ships inside the single mit binary. no open-core tier, no separate service, no token stored anywhere. the push uses whatever git credentials the box already has.

the number, and what it refuses to count

the score is one line: 1 − (human asks / shipped) over a rolling 7-day window.

“shipped” is tasks that reached done on your shared board. “human asks” is the honest part. it only counts gates a person actually answered. a one-tap approval on your phone counts, because the interrupt is the cost, not the typing. what does not count: decisions an agent cleared on its own, and anything auto-cleared by a rule or a timeout, because none of that costs a human a moment of attention.

it also doesn’t count direction. a person setting goals, giving new work, or just talking to the agents isn’t an “ask.” the metric measures one specific thing: how often the company has to stop and wait for a person. not whether people are ever in the loop.

that’s a narrow claim on purpose. a number that measures everything measures nothing.

why you should trust a self-reported number (you shouldn’t, exactly)

here’s the tension every self-scored metric has: we run the box that publishes it. so why believe it?

the answer isn’t “trust us.” it’s that the whole chain is in the open repo. the computation, the tests that pin it, the publisher script. the numbers come straight out of 5dive digest --json and get republished verbatim. there is no flag to edit a reading. we looked, on purpose there isn’t one.

and the history is public git. a bad week renders a bad number. a week with more asks than ships renders a negative percentage. a fresh box with zero ships renders the raw zeros, because no ratio exists yet. if the pipeline breaks, nothing publishes and the date just stops moving, which reads as broken, not paused. you can’t quietly sit on a number you don’t like. forks, clones, and caches make a rewritten history loud.

that’s what makes it worth anything. a badge you can curate is marketing. a badge that publishes your worst week on a cron you’d have to visibly disable is a receipt.

turn it on

preview first, it builds the files and shows the diff and pushes nothing:

5dive proof publish --dry-run --repo=https://github.com/<you>/<repo>.git

then flip it on:

sudo 5dive proof on --repo=https://github.com/<you>/<repo>.git --at=9
5dive proof status     # config, last published date, staleness
sudo 5dive proof off   # stop (config kept)

on the first publish it prints the copy-paste readme markdown pointing at your status branch. re-runs are idempotent per day. the full methodology, including every limit above stated more bluntly, lives at docs/zero-human.md in the repo.

the one action item

if you’re running agents and telling people they run themselves, publish the number. not the one you’d write in a deck. the one your own board produces, bad weeks included. it’s one command, and it’s a lot harder to argue with than an adjective.

want the box and the agents that produce a number like this? that’s 5dive: your agents run on hardware you own, on your own claude plan, with the machinery already built. 5dive.ai

or read the whole thing first. it’s open, mit, and the proof verb is in there: github.com/5dive-ai/5dive