Docs

Sigbound AI

The platform's reader. It reads every head against the goal, cites every claim, and never approves.

What it does

On every push of an attempt, Sigbound AI reads the exact head: the diff, the files the goal names, and whatever else it needs from the tree. It writes three things into the evidence:

  • What it does. One paragraph, every claim linked to a line of code.
  • Against the goal, line by line. Each done-means line scored pass, fail or cannot verify, with the line of code that decides it.
  • A verdict. Pass or fail on the whole, or needs a human when it could not verify.
Sigbound AI's read of attempt 6.2: three done-means lines, each cited
attempt 6.2 · three lines, three citations

What it never does

It never approves, lands, or edits. It hands you the decision. A line it cannot verify says so instead of guessing, and where branch protection requires its verdict, an unread head blocks rather than passes.

Cost

Reads run on the platform's key. Each receipt records what its own read cost. The space report shows agreement with human decisions and total cost over a window:

curl -H 'Sigbound-Version: 2026-08-14' -H 'Authorization: Bearer $SIGBOUND_TOKEN' \
  'https://api.sigbound.com/orgs/<space>/sigbound-ai?days=30&repo=<repo>'

From the terminal

sigbound do get_goal_attempt org=<space> repo=<repo> number=7 attempt=1   # gives the attempt id
sigbound do get_verdict org=<space> repo=<repo> attempt=<id>

Not yet

Bringing your own model key is on the roadmap; today every read runs on Sigbound's.