The reader that never approves.

On every push, Sigbound AI reads the exact head against what done means and cites every claim to a line of code. 12¢ a review, written on the run.

Request an invite
Evidence for attempt 7.1

Every claim links its code.

What it does, in one sentence with its citation. Then each done-means line, scored, with the line that proves it. Then the checks.

Evidence for attempt 7.1: Sigbound AI's cited read, each done-means line scored, then the checks
The Evidence tab of goal 7: pass, one done-means line proven at README.md line 30, three checks green.

What it does

Four things, on every push to an attempt.

Reads the exact head

Not the diff: the tree at the head that was pushed, against each done-means line of the goal.

Cites every claim

Each sentence it writes links to a file and a line. A claim it cannot cite, it does not make.

Answers in three words

Pass, fail, or needs a person. When a line cannot be verified from the code, it says so instead of guessing.

Never approves

The verdict is evidence for a person. The decision, and the merge, stay with the gate and with you.

Sigbound AI · attempt 7.1

What it costs is on the run.

12¢ a review, per review and not per seat. Every read is a run of its own, with its verdict, its duration and its cost in dollars on the page. The read of attempt 7.1 cost two cents.

The Sigbound AI run for attempt 7.1: verdict pass, 54 seconds, cost $0.021, one read and the model usage step
The run page of one read: verdict pass, 54 s, $0.021, on the platform's key.

When it runs

On every push to an attempt, and only when there is something new to read.

Every push

A new head is a new read. The previous verdict stays on the record; the goal page scores the latest one.

Outside the fence, nothing to read

A goal can say where its work lives. A push that changes nothing inside that fence carries the previous read forward to the new head, every citation still resolving, at no cost. A push that touches nothing in the fence at all gets a needs-a-person verdict that says so.

Insights · 30 days

Every read, counted.

How many heads it read, how often a person agreed with it, how often a person overruled it, and what the month cost. On this repository: 14 reads, 6 decided by a person, all 6 agreed.

Repository Insights over 30 days: the pulse of attempts and goals, and Sigbound AI's 14 reads with 6 of 6 agreed by the reviewer
Insights for snake-sigbound-2 over 30 days: 14 reads, 6 decided by a person, 6 of 6 agreed.

What it is not

The lines it does not cross, and the product is built so it cannot.

It cannot approve.

No verdict moves a branch. A pass is a reason to look; a fail is a reason to send it back.

A person decides.

Whoever made the attempt cannot approve it, and Sigbound AI is never that person either.

The gate re-verifies.

What lands is the bytes the gate verified again at landing, not what anyone said about them.

It does not name its model.

It runs on the platform's key. What you see is the read, the citations, and the cost.

Read the rest

Sigbound AI in the docs

The read, the citations, the verdicts, the cost, and the fence. Sigbound AI

Where the verdict ends up

Every read is quoted in the signed merge record of the attempt it read. Trust