Redback
The agent that works the remediation queue. It researches what the quality checks find and files a proposal with its source. A person decides what reaches the catalogue.
The tools it uses, the AI providers behind it, and what is still being built
Who it is
When a check fails, the failure becomes a task, and Redback is what picks those tasks up. It runs on a schedule and on demand, a small batch at a time, and every run is recorded with what it read, what it asked and what it spent.
One agent, many runs. Each run carries an identity of its own, the agent's name and six hex characters, such as redback-12443d. That is the name a claimed task shows, and the one the spider on the trust page answers to when you point at it.
What it works
A task has a kind, and the kind decides the work: a missing figure, two statements that disagree and an unlisted spelling are different jobs.
| Kind | What was found | What Redback does |
|---|---|---|
gap | A record is missing a figure. | Researches it on the open web, reads the passages that name the field, and files the value with the sentence that states it and the page it came from. |
contradiction | Two facts in one record disagree. | Decides whether there is a fault at all. It never proposes the corrected value: a record that disagrees with itself cannot say which half is right. |
vocabulary | A value is outside the accepted list. | Compares the spelling with the list the task carries and answers in one of a few fixed words. One decision covers every record holding that spelling. |
asset | A record has no picture. | Not Redback's. Finding and clearing an image is a person's work. |
How it works a gap
A gap is the one kind researched on the open web, so it is where the checks are strictest. Each stage below is the pipeline as it runs; the figure beside it is the same spider doing it.
01It claims a task
A run takes the next open task its policy allows and holds a claim on it, so no second worker picks up the same one. A claim held too long is returned to the queue.
02It searches and reads
Code words two searches from the record itself, the full name with its maker and part number and then the part number alone, reads the pages, and cuts each to the passages that name the field. The model gets those passages and one question: which of them states it, and in what words.
03Code checks the answer
Three checks a model cannot talk its way past: the quote is on the page as fetched, the figure is in the quote, and the unit converts. A length stated without a unit is refused.
04A second source, a second model
It looks for the same figure on another site among what it read, and a model from a different provider answers yes, no or unsure. A no, an uncorroborated unsure or low confidence goes to a stronger model; a plain not found never does.
05It waits when a provider cannot serve
A busy host or a rate limit says nothing about the task, so the task is deferred until the provider's own retry time rather than failed, and the run stops, since every task behind it would meet the same wall.
06It files a proposal, and a person decides
The answer is filed with its source, its quote and what the checks found. A task Redback closes means a proposal is waiting for review, not that the record changed.
What it never does
The rules below are why an agent working the catalogue can be trusted with it.
- It never writes to the catalogue. Every change is a proposal a person reviews and makes.
- It never files a figure whose quote is not on the page it fetched. A fabrication becomes a dropped answer, not a citation.
- It is never shown our value while it reads a page. A model that has seen the answer is not a witness.
- It never lets a model's confidence decide anything. What kind of page it read is recorded for the reviewer, not scored.
- It never guesses a vocabulary or a corrected value. Without the list, or without a source, the task is skipped.
The companion on the site
On the trust page Redback walks the grid that explains this pipeline, stopping on each stage to do what that stage does. It is drawn in code, not played from a file.
- Each foot is pinned to a point on the page and steps only when it falls a stride behind, with the knees placed by inverse kinematics, so it can walk any curve at any speed without its feet sliding.
- Walking, working, reading, idling and hanging are one drawing, the logo's proportions measured and mirrored, so it never changes shape between them.
- Its path is measured off the page and runs along the rules between the cells, climbing each edge it crosses. It works on a perch, sometimes idles, and sometimes lowers itself on a line from the bottom corner of a card.
- It never moves because you scrolled. If you can see it but not where it is going, it waits; if you cannot see it, it carries on and walks into view.
- Now and then it casts a thread to a point nearby, and a line cast ahead stays out until it gets there.
Read the queue yourself
How many tasks are open, and in which state, is public, and so is the gap report the tasks come from. Both are in the data quality section of the API reference, and neither needs a key.