Redback, the agent that researches the catalogue
Redback is the agent that works on the catalogue's gaps and conflicts. The quality checks find a missing figure or a disagreement; Redback researches it on the open web, reads the page that states the answer, and files a proposal with the source and the sentence it came from. A person decides what reaches the catalogue.
It is in active development. What runs today, what is being built and what is only planned are listed at the foot of this page, and this page changes when they do.
What it is
- One agent with one job: the remediation queue. When a quality check fails on a record, the failure becomes a task, and Redback picks tasks up.
- It runs on a schedule and when a person starts it, a small batch at a time.
- It proposes and never edits. A task it closes means a proposal is waiting for review, not that the record changed.
- Every task keeps its transcript: the searches, the pages read, what the model was asked and what it answered. The reviewer reads that, not a summary of it.
Each run carries an identity of its own, the agent's name and six hex characters, such as redback-3f9a2c. That is the name a claimed task shows.
The tasks it works
| Task | What Redback does |
|---|---|
gap | A record is missing a figure. Redback researches it and proposes the value with the page and the quote that state it. |
contradiction | Two facts disagree. Redback decides only whether there is a fault: the record is wrong, the check is wrong, or it cannot tell. It never proposes the corrected value. |
vocabulary | A value is outside the accepted list. Redback says whether it is a second spelling of a listed value, a new value, or the wrong question, and proposes no edit. |
asset | A record has no picture. Not Redback's work: finding and clearing an image is a person's. |
How one gap is researched
A gap is the one task researched on the open web, so it is where the checks are strictest. The model's part is small on purpose: it reads, and code checks what it read.
It claims a task
A run takes the next open task its policy allows and holds a claim on it, so no second run takes the same one.
Code searches and reads
Code writes the searches from the record itself and reads the record's own citations and the results. Each page is cut to the passages that name the field.
The model reads
The model gets those passages and one question: which passage states the value, in what words. It is never shown the value we hold.
Code checks the answer
The quote must be on the page as it was fetched, the figure must be in the quote, and the unit must convert. A length with no unit is refused.
A second source, a second model
It looks for the same figure on another site, and a model from a different provider answers yes, no or unsure. A no, or low confidence, goes to a stronger model.
Tools, when reading found nothing
Only when the research found nothing to read does a model get the tools below and a few turns to look for itself. Its answer faces the same checks.
It files, or says why it cannot
The result is a proposal with its source, its quote, the kind of site it came from and what each check found, or a finding that no source states the figure, which rests the task for a while before it is tried again.
The tools it can use
A tool is how a model asks for something it cannot know from the task. The model never touches the network: it names a tool, our Worker decides whether to run it and runs it, and the model gets back the result. There are twelve tools in five groups.
| Group | What it does | The rule it keeps |
|---|---|---|
| Reading pages | Fetches a web page or a PDF, such as a manufacturer's spec sheet or manual, and returns its text. | Fetched by our Worker under our own user agent. robots.txt is honoured and a refusal is said, not hidden. Private addresses and anything but http and https are refused. |
| Searching the web | Searches the web through several search providers, one taking over when another cannot answer, and can search a single site, such as the maker's own. | A search result is where to look, not an answer. Only a page actually read can be quoted. |
| Reference works | Looks up Wikipedia articles, their infobox fields, and Wikidata entries, to learn which figure to expect and which maker to ask. | A lead, never a citation. Nothing in them came through a fetched page, so there is no quote to check. |
| Our catalogue | Reads a record, searches the catalogue, and reads a record's relations, such as the manufacturer of a firearm. | Read only. It is how a contradiction is judged, since two records that disagree need both read. |
| Helpers | Converts units, and keeps short notes across the turns of one task. | Arithmetic is done in code, never by the model. A note is a reminder, not evidence. |
The sources it weighs
Every page Redback reads is sorted by what kind of site published it, and the kind is recorded on the proposal for the reviewer. The order below, highest first, is the one the catalogue uses throughout: a figure is weighed by the best kind of source that states it, never by how many sites repeat it. A site that fits none of these is recorded as other.
manufacturer: The maker's own product page, manual, datasheet or catalogue. The authority on anything the maker publishes.standards_body: SAAMI, C.I.P. or a NATO standard. The authority on cartridge dimensions and pressures.government: A procurement document, a technical manual, a service's own site, a national museum. Any <code>.mil</code> or <code>.gov</code> host is classed here by rule.reference: An independent reference that names where its figures came from: an encyclopaedia, Modern Firearms, Forgotten Weapons, a museum collection, Jane's.aggregator: A site that compiles figures from other sources without saying which. Useful for coverage, weak as evidence.press: A magazine, review site or blog: a journalist's own measurement, or a maker's figure repeated.retailer: A listing written to sell the item. Evidence the item exists and what the market calls it; its figures are copied from elsewhere.community: A forum, wiki or fan site. Evidence of existence and naming only, never the sole source for a figure.
Working across AI providers
Redback is provider independent. Every model call on the platform goes through one internal layer that sits in front of several AI providers, so a model is chosen for each piece of work on its merits and can be replaced without touching the agent.
- Models are matched to the work. Each kind of task has a model suited to it, reviewed by an operator and changed without a release.
- No single provider is a dependency. Every task has a fallback path through models from other providers, so an outage at one does not stop the queue.
- Verification is independent. The second opinion on a figure comes from a different provider than the one that read the page, so the two do not share a blind spot.
- Usage is governed. Budgets are set in advance and enforced before a call is made, never reconciled after it.
- Failure is contained. A model call never fails a task: when a provider is unavailable the task is held and resumes when service returns.
- Every call is audited. Which model answered, and what it cost, is recorded with the run.
What it never does
- 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.
- It is never shown our value while it reads a page.
- It never lets a model's confidence decide on its own. The source kind is recorded for the reviewer, not scored.
- It never guesses a vocabulary value or a corrected figure. Without a list or a source, the task is skipped.
In active development
Redback is being built in the open, one task kind at a time. This is where each part stands.
| Work | Where it stands |
|---|---|
| The task queue fed by the quality checks | Running |
| Researching gaps on the open web | Running |
| Deciding contradictions and vocabulary values | Running |
| Tools, with the tool-using fallback | Running, with more tools added as tasks need them |
| Applying approved proposals in reviewed batches, and measuring what they change | In development |
| Enrichment from Wikimedia: other names, related records, new contradictions to check, missing firearms | Planned |
| Picture tasks | Planned, and a person's work rather than Redback's |