One Redback task, end to end

The case

The record is the Glock 21 SF, and the field is its empty weight. The run is taken from the blind measurement: the figure we hold was hidden, the page it came from withheld and the record's own citations left out, so Redback met it as an empty field with nothing to start from but the record's name and maker. Everything below is what that run did and produced.

The whole path

A task can end in three places: discarded with the rule that refused it, filed as a doubt for a person to settle, or filed as a proposal. Only a person turns a proposal into a change to the catalogue, and a rejection sends the task back with the reviewer's note.

flowchart TD
  finding["A check finds a fault"] --> task["A task on the queue"]
  task --> run["A run claims it with<br/>its record's other tasks"]
  run --> gather["Code searches, reads pages<br/>and cuts excerpts"]
  gather --> reader["A model reads the excerpts"]
  reader --> held["Code finds the quote on the page"]
  held --> checks{"Another site, a second model,<br/>the evidence rules"}
  checks -->|"refused"| discarded["Discarded, with its rule"]:::muted
  checks -->|"doubt"| stronger["A stronger model reads again"]:::warn
  stronger --> proposal["A proposal with its full trace"]
  checks -->|"passed"| proposal
  proposal --> person{"A person reviews"}
  person -->|"accepts"| catalogue["A pull request to the catalogue"]:::ok
  person -->|"rejects, with a note"| task

Every stage, and what it produces

The table follows a gap task. A verification skips the search and reads only the maker's page; a contradiction and a vocabulary spelling skip the reading and ask a model to judge what the record already holds. Every task ends in the same filing and the same review.

The run, stage by stage

The same stages, with what each one produced for the Glock 21 SF.

1. Search. One query for the model's name and the field, with the maker's site first. Among the results was Glock's own page for the model.
2. Read pages. Three pages were read. Glock's page is classed as the maker's own, the strongest kind of source, and its technical data table was read as stated facts rather than as prose.
3. Cut excerpts. Three excerpts were cut. Glock's technical data table lists three weights one after another (740 g without magazine, 825 g with an empty magazine, 1100 g with a loaded one), and the excerpt from its page was cut around them.
4. Read. The reader model named that excerpt and read the figure as 740 g, from the line labelled without magazine.
5. Hold. Code found the quote on Glock's page as it was fetched, and 740 in the quote. The page states grams, so nothing was converted.
6. Corroborate. No other site read in this run stated the figure, so the answer rests on the maker's page alone.
7. Check. A second model, from a different provider, was shown the excerpt and the figure and answered yes.
8. Evidence rules. The source is the maker, so it is citable. The quote is on the page. The page is titled for the Glock 21 SF, so it names this model. The magazine rule reads only up to the next label, so the 825 g on the next line, labelled with empty magazine, does not count against a figure labelled without one. The figure is within bounds. It passed every rule.
9. Escalate. Nothing was in doubt, so no stronger model was asked.
10. File. The result was 740 g at a confidence of 0.95, with the reasoning: read off the maker's page, quote present as fetched, confirmed by a second model. In the measurement it was then scored against the figure we hold, 740 g, and counted right.

Where every output goes

One run leaves more than its answer. Each output has one place, and the ones that change what the next run does are either measurements or a person's decisions.

flowchart LR
  run["One task's run"] --> proposal["The proposal and its trace"]
  proposal --> review{"A person"}
  review -->|"accepts"| catalogue["The catalogue, by pull request"]:::ok
  review -->|"rejects, with a note"| next["The next attempt's question"]:::warn
  run --> calls["A log line per model call"]
  run --> kept["Search answers and page text,<br/>kept a day"]:::muted
  run --> sites["How each site read"]
  sites --> catalogue

Reading a proposal

A proposal is written to be read top down. Its first line is the reasoning, and a doubt, when there is one, comes first in it: what the rule saw and why it held the figure back. Then the figure as the page stated it and in the stored unit, the quote, and each page that stated it, every one a link to open. Then the checks: whether a second site agreed, what the second model said, whether a stronger model was asked, and what the evidence rules decided.

The trace is underneath, for when the answer is not enough: every question put to a model, the excerpts it was shown, every page and search with what came back, and the provider and model of each step. A figure a reviewer doubts can be followed back to the exact words a model was given.