eleticle

Core No. 15Logged 21 Nov 2026Code & data →

Which local model finds the whole answer in a mine’s reports?

SAG-Wiki 0.3.11 · five public Cadia records, 576 pages · keyword search and six local models · no connection beyond the workstation · every list checked by Claude readers against the key

We searched public records on Cadia’s molybdenum plant with SAG-Wiki, an open-source document search application, by keyword and with six local models that judge each passage. For twelve questions, including where the plant gets its feed and when it first made filtered concentrate, we checked whether the first five passages returned held all the evidence.

What we found

  • 576pages indexed and searched, every model running, with no connection beyond the workstation
  • 48 of 49searches for an answer that sits in one passage had it in the first five (7 questions, 7 ways of searching)
  • 6 of 21searches for the feed route and the first filtered concentrate together, from two records (3 wordings, 7 ways of searching), showed both in the first five
  • 11 of 11answerable questions had all their evidence in the first five with Qwen3-Reranker-0.6B on all 576 pages, the only way of searching with both facts of the two-record question in the first five in all three wordings; 9 with MiniLM and KnowLine-4B-Gen5
  • 7 of 7ways of searching returned passages for one shift’s recovery; none of the first ten held it, because the records do not
  1. Method
  2. 01The question
  3. 02Replay a search
  4. 03All questions
  5. 04No answer
  6. 05Speed
  7. 06At your site
  8. Sources

Method

Public records, a closed machine, one count

  • RealFive public recordsNewcrest’s quarterly reports for the December 2021, March 2022 and June 2022 quarters, and Newmont’s Cadia technical report summaries of February 2024 and February 2026, filed with the US SEC. Text taken out on the workstation, one heading per page: 576 pages, 174,000 words.
  • OfflineNothing beyond the machineEach run started in a new Linux network namespace with only the loopback connection (the machine talking to itself), and refused to start if it could see any other. Documents, model weights and software were downloaded beforehand.
  • SearchedSAG-Wiki 0.3.11, two waysFast search ranks passages by keywords (BM25), with no model. Deep search hands keyword candidates to a judging model, keeps those judged needed evidence, searches again and asks again: a median of 112 judgements a question. SAG-Wiki’s default judge is a hosted service; here it was replaced by a service on the same machine.
  • JudgesSix local modelsThree rerankers score each passage against the question alone: TinyBERT, MiniLM, Qwen3-Reranker-0.6B. Three decision models, built for this kind of search, also see the evidence already kept: Hindsight hs-decide-1-flash, Liquid d1-2B and KnowLine-4B-Gen5, the open model the method’s authors tested.
  • ScoredWhole answer in the first fiveBefore any run, a key named the pages that prove each fact and the words that must be visible. Script first, then every list read by Claude readers blind to the way of searching: all 202 judgements agreed, and the 84 on where each fact of the re-asked two-record question first shows. Each ran twice from a fresh index; the lists were identical.

Limits: one site’s public records; twelve questions we wrote after reading the pages (a metallurgist’s unseen questions would test more); SAG-Wiki’s default settings, except smaller decision batches for d1-2B and, for KnowLine, a longer wait and one decision batch at a time instead of 20; one workstation. The records are public and may be in the models’ training. Only passages are tested: no answer was written by a language model.

The questions and key, the decision service, the scripts that extract the text, run every search offline and score it, every list as report, page and line locators, and the protocol with its reviews are in the article’s public repository. The records are not copied there: download them from the links under Sources, then the scripts extract the text.

01 · The question

Two records hold one answer

Where does Cadia’s molybdenum plant get its feed, and what does Newcrest’s December 2021 quarterly report say had been achieved? Newmont’s plant design answers the first half, the quarterly report the second. First asked as “by the end of 2021”, which no record answers (Newcrest dated the milestone to January 2022), the question was asked again with its date on 22 November 2026 and every way of searching run again.

The circuit§14.3.3, p. 14-3
The two copper concentrators send copper-gold concentrate slurry down the overland pipeline to the Blayney filter plant. The molybdenum plant takes its feed from a take-off on that pipeline. Its molybdenum concentrate is thickened, filtered, dried and bagged; its copper-rich rougher tails are thickened and returned to the pipeline. Highlighted: the take-off.

“The molybdenum plant receives feed from the overland copper concentrate pipeline that transports concentrate slurry from the copper concentrators to the Blayney concentrate filter plant.”

Newmont, Cadia technical report summary, February 2026 · §14.3.3, p. 14-3
What had been achievedJan 2022
Filtered moly concentrate production began in January 2022 (Newcrest’s half-year results); the December-quarter report of 28 January 2022 says it had been produced, with commissioning continuing. Commissioning was completed in the March 2022 quarter, and the first shipment was delivered in June 2022. The technical reports give only the year: commissioned in 2022. Selected: Filtered concentrate.

The December-quarter report, released on 28 January 2022, says commissioning was continuing and a filtered molybdenum concentrate had been produced. Newcrest’s half-year results presentation dates the start of that production to January 2022: after the quarter, not within it.

Newcrest quarterly report, December 2021 quarter (released 28 Jan 2022), p. 3; half-year results (17 Feb 2022), slide 11 · reported 28 Jan 2022 · in our words

02 · Replay a search

The first five passages, in order

Each list replays a search recorded on 21 November 2026 (the two-record question on 22 November, asked with its date), in SAG-Wiki’s order, every passage labelled with its report and page. Newmont’s text is quoted, around the evidence where a passage earned a fact; Newcrest’s reports are described in our words, as their publisher’s terms ask. Places and marks refer to each whole recorded passage.

Replay · recorded offline runsMissing: feed route

“According to Newmont’s February 2026 technical report and Newcrest’s December 2021 quarterly report, where does Cadia’s molybdenum plant get its feed, and what does the quarterly report say had been achieved at the plant?”

  1. Newcrest quarterly, Dec 2021 qtr · p. 3

    Cadia costs; then the molybdenum plant: still in commissioning, filtered molybdenum concentrate already produced, work on throughput and quality; then a permit to process 35 Mt a year.

    filtered concentrate
  2. Newcrest quarterly, Dec 2021 qtr · p. 2

    The group expects to meet full-year guidance after maintenance and Cadia’s SAG mill motor project.

  3. Newcrest quarterly, Dec 2021 qtr · p. 19

    Competent Person’s Statement

  4. Newcrest quarterly, Dec 2021 qtr · p. 19

    For personal use only

  5. Newcrest quarterly, Dec 2021 qtr · p. 19

    A note on sources: the group's 2030 production target comes from an October 2021 growth release.

Keyword, all 576 pages: no feed route in the first ten; the filtered concentrate at place 1.

03 · All questions

Single facts were easy; feed and first concentrate together were not

One square per question and way of searching. Seven questions need a single fact and one needs two dates from two quarterlies; nearly every search had them in the first five passages. The first three are the question above in three wordings, and need the plant design and the December-quarter report.

Whole answer in the first five passagesAll 576 pages
How many of the 11 answerable questions had the whole answer in the first five excerpts, all 576 pages: Keyword 8; TinyBERT 8; MiniLM 9; Qwen3-Reranker 11; Hindsight 8; d1-2B 7; KnowLine 9.

Questions: 01–03 the feed and the first filtered concentrate (naming both reports, in site shorthand, and requiring the 2026 edition; asked with their date on 22 November); 05 the quarter commissioning finished; 06 the first shipment; 07 the copper tails; 08 the design capacity; 09 the grade; 10 the reagents; 11 dewatering and bagging; 12 commissioning and first shipment together. 04 has no answer: section 04.

All 576 pages: most complete, Qwen3-Reranker, 11 of 11. Both facts of 01–03 in the first five: Qwen3-Reranker in 3 of 3 wordings, KnowLine 2, MiniLM 1; the others in none.

Keyword search, all 576 pages

The first filtered concentrate came first in all three wordings, and the feed route was not in the first ten in any: all 30 of the first ten passages came from the quarterly reports. The re-run changed the wording as well as the date, so it does not show why.

Deep search, all 576 pages

Of the 18 lists, 8 had both editions’ plant-design pages in the first five, and 7 repeated the same feed sentence from both. Nine had the first filtered concentrate in the first five.

04 · No answer

Every search returned passages anyway

What was molybdenum recovery on day shift on 14 March 2022? That needs plant data: feed and concentrate tonnes and molybdenum assays for the shift, lined up in time. The five records hold none of it, and every way of searching still returned at least ten passages.

One shift’s recovery · first ten passages · all 576 pagesnone for a shift
Keyword: 10 passages returned, 0 of the first ten show a molybdenum recovery figure, none for a shift; TinyBERT: 13 passages returned, 2 of the first ten show a molybdenum recovery figure, none for a shift; MiniLM: 13 passages returned, 2 of the first ten show a molybdenum recovery figure, none for a shift; Qwen3-Reranker: 11 passages returned, 2 of the first ten show a molybdenum recovery figure, none for a shift; Hindsight: 13 passages returned, 3 of the first ten show a molybdenum recovery figure, none for a shift; d1-2B: 27 passages returned, 6 of the first ten show a molybdenum recovery figure, none for a shift; KnowLine: 77 passages returned, 0 of the first ten show a molybdenum recovery figure, none for a shift.

Rust: a passage showing a molybdenum recovery figure from the technical reports, such as the life-of-mine forecast (about 72 % in the 2024 edition, 67 % in 2026) or the recovery formulas. None is for a shift or a day. Two ways of searching had one in the first five.

A search returns its best passages whether or not the answer is there. Read in a hurry, a life-of-mine forecast can pass for a shift figure; a search screen for a site needs to say what the records cannot answer.

05 · Speed

Seconds on one workstation

Median time from question to list over the twelve questions on all 576 pages, first run. Keyword search needs no model; every model ran on an RTX 4080 SUPER memory-modified to 32 GB (the retail card has 16 GB), beside its other work, KnowLine with the card’s speech-recognition service paused.

Median time per questionlog scale
Median of the 12 questions, first run: Keyword 0.010 s on the CPU; TinyBERT 0.110 s on the GPU; MiniLM 0.140 s on the GPU; Qwen3-Reranker 1.140 s on the GPU; Hindsight 2.725 s on the GPU; d1-2B 3.695 s on the GPU; KnowLine 46.745 s on the GPU.

Keyword times are from SAG-Wiki’s local web interface, already running; deep-search times include starting its command-line tool. About 7.5 GB of the card was free beside its other work: d1-2B fitted only with SAG-Wiki’s decision batches cut from 32,000 to 16,000 tokens, and KnowLine-4B-Gen5, whose process peaked at 21.6 GiB here, ran while the speech service was paused, one decision batch at a time. On the 24-core processor KnowLine had not finished one question in a quarter of an hour.

06 · At your site

Before a site relies on offline search

  1. Test the models on your own questionsOn all 576 pages of this archive a 0.6-billion-parameter reranker had the whole answer in the first five for 11 of 11 questions, the only one with both facts of the two-record question in all three wordings; the three decision models, up to 4 billion parameters, 9, 8 and 7. A model’s standing on public tests does not settle it for your records.
  2. Keep one edition, or label itThe same plant-design paragraph from the 2024 and 2026 reports often took two of the first five places.
  3. Check how tables come throughTable cells came through as one-word passages, “Metric” or “0.7934”, each taking a place on the list.
  4. Make the missing visibleAsked for one shift’s recovery, every search returned passages, none holding it, with nothing on the list to show that. Point readers to the historian and the lab for what reports do not hold.
  5. Keep the source beside the passageReport, edition and page on every passage let a reader check the answer, and see when two dates describe different milestones.
  6. Run a field test before relying on itHave a metallurgist write unseen questions in the site’s own shorthand, a second person name the passages each answer needs before the run, and time how long a reader takes to check the whole answer, from the right edition, on the site’s machine. Measure the graphics memory free under the site’s normal workload first: KnowLine-4B-Gen5’s process peaked at 21.6 GiB here.

Sources

  1. Newcrest Mining. Quarterly report for the quarter ended 31 December 2021, 28 January 2022, page 3. PDF.
  2. Newcrest Mining. Half-year results presentation, 17 February 2022, slide 11: filtered molybdenum concentrate production from January 2022. PDF.
  3. Newcrest Mining. Quarterly report for the quarter ended 31 March 2022, 28 April 2022, page 3. PDF.
  4. Newcrest Mining. Quarterly report for the quarter ended 30 June 2022, 21 July 2022, page 3. PDF.
  5. Newmont. Cadia Valley Operations, New South Wales, Australia: Technical Report Summary, February 2024. Exhibit 96.6 to Form 10-K, filed 29 February 2024, §14.3.3. SEC EDGAR.
  6. Newmont. Cadia Valley Operations, New South Wales, Australia: Technical Report Summary, February 2026. Exhibit 96.2 to Form 10-K, filed 19 February 2026, §14.3.3. SEC EDGAR.
  7. SAG-Wiki, open-source document search, MIT licence; CLI 0.3.11 from npm. github.com/Zleap-AI/SAG-Wiki.
  8. Wu, Y., Liang, X. and Li, G. (2026). SAG2: embedding-free multi-hop retrieval with BM25 and decision models. arXiv:2610.22588 v2, Appendix G. arxiv.org/html/2610.22588v2.
  9. Models, from their Hugging Face pages at the revisions in the repository: TinyBERT, MiniLM, Qwen3-Reranker-0.6B, hs-decide-1-flash, d1-2B, KnowLine-4B-Gen5.