eleticle
Core Log · 2026

Articles from 2026

16 entries, newest first. The whole log is on the Core Log.

  1. No. 16Nov 2026

    Does searching a site’s own work orders help a local LLM as much as a glossary?

    The same local LLM read No. 13's 300 held-out mining work orders with ten of the company's other work orders, found for each by an offline SAG-Wiki search, in place of a 50-line glossary. Meaning came through in 290 notes with the search, 273 with the glossary run the same day (22 against 5 kept by only one) and 283 in No. 13's glossary run.

  2. No. 15Nov 2026

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

    Five public records on Cadia's molybdenum plant, 576 pages, searched in SAG-Wiki with no connection beyond the workstation. The first five passages held the answer in 48 of 49 single-fact searches, but both the feed route and the plant's first filtered concentrate, from two records, in only 6 of 21. Every search returned passages for a shift figure the records do not hold.

  3. No. 14Nov 2026

    How does SAG2 find cited answers in a mine’s records, offline?

    SAG2, run in SAG-Wiki with the search sealed from any network, searched six public mining records for eight plant and safety questions, and a model on the same workstation answered each from the passages returned, citing them. Six of the seven searches with target pages returned those pages; 38 of 41 statements that cite a passage were fully supported by it, counting the record title sent with each.

  4. No. 13Nov 2026

    Does a site glossary help a local LLM read maintenance work orders?

    Each of three mining companies had a 50-line glossary built from 3,200 of its own notes. Across 300 held-out work orders containing shorthand, a local LLM kept the meaning of 283 with the company's glossary, against 228 without it, and neither version said work was done that the note did not report.

  5. No. 12Nov 2026

    Can a local AI assistant tell which size fraction carries the most copper?

    Given a published tailings table and Python, two local models named the fraction carrying the most of the reported copper in all 20 attempts. None said that the 8.2 % of the sample missing from the table might hold copper too.

  6. No. 11Nov 2026

    Can a local model put the right assay into the right database field?

    With each field's method, unit and statistic spelled out, a local model got 83 of 84 fields right in each of three passes over seven certificates, against 55 for rules written from one certificate; its one miss made a gold standard's standard deviation (SD) three times too tight.

  7. No. 10Nov 2026

    What does a local LLM know about minerals?

    Two local models, 976 questions rebuilt from a published test, grading checked by hand. Flash Next named a geology term from its definition 6 times in 10, but recalled a mineral's basic properties only 4 times in 10.

  8. No. 09Nov 2026

    A model predicts grinding throughput with R² = 0.962. What did it actually learn?

    Every one of the dataset's 75,000 numbers, throughput included, comes out of a random-number generator seeded with 42. The regression learned the straight-line part of the throughput formula, not a mill.

  9. No. 08Nov 2026

    Does a written skill help a local LLM read a whole plant flowsheet?

    On each of two plant drawings, Qwen 3.8 Flash Next found more streams in all three attempts with a flowsheet-reading skill than in any attempt without it, and drew fewer wrong lines, in about twice the time. The 27B model gained on one drawing of two in each of two runs, not the same one, and in its rerun took about fifteen times as long with the skill as without.

  10. No. 07Nov 2026

    What hardware does a mineral-processing site need to self-host LLMs and expert models?

    An A$43,568 machine with one 96 GB card, plus a 16 GB tools card, runs GLM-5.3-Flash at an estimated 56 tokens per second for one 128K-token conversation. For the large expert models, RAM and the software stack matter: on one computer, the right stack, with its default drafting, wrote three to five times faster.

  11. No. 06Nov 2026

    Can a local LLM read a mineral processing plant flowsheet?

    Two local models, four flowsheets. Both read the simple drawings. On a whole plant neither got every line right; the one that zoomed in piece by piece recovered more of the loops.

  12. No. 05Nov 2026

    How should you clean plant sensor data?

    Four worked examples on one feed rate tag: the right cleaning depends on what data you were given and what it is for.

  13. No. 04Oct 2026

    Just above cut-off: how certain is the ore/waste decision?

    In this four-case example, a parcel estimated at 1.02 g/t has only a 50 % chance of being above its 1.00 g/t cut‑off, and processing it is still the better call. The estimate, the chance of being ore and the expected margin are three different numbers.

  14. No. 03Oct 2026

    The stope volume matches the design. Can dilution and ore loss still be high?

    The surveyed void matched the planned 10,000 m³ while 10 % of the planned ore stayed in the footwall and hanging-wall waste took its place. Report three volumes, not the net.

  15. No. 02Oct 2026

    Can truck productivity improve while every haul route gets worse?

    Reported fleet productivity rose 15.7 % while both haul routes fell 10 %: the trucks spent more of their hours on the short haul. Compare on a fixed route mix before crediting the change.

  16. No. 01Oct 2026

    Recovery moved by two percentage points. Did the circuit actually improve?

    With the plant unchanged, a reported tailings assay of 0.08 instead of 0.10 % Cu lifts two-product recovery by 1.92 points. What to check before crediting a gain.