BRENNEN

Industries / Food, farming and forestry / Agriculture and farming / Dairy

Herd records

Keeping track of each animal: who she is, how much milk she gives and of what quality, when she is in heat, bred and due to calve, when she is sick, and where she moves.

7

Milestones achieved, from 1895 to 2010s.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1895

    Neighbors share milk recording

    Farmers around Vejen, Denmark, formed a control society to measure each member's cows for milk yield and butterfat. Landbrugsavisen, the Danish farm newspaper, says the world's first such societies started in Vejen and on Samsø in 1895. Denmark had about 1,800 societies at their peak in 1935–1960; today about 60% of its 1.5 million cows are recorded, and yearly milk per cow has risen from about 2,000 kg at the start to more than 11,000 kg (Landbrugsavisen).

    landbrugsavisen.dkicar.org

    Credit

    farmers of Vejen and Samsø

  2. 1905

    Cow testing reaches America

    In September 1905 Helmar Rabild, a Michigan state dairy inspector who had seen cow-testing societies in his native Denmark, and Henry Rozema organized a cow-testing association in Newaygo County, Michigan. Michigan State University's archive calls it the first of its kind in the United States. It had 31 members, brought a tester from Denmark, and began keeping records in December 1905.

    findingaids.lib.msu.edu

    Credit

    Helmar Rabild and Henry Rozema

  3. 1951

    Common rules for recording

    After nearly thirty years of talks, recording organizations from several European countries signed an agreement in Rome on 5–9 March 1951 to use the same recording methods, calculations and result formats, and formed the European Milk/Butter Recording Committee, which first met in The Hague on 14 July 1951. ICAR, the international body for animal recording, gives this as its own start.

    LessonRecording had spread without common rules: a 1935 survey counted 34 countries and 4.5 million recorded cows, with tests from every seven days to every two months and lactations counted as 300, 305 or 330 days, so results could not be compared.

    icar.org

    Credit

    the signing organizations, chaired by A. M. Leroy of France

  4. 1977

    Step counters spot heat

    A cow can be bred only in a short period called heat (estrus), which farmers had to spot by eye. C. A. Kiddy strapped pedometers to the lower rear leg of 68 cows and read them at each milking: cows in free stalls were about four times as active in heat as at other times, and cows in comfort stalls about 2.75 times, with little day-to-day change otherwise (Journal of Dairy Science, 1977).

    ebi.ac.uk

    Credit

    C. A. Kiddy

  5. 2000

    Every cow gets a passport

    After the BSE (mad cow disease) crisis shook trust in beef, EU Regulation 1760/2000 of 17 July 2000 required every bovine animal to wear an ear tag in each ear for life, to travel with a passport, and to be recorded in a national computer database of animals, farms and movements.

    eur-lex.europa.eu

    Credit

    the European Parliament and the Council of the European Union

  6. 2010

    US drops national animal ID

    The US Department of Agriculture's National Animal Identification System (NAIS), first proposed in 2002, was meant to trace a diseased animal back to its farm and find the animals it had met. US weaknesses showed after BSE was found in a Canadian-born dairy cow in December 2003, when trading partners were slow to reopen to US beef partly because the herd mates were hard to trace. By 2009 registration covered about 95% of poultry premises but 18% of cattle premises, and on 5 February 2010 USDA abandoned NAIS after spending $150 million since 2004, leaving states to choose their own rules for animals moved between states.

    LessonA national register that put most of its cost on cattle owners never got most of them to join.

    everycrsreport.com

    Credit

    a failure; account by the Congressional Research Service

  7. 2010s

    Robots read milk hormones

    DeLaval's Herd Navigator, a module attached to milking robots, samples milk during milking and measures progesterone, a hormone whose drop marks heat and whose steady level after breeding suggests pregnancy, on average every two days. It has been sold in Europe and Canada since the early 2010s (Bruinjé, Ambrose and LeBlanc, 2025). An independent check on 1,821 inseminations of 715 Holstein cows found that its pregnancy calls caught more than 95% of pregnancies from 27 to 54 days after insemination, but should not be relied on alone before day 41 (Bruinjé and Ambrose, Journal of Dairy Science, 2019).

    pmc.ncbi.nlm.nih.govebi.ac.uk

    Credit

    DeLaval; heat-detection model by Friggens and colleagues (2008)

Watch for

  1. Herd questions answered by AI

    Herdwatch, a livestock records app, introduced Herdi, an AI assistant in its app that it says gives "instant, reliable answers to both herd-specific and general farming questions", and now offers its Herd Intelligence to all subscribers. It has published no accuracy figures.

    Achieved whenan outside party tests its answers against farms' own records and publishes the share it gets right.

    techcentral.ieherdwatch.com

    Claimed by

    Herdwatch, September 2025

  2. A health summary per cow

    GEA showed DairyNet Intelligence at the EuroTier trade fair: AI that reads a cow's milk yield, activity, somatic cell count (white cells in milk, a sign of udder infection), milk conductivity and data from the CattleEye and CowScout systems and writes a short summary of her health, plus a test version that answers set questions such as which cows need attention. GEA says the functions are in testing, do not yet recommend actions, and will reach selected farms in late 2026 or early 2027.

    Achieved whenon working farms, an outside party compares its flags with vets' diagnoses and reports how many sick cows it caught and how many alerts were false.

    gea.com

    Claimed by

    GEA, November 2026

Open

  1. 01

    Test AI herd assistants

    A public test of the questions farmers ask about their own herd, such as "Which cows are due to calve in the next 30 days?" or "Which cows had a high cell count at both of their last two tests?". Each question has a correct answer computed by code from a herd records file, and a scoring script runs the test on any assistant.

    The brief
    Credit · unclaimed
  2. 02

    Open robot milking records

    A public dataset of per-milking records from milking robots on several farms, joined to the health events recorded for each cow (udder infections treated, lameness scores, ketosis tests), with identities removed. Add a baseline early-warning model for udder infection and a leaderboard.

    The brief
    Credit · unclaimed

Open 01 · the brief

Test AI herd assistants

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public test of the questions farmers ask about their own herd, such as "Which cows are due to calve in the next 30 days?" or "Which cows had a high cell count at both of their last two tests?". Each question has a correct answer computed by code from a herd records file, and a scoring script runs the test on any assistant.

Why it is still open

Herdwatch (September 2025) and GEA (November 2026) describe assistants that answer herd questions but publish no accuracy figures. The nearest test found, a Cornell preprint (Liu, Yang and Hostens, December 2025), ran 20 small open models through 30 questions, ten of them about two partner farms' private records, with answers graded by an expert. Searches of Europe PMC and Zenodo found no public test with answers computed from a shareable herd file.

techcentral.iegea.comarxiv.org

Built with what exists

Wageningen University & Research's simulated milking robot records on Zenodo (2025) as a first herd file; Python with pandas to compute the answer key; an open-weights model such as Mistral Large 4 (see the Note "A trillion open parameters") alongside commercial assistants; scoring that also checks whether an answer addresses the farmer's real goal, in the manner of Ai2's XYBench (see the Note "Answering the question behind the question").

  1. First step

    Write 50 questions with two dairy farmers and compute their answers from the Zenodo file.

  2. Done looks like

    At least 300 questions over at least two herd files, one simulated and one real with the farmer's consent and identities removed; exact scoring for number and list answers; and published results for at least three assistants.

Open 02 · the brief

Open robot milking records

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public dataset of per-milking records from milking robots on several farms, joined to the health events recorded for each cow (udder infections treated, lameness scores, ketosis tests), with identities removed. Add a baseline early-warning model for udder infection and a leaderboard.

Why it is still open

The authors of CowScreeningDB, a 2024 public lameness dataset, name "the unavailability of a public dataset which is currently either commercial or privately held" as the main limit on such work; their own set covers 43 cows, with lameness simulated by a block on one claw. Wageningen's 2025 data-space demonstrator had to use a dummy robot dataset that "does not represent real farm data". Zenodo and Europe PMC searches found no real multi-farm robot dataset with health labels.

arxiv.orgzenodo.org

Built with what exists

the herd software's own data export on each farm; Python; LightGBM for the baseline model; Zenodo or Hugging Face Datasets for hosting under an open license.

  1. First step

    Ask one robot-milking farm and its vet for one year of milking and treatment records, and count how many treated cases can be matched to cows and dates.

  2. Done looks like

    At least five farms, each with at least one full year of milkings and at least 500 recorded udder infections in total; the baseline's share of infections caught and its false alerts per 1,000 milkings reported on a farm left out of training; all published under an open license.