Research timeline

← ARI3 overview

Every dated step happened on that date (UTC). Planned steps carry no date.

  1. 2026-05-04

    Repository started

    A headline crawler and a tool server for querying it.

  2. 2026-05-18

    First real crawl

    96 headlines from two outlets. It later became the first published report.

  3. 2026-09-23

    Dated collection begins

    RSS collection every 4 hours into a local database. The evidence gate runs on every change.

  4. 2026-09-23

    First sourced report

    The 2026-05-18 crawl, re-checked against the articles behind it, published with every claim linked.

  5. 2026-09-23

    First human labels, INDUSTRIA trained

    145 labels by the editor. v0.0.1 reaches 0.882 held-out accuracy.

  6. 2026-09-24

    INDUSTRIA committed, EXP-002 pre-registered

    v0.0.1 (frozen the day before) and the next experiment's pass criteria committed publicly, one second apart.

  7. 2026-09-26

    DISCIPLINA frozen

    350 labels, 50 of them hard cases. All five pre-registered hypotheses supported.

  8. 2026-09-26

    Lexicon v1 loaded

    73 style terms chosen by the editor, loaded and matched against stored items. Seven ambiguous terms are kept but left out of every count.

  9. 2026-09-26

    EXP-003B: forecast detection fails its test

    The new head never flagged a forecast (balanced accuracy 0.50, pass mark 0.75). It is frozen and not used.

  10. 2026-09-30

    First-seen times recorded as they happen

    From this point the store records when ARI3 first holds each item. Earlier items carry reconstructed times, labeled as such.

  11. 2026-10-01

    EXP-004: grounded retrieval tested once

    The frozen hybrid retriever reached Recall@10 0.770 and Hit@10 1.000 on 12 answerable held-out questions.

  12. 2026-10-02

    EXP-003A: v0.0.2 on newer articles

    Coverage 0.927 and not-sure share 0.327 met their pass marks. Accuracy 0.833 missed its pass mark of 0.85.

  13. 2026-10-03

    INTEGRITAS released (v0.0.3)

    The release records both EXP-003 results and a frozen forecast head that is not used. The style classifier is still the frozen v0.0.2 model.

  14. LaterPlanned

    PROVIDENTIA

    The system begins looking forward.

  15. LaterPlanned

    CONCORDIA

    Independent models act in concert.