ARI3LLA INDEX

ARI3

ARI3 is the model system behind ARI3LLA INDEX, run as a small research program. Each experiment is a permanent notebook entry with its question, hypotheses, method, results, costs and open questions. Every model version is frozen, hashed and committed to the public repository before its results appear here.

ARI3 v0.0.3 INTEGRITAS is the current research release. The style classifier is still the frozen v0.0.2 model. v0.0.3 adds that classifier's evaluation on articles published after its freeze, the integrity and provenance work behind that test, and a frozen forecast-detection head that did not meet its criterion and is not used in any count. The style classifier reads a news item's headline and feed excerpt, judges whether the item is about style, and says when it is not sure. No ARI3 version forecasts, ranks trends or writes reports. The principles behind the program are on the research philosophy page.

Experiment registry

Entries are permanent. A correction is added as a dated note, never by editing the result.

EXP-004Grounded retrievalClosed
On the 12 held-out test questions with a resolved relevant item, the frozen hybrid retriever reached macro Recall@10 0.770 and Hit@10 1.000.
Read the notebook entry
EXP-003INTEGRITAS v0.0.3ReleasedPre-registered
The evidence can be trusted.
Both parts ran once. Frozen v0.0.2 kept its coverage on later articles but missed its accuracy pass mark, 0.833 against 0.85. The forecast detector never flagged a forecast.
Read the notebook entry
EXP-002DISCIPLINA v0.0.2FrozenPre-registered
The method emerges.
35 hard cases raised confident predictions from 66% to 82% at equal training size.
Read the notebook entry
EXP-001INDUSTRIA v0.0.1Frozen
The work begins.
The first ARI3 model made no confident mistakes on its 34 held-out items.
Read the notebook entry

State of the project

Completed

Dated collection
97 feeds read every 4 hours into a local database, respecting robots.txt
Evidence gate
No report is published without a dated article link for every claim
Perception v0.0.1 and v0.0.2
Frozen, hashed and public
Pre-registration
Every experiment that compares versions commits its pass criteria before its data
Grounded retrieval (EXP-004)
A frozen retriever, tested once on held-out questions
INTEGRITAS (v0.0.3)
Released 2026-10-03. Both parts of EXP-003 ran once, and the results include a failed accuracy criterion and a failed forecast detector

In progress

Lexicon v1
73 style terms chosen by the editor, loaded 2026-09-26 and matched in items stored up to then. Seven ambiguous terms are kept but left out of every count until their uses are checked one by one
Grounded answers (EXP-005)
A prompt and answer schema that answer questions only from retrieved ARI3 items are in development. In the two latest development batches, of five and eight questions, every claim was supported by its cited items, judged only against the evidence each answer was given. A side-by-side test of a prompt that states the schema's size limits found no difference from the current prompt, so the current prompt stays. These are development results, not an estimate of general performance. No final evaluation has run
Forecast ledger
Database tables built. No forecast has been made

Research

Term extraction
Counting lexicon terms across stored articles, with sense checks for ambiguous words
Style R₀
A Hawkes process model of how terms spread between sectors. Designed, not built
Latent salience
A Kalman filter estimate of attention beneath each source's noise. Designed, not built

Not started

Flow between sectors
Optimal transport
Event lift
Synthetic control
Adaptive collection
A reinforcement learning crawler

Release names

Released and in-progress versions link to their notebook entries. Planned names mark intent. None has a date.

  1. v0.0.1INDUSTRIA · The work begins.Released
  2. v0.0.2DISCIPLINA · The method emerges.Released
  3. v0.0.3INTEGRITAS · The evidence can be trusted.Released
  4. v0.0.4PROVIDENTIA · The system begins looking forward.Planned
  5. v0.0.5CONCORDIA · Independent models act in concert.Planned
  6. v1.0FIDES · The system earns trust.Planned
  7. v2.0AUCTORITAS · Others build upon it.Planned

Checking a release

Every entry lists its commit, whose time GitHub records, and the SHA-256 hash of each file. A file downloaded from the linked commit should produce the listed hash. The source is at github.com/arikthehacker/fashion-trend-crawler.