AI Agents for Cosmological Anomaly Detection Test Catalogue

This workspace contains a frozen static catalogue for agent-generated CMB anomaly tests.

The catalogue preserves agent-generated hypotheses, code, results, and summaries. Literature comparison text is included as provenance and should be checked before manuscript use.

Models

• Mimo V2.5 Pro (run_directory): runs/cmb_test_run_mimo
• GPT5.5 (run_directory): runs/anomaly_agent_gpt_run

Local preview

python3 -m http.server 8000 --directory docs

Then open http://localhost:8000.

Regenerate

python3 scripts/build_test_catalogue.py

The default builder reads runs/cmb_test_run_mimo and labels it Mimo V2.5 Pro. To combine multiple model runs, pass each input with a model label:

python3 scripts/build_test_catalogue.py \
  --run-dir 'runs/cmb_test_run_mimo::Mimo V2.5 Pro' \
  --run-dir 'runs/another_model_run::Other Model'

Single JSON exports are also supported:

python3 scripts/build_test_catalogue.py \
  --run-dir 'runs/cmb_test_run_mimo::Mimo V2.5 Pro' \
  --json 'runs/model_b_results.json::Other Model'

The builder copies each test's figures and raw artifacts into docs/, and writes:

• docs/index.html: paper-style landing page.
• docs/catalogue.html: browsable static catalogue.
• docs/tests/<model-id>/<test-id>/: permanent detail page for each test.
• docs/data/tests.json: machine-readable registry.
• docs/data/tests.csv: compact table for audit and downstream analysis.
• docs/data/raw/: sanitized per-test JSON records preserving generated fields.
• docs/data/statistics/: Planck and simulation statistic arrays.

The main pages show the primary p-value for the pre-specified hypothesis tail and an anomaly-significance strength derived from it. Audit fields, including one-/two-tailed p-values, hypothesis-tail sigma, anomaly sigma, and raw agent-reported sigma, are preserved in docs/data/tests.json and docs/data/tests.csv.

Publishing

The docs/ directory is ready for GitHub Pages, Netlify, Cloudflare Pages, or any static file host. For GitHub Pages, publish the repository from the docs/ folder on the default branch.

Paper citation snippet

The full catalogue of agent-proposed tests, including generated hypotheses,
analysis code, numerical results, and diagnostic plots, is provided as an
online resource in the frozen Test Catalogue v1.0.