TOCSIN

"A strange game. The only winning move is to check whether anything actually beats the base rate."

A conflict-forecasting system: reference-class base rates computed from the UCDP record, operational questions that resolve themselves from the next data refresh, and calibration scoring of your forecasts against the prior you started from.

TOCSIN is the forecaster; the analyst audits it and makes it race. Three parts:

1. The model. Ask for a probability and get an explicit reference-class ladder — the unit's own history, its region, the world — conditioned on recency and episode age (conflict is sticky: ~0.17 unconditional becomes ~0.85 for countries violent last year), with empirical-Bayes shrinkage and the class sizes in your face. If the class is thin, you see that. Every prediction shows its work; the walk-forward backtest measures its calibration on ~280k historical unit-years. No astrology.
2. The testable record. Predictions become questions with hard UCDP-pinned criteria that resolve mechanically from the next data refresh — the model's track record accumulates in git, timestamped, with no human in the resolution loop.
3. The arena. Challenger forecasts — other models (VIEWS), naive baselines, or an analyst override — get logged against the same questions and scored head-to-head on identical resolved outcomes: paired Brier, challenger vs engine. That's the question the whole system exists to answer: does anything out there carry information the base rate doesn't?

tocsin/
├── pipeline/          pull (UCDP + G-W) → build (tables, registries) → validate
│                        + acled (aggregates) + site_export (render-ready JSON)
├── engine/            reference-class ladder, recency buckets, EB shrinkage,
│                        walk-forward backtest
└── journal/           question store, auto-resolution, Brier/log/calibration
data/
├── meta.yaml          release + coverage bounds (annual vs candidate)
├── registry/          states, conflicts, dyads, non-state, one-sided (YAML)
├── tables/            country-year, dyad-year, country-month, dyad-month (CSV)
├── backtest.yaml      the engine's own measured reliability
└── site/              JSON exported for the site (`tocsin export`)
questions/             the journal: one YAML per question, git-timestamped
site/                  Astro site: world risk map, Global War Index, WWIII panel,
                         country pages, the journal — Catppuccin Latte/Mocha
docs/                  method (read the caveats), data model, roadmap
tests/                 stdlib unittest; `make verify` gates data + journal

make site-dev (or make site for the static build; needs npm install in site/ once). The board leads with the Global War Index and a world risk map; the watchfloor ranks where conflict is diverging from its own base rate right now (heating/cooling/onset, with ACLED corroboration); the dyads page lists continuation and termination priors; country pages carry the reference-class ladder, walk-forward prior history, deaths, monthly tempo, an 80-year activity strip, and coup history/risk; the methods page shows the walk-forward reliability curves and the VIEWS arena. Catppuccin Latte/Mocha with a toggle; every chart has tooltips and a table view; all SVG is rendered at build time — zero client JS beyond the tooltip layer and theme switch.

$ tocsin pull && tocsin build        # ~280 MB of UCDP into sources/, tables into data/
$ tocsin rate --country Ethiopia --threshold 25
Ethiopia (country 530) — P(sb deaths ≥ 25 in a calendar year)
as of 2026 · bucket: active (history through 2025) · substrate 1989–2025
level     units    yrs   hits    rate       M  posterior
self          1     33     31  0.9394       —          —
region       52    460    379  0.8239     5.4     0.9231
global      181   1112    940  0.8453     5.0      0.927
headline: p = 0.9231  (region posterior, bucket-conditional)

$ tocsin ask --pair "Venezuela,Guyana" --year 2027         # model's prediction stored
$ tocsin call 2027-001 0.03 --source views                 # challenger number logged
$ tocsin resolve                   # grades due questions from the data (provisional
                                 #   on candidate months, confirmed each release)
$ tocsin score
engine (prior)      : Brier 0.041  log -0.113  (n=23)
challenger          : Brier 0.058  log -0.171  (n=23)
challenger vs engine: ΔBrier +0.017 on 23 paired questions — the engine held

tocsin status, tocsin list, tocsin show ID browse the journal; --manual questions resolve by hand; tocsin resolve --id X --void "reason" retires a broken question. Scored forecasts are the last ones made before a question's deciding date — once the threshold crosses, the book is closed (that applies to stale engine priors too).

• Every prior ships with its full ladder: class definitions, counts, conditioning bucket, shrinkage strength. Sensitivity is visible, not hidden.
• Questions are operational by construction (UCDP measure + scope + threshold
  • window), so resolution is mechanical, and ambiguity dies at ask-time.
• Candidate-based resolutions are marked provisional and re-graded on the next annual release; flips are recorded, and the score moves.
• Dyad rates are recurrence rates over UCDP-observed dyads — the engine says so out loud. Read docs/method.md before trusting any number.

The system's findings are written up as a preprint — Transparent Base Rates Approach the State of the Art in Conflict Forecasting: the 12% transparency tax, five preregistered covariate failures, the metric-blindness decomposition, and the equal-weight pool that beats every model. Every number regenerates from this repository.

To cite this project, see CITATION.cff or use the archival concept DOI 10.5281/zenodo.21634193, which always resolves to the latest archived version (v1.0.0: 10.5281/zenodo.21634194). Each release carries the CC BY 4.0 dataset bundle (tocsin-dataset-*.tar.gz, with codebook); rights and attribution for the layered data are documented in DATA-RIGHTS.md.

Python ≥ 3.11 and PyYAML; nothing else. pip install -e . for the tocsin entry point (or run python3 -m tocsin …). make install-hooks turns on the pre-commit gate (make verify: data validation + unit tests).

Data terms and required citations: DATA-RIGHTS.md. UCDP annual releases land mid-year; candidate GED lands monthly with ~6 weeks of lag — tocsin pull && tocsin build && tocsin resolve is the maintenance loop.

Optional: with myACLED credentials in the repo-root .env (ACLED_USERNAME/ACLED_PASSWORD), tocsin acled pulls ACLED's aggregate files (country-month political violence, country-year fatalities, weekly Admin-1 regionals) into sources/acled/ — tempo signals for the future watchfloor, not resolution authorities (different ontology than UCDP; see docs/data-model.md).

TOCSIN: the alarm bell rung when war approaches — from medieval French toquassen, "strike the bell." A watchfloor that rings when the record says risk is rising deserves the word coined for exactly that. (Developed under the codename WOPR, after the WarGames computer that learned about unwinnable games by playing itself; renamed to a word nobody owns. Five covariate families tested against held-out history, five absorbed by the base rates — the humility fits either name.) The scoring loop is the point. Live at tocsin.karazajac.io.