- packages/core/src/solver.js: decision tree with perfect recall over the community form's choices; chance events (first dragon, deaths, shadow, meadow dragon, encounter outcome) depend on earlier choices through hierarchical Dirichlet back-off with empirical-Bayes concentration per event; backward induction for P(colour this attempt) and Dinkelbach for P(colour) / (P(colour) + P(other colour)); 200 posterior draws for credible intervals and "still best" rates; cross-fitted held-out check. - Only genuine attempts inform encounter outcomes (goal / Malek intent). - data/public/policy.json (all runs and a low-bias sample) and the readable docs/POLICY.md; published on Pages with solver.js for the bot. - Bot advisor: re-solves from any history, maps advice onto screen text, reports margins so ties can be randomised. docs/BOT_INTEGRATION.md. - Tests on a synthetic world with known answers; daily source update now rebuilds the policy. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UCzTd31GphpMXDhQqWrCMQ
5.6 KiB
Threshing Atlas
Route probabilities and path finding for the Dragonkind threshing: which choices lead to which dragon, how often, how sure we can be, and (in progress) the best route to each colour.
Unofficial fan project. Not affiliated with Dragonkind or the Fourth Wing publishers.
Layout
apps/atlas/ the web page: valley map, route flow, charts, Dragon Register, Reddit post generator
packages/core/ @threshing-atlas/core, no dependencies, runs in Node and the browser:
src/engine.js coding rules and statistics for the community form
src/graph.js the shared route graph: scene graph, decision-variable graph,
bot-to-form crosswalk, Beta credible intervals
src/solver.js best decisions per dragon colour (history-dependent decision tree,
empirical-Bayes back-off, two objectives) + the bot advisor
packages/ingest/ @threshing-atlas/ingest: DragonKind bot importer (fingerprints and choice text only)
research/python/ independent Python implementation of the coding rules and statistics
tests/ JS/Python parity tests (they fail loudly if the two disagree)
scripts/build-site.js assembles the static site into dist/
scripts/build-graph.js combines every source into data/public/graph.json
scripts/build-policy.js solves for every colour -> data/public/policy.json; policy-report.js -> docs/POLICY.md
data/raw/ source exports (form .xlsx; the register export stays local, git-ignored)
data/public/ what the page loads: form_responses.json, register.json, graph.json
data/sources.json where the community spreadsheets and the PDF are downloaded from
data/derived/ coded runs (runs_clean.csv)
data/bot/ anonymised DragonKind bot runs, added by pull request
docs/ ROADMAP.md, BOT_DATA_PROTOCOL.md, REDDIT_POST.md
Run it locally
Requirements: Node 18+, Python 3.9+ (pip install -r research/python/requirements.txt).
npm run dev # builds dist/ and serves it at http://localhost:8000
npm test # JS engine + Python reference + parity checks
New data from the community sources
A GitHub Action (.github/workflows/update-sources.yml) runs every day. It downloads the Threshing Routes
responses, the Dragon Register and the route PDF, and if anything changed it rebuilds the data, runs the
tests and opens a pull request that says what changed: new responses, new or renamed questions, and for
the PDF the pages and lines that are new. The PDF itself is never committed, only a fingerprint
(data/derived/pdf_fingerprint.json). If a link stops working, the Action opens an issue. You can also
run it by hand from the Actions tab, or locally with npm run check-sources.
Best routes (the solver)
npm run policy rebuilds data/public/policy.json and the readable docs/POLICY.md (about 2.5 minutes:
200 posterior draws for two data samples). For every colour it gives the best decision at every point under two
objectives: the chance of bonding that colour on this attempt, and the chance the account eventually ends on that
colour when you die on purpose whenever a run goes wrong. Chance events (which dragon appears, whether the shadow
shows) may depend on earlier choices; how much is learned from the data. See the top of
packages/core/src/solver.js for the model and docs/BOT_INTEGRATION.md for how the bot uses it.
Bot data
The private DragonKind bot's files are converted locally, and only the anonymised result is committed:
DK_ACCOUNT_SALT=... npm run import-bot -- --bot /path/to/DragonKind --tag 2026-11-01
npm run graph
This writes data/bot/scenes.json and data/bot/trajectories-<tag>.jsonl: scene fingerprints, short labels,
choice text and outcomes, with no game prose and no account names. The importer refuses to write anything that
looks like prose. See docs/BOT_DATA_PROTOCOL.md for what the bot should log.
Updating with new form data by hand
In the browser: open the page and use Load a newer export with the responses spreadsheet
(Google Sheets → File → Download → Microsoft Excel). Everything recomputes. Questions are matched by
wording; a renamed question is listed as "not found", and then COLUMN_KEYS in
packages/core/src/engine.js and research/python/clean.py need updating.
To update the bundled snapshot, put the new export in data/raw/form/, then:
cd research/python
python export_raw.py ../../data/raw/form/NEW_EXPORT.xlsx ../../data/public/form_responses.json
python clean.py ../../data/raw/form/NEW_EXPORT.xlsx
cd ../..
npm test
Changing a coding rule
Every rule lives in codeRow() in packages/core/src/engine.js and code_row() in
research/python/clean.py. Change both and run npm test.
Statistics in one paragraph
Rates use 95% Wilson score intervals. Each answer is compared with the other answers to the same question by odds ratio (Woolf CI, +0.5 Haldane correction for zero cells) and Fisher's exact test, with Benjamini–Hochberg q-values. Because Blue- and Black-chasers report rare sightings far more often than players who log every run, key outcomes are also modelled with ridge-penalised logistic regression (λ = 1) that includes data source and chase goal, and route rankings use a Mantel–Haenszel odds ratio stratified by the same factors. The page's Methods section has the details.
Licences
Code: MIT (LICENSE). Data produced by this project: CC BY 4.0 (DATA_LICENSE.md), which also lists
third-party sources that keep their own terms.