Step 2 of the Python-first, declarative core. - core/spec/game.yaml: the game as data: every decision and its options, every roll and the earlier steps its odds may depend on, every bonding encounter, and the flow of a run from the entrance to the dragons (set / keep / drop / go / end / encounter). Also how the form's coded answers map onto the game. - core/atlas/model.py: what each coded run reveals about each step, count tables, empirical-Bayes shrinkage, and the posterior draws (same random numbers as before, reproducible from the seed). - core/atlas/solver.py: one interpreter of game.yaml used three ways (solve, evaluate a fixed policy, export the published tree) instead of the tree written out five times. Vectorised over the 200 posterior draws: the build takes about 2 minutes instead of 4, with no memory flag. - planner/build_policy.py and policy_report.py replace the JS build and report. Checked against the JavaScript model: every number in policy.json agrees to within 4e-12 (last-digit differences in log/cos). Three "still best" shares differ by one draw in 200, at near-ties under the "eventually" objective. POLICY.md is otherwise identical. tests/parity/compare_solver.py keeps checking JS against Python in CI until the site's JS advisor replaces solver.js; tests/core/test_game.py checks game.yaml is complete and consistent. Claude-Session: https://claude.ai/code/session_01UCzTd31GphpMXDhQqWrCMQ Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
3.0 KiB
3.0 KiB
Roadmap
The aim is to move from summary statistics to path finding: for each dragon colour, the decision policy that maximises the chance of bonding it, with honest uncertainty.
- Repo and packaging (done): monorepo, CI, GitHub Pages, JS/Python parity tests.
- Graph format (
core/, first version done): typed scene, chance and terminal nodes; history-context keys; per-source counts (bot, form, PDF, Discord); fingerprints and short labels only, no game prose. Importers incollection/botfor the bot'sknowledge.json,runs.jsonlandencounters.jsonl(done), the form (done, via the crosswalk ingraph.js), and a PDF scene crosswalk (to do). Daily source checks for the form, the register and the PDF (done). - Context learning (
analysis/): tests of history dependence (G-tests, BH); variable-order Markov / context tree with BIC; temporal-tiered DAG structure learning (BDeu/BIC); latent "character" model vs specific triggers vs memoryless baseline, compared on held-out likelihood; hierarchical Dirichlet back-off pooling. - Solver (first version done:
core/solver.js,analysis/reports/POLICY.md; still to do: richer history models from step 3): backward induction on the augmented state (scene, history summary), i.e. an influence diagram with perfect recall. Two objectives side by side per colour:- P(target colour on this attempt);
- eventual success when deliberate deaths are free: maximise p_X / (p_X + p_other bond), solved by Dinkelbach iteration. Monte Carlo credible intervals, top-k routes, required steps and bottlenecks.
- Valley graph UI: every option at every point, with P(survive) and P(bond) per colour, the policy overlay for a chosen colour, and untested branches marked.
- Bot loop (advisor and sync published, see
docs/BOT_INTEGRATION.md; experiments.json to do): the atlas publishespolicy.jsonandexperiments.json(ranked by value of information); the private DragonKind bot consumes them and returns anonymised runs.
Python first, declarative rules (November 2026)
The repo is moving to Python, with the rules and the game written as declarative spec files. JavaScript stays only where it has to run: the site's drawing and interaction, and a small route advisor the site and Ken's bot share.
- Coding rules:
core/spec/form.yaml+core/atlas/coding.py, with rule tests. Done. - Route model in Python, built from
core/spec/game.yaml. Done (the bot's text,core/spec/crosswalk.yaml, moves with the bot importer and the JS advisor in step 4). - Fix the coding bugs found in the core review (letting the Red pass, "left your squadmates" before the redesigned form, missing bot text), each with a rule test.
- The site shows precomputed statistics with fixed filter presets; the upload button and
engine.jsgo; a small JS advisor that reads the Python-built model replacessolver.js. data/dictionary.yamlfor every published column, and Croissant metadata for the datasets.