- planner/knowledge.yaml: what the game's developer has said (paraphrased, dated, graded), what the Atlas does
with each statement, and the open questions with the experiment that tests each.
- analysis/black/evidence.py -> analysis/reports/BLACK.md: counts from the runs. On the same recorded route the
shadow shows on 21 of 79 first attempts and 1 of 80 later ones; no single route; Blue then Black after the shadow.
- game.yaml: the shadow is a guard (not a roll); what follows it is the Blue, then the Black; on-screen text for
the squad's direction screen.
- planner/experiments.yaml v2: one exact path on every fresh account (fresh-replicate), then one choice changed
at a time (fresh-ablate); shadow-again, second-attempt-same-path, veteran-repeat, veteran-variety; what the bot
records about the account; the sequence to follow once the shadow shows.
- solver.js: assign() takes account facts, account offsets and cycled arms; recommendOption() answers the
squad screen (stay + direction) in one pick.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014jqCC2LZ4x1WHQ1V57nTsv
- game.yaml: `givens` (first or later attempt) and `distinct` levels that are not pooled; the shadow roll
depends on the attempt first. Form submissions show the shadow on 9% of first attempts and 0.3% of later ones.
- Reporting bias: a roll whose bias cannot be measured can borrow another roll's (`borrow: {shadow: meadow}`),
so form submissions inform the shadow again. An assumption, stated in planner/policy.yaml.
- policy.json: targets.<colour>.firstAttempt (this attempt, eventually). Black: about 1.4% on a first attempt,
0.07% on a later one.
- solver.js: advisor options `attempt`, `fix`, sampling "uniform"; assign() and armFor() for the plan.
- planner/experiments.yaml + build_experiments.py -> data/public/experiments.json (progress and status per
experiment, rebuilt from the bot's runs).
- python -m collection cycle: import bot runs, refit, re-plan, deliver the kit; .gitea/workflows/bot-cycle.yml.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014jqCC2LZ4x1WHQ1V57nTsv
A form row recording a Green bonded at the shadow Blue let the "eventually" objective
find a route whose only possible bond was that Green (ratio 1, near-zero chance).
game.yaml now lists the possible bonds per encounter (own colour; the two Greens also
Orange and Brown); anything else is a data-entry error and is not counted.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UCzTd31GphpMXDhQqWrCMQ
- game.yaml: chance events can have `requires` (a gate) and `otherwise`; the shadow
requires "Investigate the scream" and "Help your squadmate" (the route PDF's Black
Dragon Gates; spreadsheet: 78 shadows in 4199 gated tree runs, 2 in 646 others)
- model and advisor: a closed gate fixes the roll; runs off the gate are not counted
- form.yaml: "Protect / Help Squadmate" (new answer) -> Protect squadmate
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UCzTd31GphpMXDhQqWrCMQ
- core/atlas/bias.py: for each rare roll (shadow, meadow Blue) and each source group,
the factor by which the group over-reports the rare outcome relative to complete
logging (bot, bulk logs), with a 90% range; rare outcomes from biased groups count
1/rho; a roll the complete logs have (almost) never shown is learned from them only
- planner/policy.yaml: groups, rolls and thresholds; estimates published in
policy.json under model.reportingBias
- fit_alpha: a roll seen only one way is pooled fully
- docs: goal accounts follow the best route, research accounts use Thompson sampling
(decisionMode atlas_thompson)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UCzTd31GphpMXDhQqWrCMQ
- planner/policy.yaml: which sources the route model learns from (spreadsheet,
reported, bot) and the low-bias sample
- bot runs: deaths coded by stage, Malek intent for runs with a target
- model: survival of the first dragon is not counted when nothing after it is
known; the meadow dragon falls back to what the bot saw
- solver.js: advisor(..., {sampling: "thompson", seed, draws}) follows one
posterior draw per attempt and reports each pick's propensity
Not shipped yet: mixing sources lowers the fitted shrinkage for the shadow roll
(alpha 500 -> 50 in the low-bias sample), see the conversation of 2026-11-09.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UCzTd31GphpMXDhQqWrCMQ
- the importer reads data/accounts/*/runs.jsonl and logs/accounts/*/encounters.jsonl;
with account folders, the top-level files (a mirror of the first account) are skipped
- run records that carry their steps (schema 2) become trajectories directly, keeping
what the bot saw (shadow, meadow dragons), propensities, decision mode and policy id;
older runs are still rebuilt from the decision log; bonded runs without steps are kept
- account names become salted hashes (DK_ACCOUNT_SALT, or a private salt in data/raw)
- runs_clean.csv: new bot_mode column; saw_shadow and saw_blue_meadow from what the bot
saw; deliberate burns say why
- crosswalk: the valley Red, leaving squadmates, squad direction, hold your ground,
hide/approach reactions, the wounded cadet's other two choices
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UCzTd31GphpMXDhQqWrCMQ
- collection/pipeline.yaml declares every source (extract kind, transform,
privacy rules, load target, run source); `python -m collection update|build`
- extract keeps data/raw/<source>/current + previous only (git-ignored);
data/manifest.json records the current raw checksum and change date
- transforms are pure; validate checks runs against data/dictionary.yaml
before load; load writes byte-stable files only when they change
- data/public/runs_clean.csv: every run from every source with a new
`source` column (reddit spreadsheet, reddit thread, reported, bot,
reddit pdf); the sheet's own row label moves to `entry`; run_id is text
- bot importer ported to Python (same output as bot.js on the fixture);
labels and crosswalk in collection/specs/bot.yaml, written into graph.js
- Reddit through the official API (result links to add by hand); the 12
result pages move to collection/manual/reddit_results.yaml
- "Report a run": a GitHub issue form generated from
collection/specs/report_form.yaml, collected as source "reported", plus a
site tab pointing to it, the community spreadsheet and the megathread
- analysis and the route model read runs_clean.csv (sources in
measures.yaml); stats.json and policy.json unchanged apart from the
factor's new name
- daily workflow runs the pipeline; the raw store is kept between runs in
the Actions cache, encrypted with RAW_ARCHIVE_KEY
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UCzTd31GphpMXDhQqWrCMQ
- data/dictionary.yaml: every published file, and every column of runs_clean.csv
with its type, meaning and allowed values
- collection/publish_metadata.py writes data/croissant.json (Croissant 1.0) from
it, with a sha256 per file; `npm run metadata`, run by the daily update
- tests/core/test_dictionary.py: every column and value described, types hold,
croissant.json up to date and checksums match
- attempt is now written as a whole number (was 1.0, 2.0); no statistic changes
- the site ships both files and the Methods text links them
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UCzTd31GphpMXDhQqWrCMQ
Step 4b of the Python-first core.
- core/solver.js is now a 250-line advisor (was 900 lines): it reads the game (core/spec/game.yaml,
published inside policy.json as model.game) and the count tables, and re-solves from any point in
a run with the same interpreter as the Python model. Same API for the site and Ken's bot
(advisor, recommendOption, solverHistory). The bot's on-screen text for each option moves to
game.yaml (bot_text); the deadly rolls carry their risk labels there too.
- tests/core/advisor.test.js: the advisor gives exactly the Python model's values and choices for
every colour and goal (to 1e-12, every decision on every published route), walks a run in the
game's order, and maps advice onto the bot's screen.
- Removed core/engine.js and everything that only compared it with Python (tests/parity/run_engine,
compare_runs, compare_results, compare_routes, compare_solver, compare_stats, analysis/stats/
analyze.py). In their place: tests/core/test_stats.py checks the statistics against known answers,
test_solver.py checks the route model on a synthetic world (tests/fixtures/synthetic_policy.py).
- analysis/network/build-graph.js reads the Python-coded runs; graph.json is byte-identical.
- docs/BOT_INTEGRATION.md: what changed for the bot (history keys, new choices, re-sync all three files).
Claude-Session: https://claude.ai/code/session_01UCzTd31GphpMXDhQqWrCMQ
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
1. Letting the Red pass. Only climbing down quickly puts you above the Red, so a tree run that
passed on the Red and then met the Greens let it pass. Those 16 runs were coded as slow
descents; they are now "Descend quickly: let the Red pass" and count for the Greens'
encounter, so the route model's "let it pass" option rests on its own runs. A tree run with
both the Red and the Greens attempted (impossible in one run) is now flagged.
2. Leaving your squadmates. The form only offered "> Left them" from 2026-10-27 (0 of 504 squad
runs before chose it, 40 of 144 after). Earlier and undated squad runs are now unknown rather
than "stayed together" (form.yaml: left_squadmates_asked_from).
3. Bot text. The wounded cadet's other two choices map onto the form; a test checks every choice
the bot has logged maps onto a form variable (encounter answers excepted, not modelled yet).
Each rule has a test in tests/core/test_coding.py. Data rebuilt (runs, graph, policy, report);
the headline odds move by at most 0.02 points. The Methods tab describes both rules.
Claude-Session: https://claude.ai/code/session_01UCzTd31GphpMXDhQqWrCMQ
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
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>
Step 1 of moving the repo to Python with declarative rules.
- core/spec/form.yaml: every question (how it is found in an export), every answer and what it
is coded as, missing answers, email redaction, burn order, first-dragon pools.
- core/atlas/coding.py reads it and replaces core/python/clean.py (python -m atlas.coding).
- tests/core/test_coding.py: 16 rule tests, one hand-written row per rule, checking the rules
are right rather than only that JS and Python agree.
- analysis/stats/analyze.py takes the colours and missing answers from the spec instead of its
own copy; the daily update job codes rows through the package.
- pyproject.toml for the package; pyyaml and pytest added to requirements.
The coded runs are byte-identical to before (5,988 runs); all parity checks still match.
Claude-Session: https://claude.ai/code/session_01UCzTd31GphpMXDhQqWrCMQ
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Moves only; no behaviour changes. Every file now lives under the part of the project it
belongs to, each section has a README, and docs/PRIVACY.md collects the rules all parts follow.
core/ engine.js, graph.js, solver.js, python/clean.py
collection/ sources/ (form, register, PDF check, daily update), bot/ (importer)
analysis/ stats/ (Python statistics), network/ (build-graph.js), reports/ (POLICY.md)
planner/ build-policy.js, policy-report.js
site/ atlas/ (the page), build.js, qa/
tests/ core/ (unit), parity/ (JS vs Python), fixtures/
Checked against the previous layout: tests pass; coded runs, test outputs, graph.json and
policy.json are byte-identical; the built site differs only in comments that name file paths.
The published file names Ken's bot downloads (solver.js, graph.js, data/policy.json) are unchanged.
Also removes the undefined/ folder of QA screenshots committed by mistake (the QA scripts now
default to a temp folder) and test.txt, and drops the unused npm workspaces.
Claude-Session: https://claude.ai/code/session_01UCzTd31GphpMXDhQqWrCMQ
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>