Commit Graph
7 Commits
Author SHA1 Message Date
ekcSchianoandClaude Fable 5.2 ab3664f369 Route model: first attempts are their own shadow roll; experiment plan and automated bot cycle
- 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
2026-11-12 03:31:33 +00:00
ekcSchianoandClaude Opus 5.5 8f1f0383fe Game: which dragons each encounter can give; impossible bonds are not counted
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
2026-11-10 02:36:49 +00:00
ekcSchianoandClaude Opus 5.5 64fd3af235 Game: the shadow is gated on investigating the scream and helping
- 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
2026-11-10 02:33:07 +00:00
ekcSchianoandClaude Opus 5.5 2271fe0976 Route model: correct for reporting bias
- 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
2026-11-10 02:29:09 +00:00
ekcSchianoandClaude Opus 5.5 c0355616b6 Route model learns from bot runs; Thompson sampling in the advisor (work in progress)
- 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
2026-11-10 02:29:09 +00:00
963b08c522 Core: fix three coding bugs from the core review, with rule tests (#11)
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>
2026-11-03 21:12:34 +00:00
cc86343668 Core: the route model in Python, built from a declarative game.yaml (#10)
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>
2026-11-03 21:04:21 +00:00