RL agent ladders for cold-start skill prediction: sprint, minesweeper, and what transferred - #1

Merged

mwburke merged 68 commits into

mainfrom

claude/game-rl-agents-skill-e87e58

Aug 10, 2026

Commits on Aug 7, 2026

Add Othello, backgammon and Leduc ladders with per-game telemetry

mwburke
and
claude
committed
ad3bd0f
Correct the backgammon and Othello feature sets against measurement

mwburke
and
claude
committed
3bd7c83
Add experiments plan

mwburke
and
claude
committed
5eb063c
Reframe the plan as two committed tracks, not one gating the other

mwburke
and
claude
committed
fc7fdc4
Drop riichi mahjong from the plan

mwburke
and
claude
committed
afa87df
Add E1: real-player ingestion and the cross-game transfer test

mwburke
and
claude
committed
8b08bdc
Add SkillCraft as a second game, and the ablation that qualifies the result

mwburke
and
claude
committed
31ff9ef
Add Lichess as a third game; the E1 shape hypothesis is refuted

mwburke
and
claude
committed
ae6e632
Add NetHack; efficiency is the portable axis, speed is not

mwburke
and
claude
committed
0a531c9
Record E1 as answered, and the Atari Grand Challenge dead end

mwburke
and
claude
committed
5acd8ce
E2: build ladders by handicapping, and the generator turns out to matter

mwburke
and
claude
committed
aea7d3d
E2 complete: all three handicaps, and a refuted prediction

mwburke
and
claude
committed
988a72a

Commits on Aug 8, 2026

Add a distillation ladder, and control E2 for tier spacing

mwburke
and
claude
committed
c8db320
E2 with five generators: undertrained RL checkpoints are the outlier

mwburke
and
claude
committed
1e092e4
Qualify the README headline with what E2 found

mwburke
and
claude
committed
c08d73e
Add WTHOR Othello, the first agent-vs-human comparison, which fails informatively

mwburke
and
claude
committed
24bd443
Scope TETR.IO down to single-player sprint, and verify the range

mwburke
and
claude
committed
b85c89c
Build tetris_sprint: an input-level 40 LINES env on a virtual clock

mwburke
and
claude
committed
85c351d
Add sprint telemetry, the DQN trainer, and the human 40L ingestion

mwburke
and
claude
committed
687e8a1
Add --resume to the sprint trainer

mwburke
and
claude
committed
1b353a9
Record the human 40L target: 396 records, every feature monotone in rank

mwburke
and
claude
committed
0917eb7
DQN cannot bootstrap sprint clears; add a scripted placement-search teacher

mwburke
and
claude
committed
693988c
Add a coupled skill dial after latency alone proved to be pure speed

mwburke
and
claude
committed
99a4a43
Add DAgger after plain cloning stalls at 99.5% agreement and 4 lines

mwburke
and
claude
committed
3bc0462
Teach the bot to play for quads; the range check was hiding that it didn't

mwburke
and
claude
committed
1b8c66e
Add --init so RL can start where cloning left off

mwburke
and
claude
committed
1925f75
Make the skill dial interpolate strategy, and revert what I fitted to humans

mwburke
and
claude
committed
268028f
Make the teacher Markov, because DAgger was asking it an ill-posed question

mwburke
and
claude
committed
d68b2a7
Record the distillation failure and the metric that hid it

mwburke
and
claude
committed
87eadc4
Measure the cold-start ladder against 396 real sprints

mwburke
and
claude
committed
790170f
Run the coherence metric on the humans first; 8.0 ranks is its noise floor

mwburke
and
claude
committed
4494116
Decide holds by two-ply piece ordering, and label it as test-guided

mwburke
and
claude
committed
ab8dfcd
Sample distilled students instead of taking the argmax

mwburke
and
claude
committed
8b45af1
Two-ply hold closes the gap; correct my false column-9 claim

mwburke
and
claude
committed
c36699a
Keystroke cloning fails; split judgement from execution

mwburke
and
claude
committed
933aea4
Reorder the sprint ladder report around what each number is worth

mwburke
and
claude
committed
c9472aa

Commits on Aug 9, 2026

Commit a real passing validation report; the old one was benchmark-only

mwburke
and
claude
committed
eaa75fa
Add DAgger to the placement student, and log the agreement that matters

mwburke
and
claude
committed
c376e81
Record the validation-gate failure and the placement diagnosis

mwburke
and
claude
committed
80ad710
Update the E5 status: latency-as-dial and plain distillation both failed

mwburke
and
claude
committed
6ee12f4
Pin the placement encoding and keep the two agreement metrics distinct

mwburke
and
claude
committed
85538c4
Teach the ladder loader about 44-way placement heads

mwburke
and
claude
committed
64a4a09
Put the E5 result before the detours

mwburke
and
claude
committed
5d18458
Record why plain placement cloning cannot work: 30x the holes

mwburke
and
claude
committed
143f67c
Mask the placement student to reachable targets; it was stalling on illegal ones

mwburke
and
claude
committed
a71f061
Commit the placement target once per piece, and drop when a move is blocked

mwburke
and
claude
committed
a1db367
Put hold back in the student's action space; without it the pathway caps at 31

mwburke
and
claude
committed
73cbdb0
Print the pathway ceiling before training, every time

mwburke
and
claude
committed
303f353
Measure the accuracy cliff that makes distillation unusable as a ladder

mwburke
and
claude
committed
71ba63e
Write up the accuracy cliff and the six faults that hid it

mwburke
and
claude
committed
9346757
Add E6 for sprint: fit on agents, predict humans, against real baselines

mwburke
and
claude
committed
607fc98
E6: the simulator loses to ranking players by pieces-per-second

mwburke
and
claude
committed
5eadb9f
The grid fixes the sign flip and confirms the verdict: pps is unbeatable here

mwburke
and
claude
committed
6a8136d

Commits on Aug 10, 2026

Correct E6: the pps baseline was never cold-start legitimate

mwburke
and
claude
committed
9dda08d
Enforce robots.txt properly; the stdlib parser got the real policy wrong

mwburke
and
claude
committed
203110e
Stop the minesweeper scrape: the only data route writes to their database

mwburke
and
claude
committed
b0b9c08
Spec the minesweeper environment

mwburke
and
claude
committed
19c2242
Write minesweeper reference.py from the spec

mwburke
and
claude
committed
d2d2ee9
Vectorize minesweeper and pass the full battery; add the fully-conv net

mwburke
and
claude
committed
8958c72
Reverse curriculum with a pluggable replay source, wired into the DQN

mwburke
and
claude
committed
e23413c
Supervised mine-probability net: dense labels instead of the sparse Q-signal

mwburke
and
claude
committed
ce6e91b
DQN v2: reprice death, weight the shaping, add an auxiliary mine head

mwburke
and
claude
committed
25df0b4
Beginner and intermediate boards, differentially pinned; outcome in replays

mwburke
and
claude
committed
7e7b59b
E4's yardstick and feature set: the deduction oracle and the telemetry layer

mwburke
and
claude
committed
746bebd
Measure the two-dial beginner ladder, oracle-judged

mwburke
and
claude
committed
66380e2
Three-dial grid: speed x judgement r = -0.03 where sprint had +0.85

mwburke
and
claude
committed
fe1e396
Track the training curves; an ignore rule aimed at checkpoints was hiding them

mwburke
and
claude
committed
031915e
Drop the invariant-violation dumps

mwburke
and
claude
committed
f4fe7a2