The exploiters ran 32.5M learner actions, not the 65.5M I recorded -- they train
one seat, so a 250x1024 run yields half of what the same shape gives the
both-seat self-play trainer. Against targets trained on 131M and 122.6M, that
makes the attacker roughly 4x underfunded.
Which means the absolute number does not support "ours is only 22.8%
exploitable". It supports exactly one claim: at a matched budget, league gives up
more. Whether the ordering survives a properly funded attacker is now the open
question, so the script takes --updates and --batch-games to run it.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XBQKgvBbxbheiTF1AVy1Sh
Winning the head-to-head says a policy is strong on average, not that it is hard
to beat. So freeze each policy, train a fresh one from scratch whose only job is
to beat that policy, and see how far it gets.
Same exploiter budget (250 updates x batch 1024, 65.5M learner actions):
ours (match stack, 131M) exploiter reaches 0.2278 [0.219, 0.237]
league (web-deployed) exploiter reaches 0.3213 [0.311, 0.331]
League gives up 9.4 more points to a dedicated attacker. Both sit far from 0.5, so
neither is a pushover -- but the caveat left open by the head-to-head is now closed
in our favour on both axes: stronger on average (0.6094) and harder to exploit.
Worth noting against expectation: league was trained *with* an exploiter structure
and we ran pure self-play, and we still come out less exploitable. Whatever the
league machinery buys, it did not buy that here.
The number is a lower bound -- a bigger exploiter would find more -- so it only
means anything as a like-for-like comparison, which is how it is used.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XBQKgvBbxbheiTF1AVy1Sh
Each piece switched off in turn, trained at identical compute, then played against
the full stack over 8192 duplicate matches. Below 0.5 means the removed piece was
doing work.
- both seats: 0.3317 [0.322, 0.342]. The biggest single contributor. Half of it is
simply sample count -- dropping the opponent seat halves the learner actions per
update -- but that is the point: self-play already produced those plies with the
same network, and the old trainer stop_gradiented them away.
- match observation: 0.4751 [0.464, 0.486]. Small but real. Since carry itself
contributes almost nothing (rounds decompose), most of this is likely the
single-round observation defects being fixed: to_move, the deck clock, and the
score_diff scale.
- privileged critic: 0.5160 [0.505, 0.527] -- turning it OFF makes the agent
significantly STRONGER. Fable called this the biggest missing idea; it is wrong.
A critic that knows the deck fits V(full state), which is not
E[return | masked obs], so the advantage picks up a component the actor cannot
act on. From the actor's side that is noise, not variance reduction. Asymmetric
critics hurting under partial observability is a known failure mode.
Defaulted off accordingly. (Reusing it as a PIMC leaf evaluator may still stand --
that is a separate claim from using it to train the policy.)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XBQKgvBbxbheiTF1AVy1Sh
Self-play works: matches converge to 146.8 plies (~49 a round) with a 91% deck-
race rate, so the stalling that static opponents induced is gone. Duplicate match
eval scores 0.4968 with a mean lead of exactly 0.0 -- same deals, same coins, both
seats, deal luck cancelling exactly.
Success criterion 1 does not pass. The carry probe is close to flat: expeditions
opened sit at 5.00 whether the policy is 60 points down or 60 points up. Wager use
does move monotonically across all six carry levels, and in the right direction
(behind -> more multipliers), but the spread is 0.31 wagers.
Two diagnoses, one of which was mine and wrong:
- Residual potential shaping was NOT the cause. Annealing it fully to zero left
the probe just as flat.
- terminal_scale is. At carry -60, tanh((margin - 60)/50) is close to linear over
any realistic round margin, and maximising E[tanh] on a linear stretch is just
maximising E[margin] -- there is no reason to gamble. Risk-seeking only appears
where tanh is sharply convex, which needs a smaller scale. Dropping 50 -> 12
widens the wager spread 0.19 -> 0.31, which is the mechanism showing up.
The probe itself is also mis-scaled: at scale 12, tanh(60/12) is 1.0, so +/-60 is
a saturated dead zone with no gradient and the policy has learned nothing there.
The measurable band is |carry| <~ 2 * terminal_scale, and the probe levels have to
be set from the scale rather than fixed.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XBQKgvBbxbheiTF1AVy1Sh
50-iter sweep on default.yaml with stronger MCTS exploration:
- c_puct: 3.0 -> 5.0 (UCB weight, more exploration of low-prior actions)
- root_dirichlet_epsilon: 0.25 -> 0.4 (more noise injected at root prior)
Standalone eval at iter 50 (30 games/opponent, all natural-end, timeouts=0):
- vs heuristic-balanced: W=0/30 S=-70.5 (PA 0.15)
- vs heuristic-aggressive: W=2/30 S=-65.1 (PA 0.14) [+10, +4]
- vs heuristic-cautious: W=1/30 S=-48.0 (PA 0.14) [+2]
3 natural-end wins vs prev trapfix baseline iter 44 (which had 0 natural
wins + 1 timeout-tie). Stall trap fixed remains true (timeouts=0 in c1).
Trade-off observed: more exploration -> higher variance. Score avg vs
cautious worsened (-32 -> -48), but win events appeared. For the
non-terminal-win objective, exploration win > score-avg loss.
Next: commit to long run (300 iter) with these params before tuning more.
Three coordinated hygiene changes; none target the diagnosed
selection-bias bottleneck. They make the codebase honestly reflect the
pure-self-play stance and reduce dashboard noise.
Bot rename (drop the unhelpful safe_ prefix; suffixes describe behaviour):
- safe_heuristic_loose -> heuristic_aggressive
- safe_heuristic -> heuristic_balanced
- safe_heuristic_strict -> heuristic_cautious
- noisy_safe -> heuristic_noisy
- passive_discard -> discard_only
Class renames in bots/: SafeHeuristicBot -> HeuristicBot,
SafeHeuristicParams -> HeuristicParams, PassiveDiscardBot -> DiscardOnlyBot,
plus loose/strict parameter constants. Backwards compatibility was dropped
intentionally per user instruction; no aliases. Active configs, docs,
scripts, tests updated. Archive directories (configs/archive,
docs/archive, runs/archive) left intact and may still reference old
names per their read-only policy. The src/.../bots/passive.py module was
renamed to discard_only.py via git mv.
Analyze plot curation (deep_cfr/analyze.py):
- Added analysis_00_core.png as the canonical daily dashboard with 10
heuristic-free metrics (loss/{advantage,strategy}; vs heuristic_cautious:
avg_score_diff0, win_rate0, avg_opened_colors, positive_expedition_rate,
bonus_expedition_rate, score_per_opened_color, policy_entropy; vs random:
win_rate0).
- Removed analysis_05_open_quality.png (bad/weak/good open rates,
recoverable score) and analysis_07_calibration.png (calibration gap,
recoverable mean) - both relied on the heuristic recoverable_score
classifier already dropped from inputs.
- Removed SELECTIVITY_PLOTS and plot_selectivity (heuristic-laden).
- SUMMARY_EVAL_METRICS no longer includes bad_open_rate or
calibration_gap.
- PlotSpec gained an opponents allowlist so the new core section can pin
a specific opponent per panel without restructuring plot_section.
Tiered evaluation cadence (EvaluationConfig):
- Added extended_opponents and extended_eval_every (default 0 = disabled).
- opponents_for_iteration(iteration) returns the core list every
eval_every and appends extended_opponents (de-duplicated) when
iteration is also a multiple of extended_eval_every.
- default.yaml now uses 3 core opponents (random, discard_only,
heuristic_cautious) every 5 iterations and 3 extended opponents
(heuristic_balanced, heuristic_aggressive, heuristic_noisy) every 50
iterations. random is the floor sanity. discard_only is the
zero-pit detector / absolute-score reference (its score is always 0,
so eval/discard_only/avg_score_diff0 directly equals the model's raw
average score). heuristic_cautious is the ceiling and the
archive-comparable benchmark used in the prior diagnostic sections.
Net eval cost reduction: roughly 50% (3 opponents x every 5 iter, plus
6 opponents x every 50 iter, vs the prior 6 x every 5).
Documented in docs/plans/deep-cfr-selectivity.md section 9.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The interleaved traversal scheduler's _regret_matching was hard-coded to
spread fallback policy uniformly across legal actions, regardless of the
configured regret_matching.all_negative_fallback. default.yaml has
shipped with all_negative_fallback: argmax_tiebreak since 618d5f8 based
on the 20-iter audit + 1000-iter empirical comparison in
docs/archive/deep-cfr-regret-fallback-audit-2026-05-07.md, but the
default scheduler was switched to interleaved in 09bbe7c, after which
the configured fallback mode silently no-op'd.
_regret_matching now takes fallback_mode and concentrates policy mass on
the lowest-index tied action when "argmax_tiebreak". Tiebreak is
deterministic; the Cython recursive traverser randomises ties using its
per-traverser RNG, which the batched policy does not have. Behaviour
matches the spirit of the recursive path (concentrate on best, do not
dilute uniformly).
Plumbed through BatchedPolicy, InterleavedTraversalConfig,
run_interleaved_traversal_batch, trainer.py, workers.py, and the
analyze_first_open_targets.py caller. Two unit tests added.
Also bumps default.yaml outcome_sampling_epsilon 0.2 -> 0.05. The
200-iter sweep in docs/plans/deep-cfr-selectivity.md section 1 showed
0.05 produced the best short-run safe_heuristic_strict score diff
(-40.01 vs -57.87 for 0.20). Recent experiments already used 0.05; the
default now matches actual experimental practice.
Neither change targets the diagnosed selection-bias bottleneck. They
align config intent with scheduler behaviour and make the default config
reproduce known-best knob settings out of the box.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Audit of the Deep CFR information-state encoding identified two tiers of
non-pure features and removed both:
- Tier 3 (judgment): is_bad_open_candidate, open_risk_score,
is_safe_continuation. Same heuristic family used to label bad_open in
evaluation, embedded as model input.
- Tier 2 (projection): recoverable_score_no_bonus,
recoverable_margin_no_bonus, min_needed_to_break_even,
cards_needed_for_bonus, has_bonus_path. Mechanical but assumption-laden
("commit and play all currently-playable cards"). The no_bonus form is
asymmetric: it amplifies the immediate -20 penalty while truncating the
+20 bonus upside, biasing the model toward the same "don't open" basin
the diagnostics already flagged.
Input dim 365 -> 297. DERIVED_PLAYABILITY_PER_COLOR 19 -> 15;
SLOT_AWARE_PLAYABILITY_PER_SLOT 12 -> 6. Test shape assertions updated.
Also adds selectivity diagnostic infrastructure used to reach this point:
- traversal.outcome_unsampled_first_open_prior_alpha config field with
signed-prior overlay on unsampled first-open advantage targets (A1).
- analyze_first_open_counterfactual.py --post-policy to swap the
policy_player rollout policy and isolate selection bias (D1).
- analyze_first_open_followup.py to inspect post-forced-open behavior
(E2): same-color play vs discard counts, other-open rate, terminal
hand composition.
Findings recorded in docs/plans/deep-cfr-selectivity.md sections 3-6.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Second survey processed the remaining 12 archives via gemini after
the first batch of 3 was accepted. 12 drafts, 0 skips, 0 errors.
Every draft carries a deterministic Last-verified header
(2026-05-08, commit 5c221fb) thanks to the post-processing fix
landed in the previous commit. All 12 accepted into docs/research/
verbatim:
deep-cfr-evaluation-profile-plan
deep-cfr-legacy-experiment-reproduction
deep-cfr-legacy-runtime-comparison
deep-cfr-performance-experiments
deep-cfr-profile-advantage-memory-split
deep-cfr-profile
deep-cfr-regret-fallback-audit
deep-cfr-v0-gap-vs-coolrl
deep-cfr-v0-plan
fast-engine-next-optimizations
post-a-optimization-calculus
test-coverage-notes
docs/archive/ is now fully covered: every entry either has a
research counterpart by stem or by tail-match.
Also extracts _dispatch_one and adds --parallel N to
scripts/librarian_survey.py. ThreadPoolExecutor over the per-archive
work is safe because subprocess.run is network-bound (no GIL fight)
and each thread writes to its own output filename. Default stays
1 (sequential); --parallel 4 is the recommended speedup for large
surveys. The two surveys above ran sequentially; future runs can
opt in.
Plan declares librarian closed for new feature work. MEMORY drift
fixup and duplicate-merge modes stay deferred until a real input
surfaces. Stage 1 (5 deterministic checks) and Stage 2 (promote +
survey) remain operational.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Spot-check of the three drafts gemini produced in the --max 3
survey smoke test: all cited file paths exist, line numbers and
function names land within 1-2 lines of actual symbols
(game.pyx:217 cdef class GameState, evaluate.py:220 batched-entropy
block, trainer.py:892 _evaluate_parallel, action_distribution at
evaluate.py:238 with cited code at line 250 inside it). Numbers
cross-checked against archives match. Conclusions preserved.
The one systemic weakness was the Last-verified commit field:
gemini left a `<short-hash>` placeholder, a literal `HEAD`, or
omitted the commit entirely depending on the call. Fixed in two
places:
1. Manually patched the three drafts before acceptance and copied
them into docs/research/.
2. Added _current_commit_sha and _post_process_draft helpers to
both librarian_survey.py and librarian_promote.py. The drafts
now go through `**Last verified:**` line normalization that
substitutes today's date and `git rev-parse --short HEAD`
before being written to disk. Future runs converge
deterministically.
Net: docs/research/ gains classic-port-notes.md,
deep-cfr-batched-evaluation.md, and deep-cfr-evaluation-profile.md.
12 archive entries remain unprocessed.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
scripts/librarian_survey.py walks docs/archive/*.md and dispatches
every entry without a research counterpart through the same prompt
assembly as librarian_promote.py. Outputs land under
runs/tmp/librarian-survey-<timestamp>/, classified into:
<stem>.md draft, ready to copy into docs/research/
<stem>.SKIP.txt LLM's one-line "not promotable" reason
<stem>.ERROR.txt CLI stderr if the call itself failed
Counterpart detection uses exact stem match plus a tail-match
heuristic so research notes that intentionally drop a domain prefix
still suppress their archive. Verified against the current tree:
docs/archive/deep-cfr-opponent-policy-network-divergence-* is
correctly recognized as already covered by
docs/research/opponent-policy-network-divergence.md.
--dry-run lists candidates and suggested research targets without
calling the LLM. --max N caps processed archives per run, useful as
a cost guard. Sequential dispatch; one LLM call per archive.
Plan updated to mark survey mode complete.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
scripts/librarian_promote.py gains a --accept flag that copies the
generated draft to the suggested docs/research/ target in the same
invocation. This is explicit per-invocation opt-in, not auto-apply:
the operator types --accept knowingly, the cp is still a deliberate
acceptance decision, just expressed in one command instead of two.
docs/research/option-a-bench-result.md is the first promoted note,
generated by gemini from
docs/archive/option-a-bench-result-2026-05-07.md and accepted
verbatim. Spot-check verified that cited file:line locations match
current source (traversal.pyx:473-475 and
inference_server.py:226-228 carry the cited code), the Last-verified
header reflects today's date and HEAD, and the durable conclusion
(sync-blocking policy boundary as the structural ceiling, not IPC
plumbing) is preserved.
This closes the end-to-end loop the librarian was designed for:
oversize check surfaced docs/performance.md, the routing pass split
durable analysis into the archive entry, the promote dispatcher
turned the archive entry into a research draft, and the human
accept step landed it as a tracked research note. Took one LLM call.
Removes the now-resolved ignore-list entry for
docs/research/option-a-bench-result.md (the file exists; the
forward reference is real).
scripts/librarian.sh exits 0 against the working tree.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
scripts/librarian_promote.py is the first Stage 2 piece: a
vendor-agnostic LLM dispatcher that drafts a docs/research/ note
from a given docs/archive/ entry. It assembles the prompt by
stitching scripts/librarian-prompt.md (system) onto the archive
body with a "draft a research note per the rules above" task
instruction, then shells out to the CLI selected by LIBRARIAN_LLM
({claude|codex|gemini}; default claude). The LLM's stdout is
captured to runs/tmp/librarian-promote-<timestamp>-draft.md for
human review — the script never writes into docs/research/ itself.
Refusal cases:
- path not under docs/archive/
- target docs/research/<stem>.md already exists (after stripping any
-YYYY-MM-DD suffix)
- archive missing
If the LLM judges the archive non-promotable, it is instructed to
return a single SKIP: <reason> line instead of fabricating a draft.
--show-prompt prints the assembled prompt without invoking the LLM,
useful for inspecting what would be sent.
Plan updated: Stage 2 v1 marked complete; remaining Stage 2 work
(MEMORY drift fixup, duplicate-merge, survey mode) catalogued.
Next concrete step is a smoke test against one real archive entry.
Adds one ignore-list entry for docs/research/option-a-bench-result.md
which appears in the plan as a hypothetical accept target.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds the last two deterministic Stage 1 checks and wires them into
scripts/librarian.sh:
- librarian_check_stale_plans.py: per docs/plans/*.md, uses
`git log -1 --format=%cs` to get the last commit date and flags
plans untouched in 60+ days. Plans should be either active or
archived; long silence is drift.
- librarian_check_memory_drift.py: validates the user-memory dir
(~/.claude/projects/<slug>/memory). Each MEMORY.md index line
must point at a real file with frontmatter (name, description,
type ∈ {user, feedback, project, reference}); no orphaned memory
files. Memory dir is derived from repo root for portability.
Both run clean against current state.
Stage 1 declared complete. Promotable-archive and duplicate-prose
detection are moved to Stage 2 in the plan because both need LLM
judgment to avoid false positives — not pattern matching.
Next concrete step is the open Stage 1 finding from yesterday:
docs/performance.md at 914 lines, due for a split into sub-topic
notes.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Third Stage 1 piece: scripts/librarian_check_oversize.py walks
non-archive markdown files and flags any over the 500-line soft cap
declared in AGENTS.md. Wired into scripts/librarian.sh.
Caught one real finding on first run: docs/performance.md at 914
lines. Splitting it into sub-topic notes is a separate cleanup task
— surfaced for the user, not auto-applied.
Updates docs/plans/librarian.md Progress + sets the next concrete
step to a stale-plan checker.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Single entry point that runs every Stage 1 check in order and
aggregates exit codes — both checks always run so users see all
findings in one pass. Two checks wired up today (lychee link
integrity, file:line citations), more land incrementally as Stage 1
grows.
Removes scripts/librarian.sh from the ignore list now that the file
exists. Updates docs/plans/librarian.md Progress + Next Step.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
scripts/librarian_check_citations.py extracts file:line references from
inline-code spans across docs/**/*.md and verifies each path exists
(and, if cited with a line number, is within range). Skips
docs/archive/ and docs/plans/archive/ which are read-only by policy.
Caught one real drift in docs/research/optimization_sequencing.md: the
note pointed at docs/plans/amp_trainer.md, which had moved into
docs/plans/archive/.
scripts/librarian-ignore.txt holds fnmatch globs for citations that are
intentionally future-tense (planned files described in the plan docs
themselves). Used sparingly so the checker stays useful as a drift
signal.
Updates docs/plans/librarian.md Progress + Next Step. Next concrete
step is a thin scripts/librarian.sh orchestrator over both checkers.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Captures the full librarian design (two-layer architecture, three-stage
pipeline, vendor-agnostic via LIBRARIAN_LLM env, propose-only / no
auto-apply) in docs/plans/librarian.md. Lands the first concrete Stage 1
piece: scripts/librarian_check_links.py, a lychee --offline wrapper
ported from ~/dev/coolrl/src/coolrl/dev/check_doc_links.py.
Also moves the librarian prompt from .claude/agents/ (Claude Code only)
to scripts/librarian-prompt.md so any CLI can load it as a system
prompt later. Fixes one stale README link the new checker caught:
docs/classic-port-notes.md → docs/archive/classic-port-notes.md.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds torch.autocast(fp16) + GradScaler around _train_advantage and
_train_strategy when run.use_amp=true and device=cuda. CPU/non-CUDA
falls back to fp32 no-op. Mitigations:
- scaler.unscale_(optimizer) before grad_clip.
- nonfinite-loss guard skips overflowing batches and counts them.
- diff.float().square() in advantage loss to avoid fp16 overflow.
- strategy mask/log_softmax kept in fp32.
New metrics: amp/grad_scale, amp/nonfinite_loss_count.
Tests: AMP CUDA smoke + CPU fallback in test_deep_cfr_trainer.py.
Bench: scripts/bench_amp_trainer.py micro-benches train phases under
synthetic replay memory. smoke.yaml result is fp32 3.22ms / AMP 3.92ms
(0.82×, regression). 100-iter A/B on default.yaml deliberately
skipped: smoke regression mirrors the 2026-05-07 torch.compile
regression dynamic (dispatch overhead > kernel benefit at this model
size) and re-confirming on the same size adds no information.
Default stays run.use_amp: false. Re-enable trigger documented in
docs/performance.md: hidden_size >= 1024 or num_layers >= 6, then run
the bench script + 100-iter A/B before flipping default.
Implements the central inference server pattern: a dedicated GPU
process owns advantage/strategy/league networks, batches policy
requests across traversal workers via shared-memory tensor pool, and
returns logits. Workers route forward calls through InferenceClient /
NetworkProxy when traversal.inference_backend == "server".
Default remains traversal.inference_backend: local. The server
backend regresses iter time ~3.8× on the inspected default config
(small-model dispatch + sync-blocking traversal capping realized
batch at ~num_workers=8 instead of the bs=64-256 needed to amortize
IPC overhead). Keeping the implementation behind the flag lets us
re-enable when (a) model size grows, (b) per-worker interleaved
traversal lands, or (c) eval becomes dominant — see
docs/performance.md "Option A Bench Result and Structural Ceiling"
for the full diagnosis.
Plumbing included:
- inference_buffers.py: shared-memory tensor pool with slot
management.
- inference_client.py: per-worker client + NetworkProxy adapter for
the existing traversal.pyx call sites.
- inference_server.py: spawn-context server process with
batch-window aggregation, weight sync, shutdown sentinel.
- bench_inference_backend.py: A/B between local and server backends
with eval/checkpoint disabled.
- test_inference_server.py: round-trip and integration tests.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Measured DeepCFRMLP forward at bs={1,4,16,64,256,1024} on RTX 3090.
Per-state cost drops 232× from bs=1 (80 µs) to bs=256 (0.34 µs) while
per-call latency stays near 90 µs through bs=256. Policy-call supply
from a real run is ~368 states per traversal and ~200k per iteration,
well above the bs=64–256 plateau, so batched inference is not
supply-limited. GPU forward is not the limiter once batching exists.
Verdict: Optimization Priorities #5 (batched traversal inference) is
worth pursuing. End-to-end gain will still be bounded by encoding and
worker-GPU coordination overhead.
- scripts/profile_gpu_forward.py: standalone profiling script
- docs/performance.md: new "GPU forward profiling for batched traversal"
experiment section with table, supply estimate, and verdict
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
coolguyandClaude Opus 4.8