openInvest Parameter Tuning Feasibility Diagnosis

Read-only assessment, no code changes. 2026-05-27. Supplementary Document (Coupling Analysis / Do-Not-Tune / Optuna Review / Type B Breakdown): 13-param-tuning-feasibility-addendum.md

---

1. Parameter Inventory

1.1 Regime Classification Thresholds (core/regime.py)

Parameter	File:Line	Current Value	Basis	Impact Scope
trend_ma_spread_pct	core/regime.py:37	3.0% (default)	MA20 vs MA120 deviation of 3% indicates a trend. No commit/ADR record, empirical initial value	Quant prompt strategy hints, Dreaming regime bucketing, verdict_review regime tagging
crash_atr_pct_min	core/regime.py:39	5.0% (default)	14-day ATR ≥5% considered abnormally high volatility. No clear basis	Crash liability exemption, Dreaming exemption, verdict_review macro_shock
crash_drawdown_30d_pct	core/regime.py:42	20.0%	30-day drop ≥20% + high volatility = sharp crash. No ADR	Same as above
crash_deep_drawdown_30d_pct	core/regime.py:46	30.0%	30-day drop ≥30% triggers crash independently (regardless of volatility). Added 2026-05-26, has commit record	Same as above
recovery_rebound_pct	core/regime.py:48	10.0%	Rebound ≥10% from 30-day low + still at low level = recovery. No ADR	Quant prompt "recovery" strategy hint
recovery_quantile_max	core/regime.py:50	0.50	Price 2y percentile <50% qualifies as "still at low level". No basis	Same as above
low_quantile_threshold	core/regime.py:52	0.20	2y percentile ≤20% = "low level". No basis	Deprecated (strategy hint no longer reads it after 2026-06 directional lock refactor; key retained for compatibility)
high_quantile_threshold	core/regime.py:54	0.80	2y percentile ≥80% = "high level". No basis	Deprecated (same as above)

Per-asset overrides (core/regime.py:69-85):

Asset	Overridden Parameter	Value	Basis
GC=F	trend_ma_spread_pct	5.0%	"Gold trends are more coherent, 3% is too sensitive" — commit comment, no quantitative validation
GC=F	crash_atr_pct_min	3.5%	"Gold's long-term ATR is ~1%, 5% would never trigger" — empirical judgment
NDQ.AX	trend_ma_spread_pct	4.0%	"ETF is smoother than single stocks" — empirical judgment
BTC-USD	trend_ma_spread_pct	8.0%	No basis
BTC-USD	crash_atr_pct_min	8.0%	"Crypto ATR is 3-5%, 8% prevents false positives" — empirical judgment
ETH-USD	trend_ma_spread_pct	8.0%	No basis
ETH-USD	crash_atr_pct_min	8.0%	No basis

1.2 Committee Sanity Check Thresholds (core/committee/cio_parse.py)

Parameter	File:Line	Current Value	Basis	Impact Scope
buy_confidence_overdrive	core/committee/cio_parse.py:84	0.95	"60-day sample signal is not significant, >0.95 historical hit rate is actually lower than 0.7-0.8" — commit comment, data-supported	BUY→ACCUMULATE downgrade trigger line
buy_confidence_downgrade_to	core/committee/cio_parse.py:85	0.6	"Expresses original LLM overconfidence" — design intent, no quantification	Confidence value after downgrade
alloc_cny_ceiling	core/committee/cio_parse.py:86	100,000	"Monthly investment ~¥10k, a single amount exceeding 10x is almost certainly an LLM error" — empirical	alloc clamp upper limit
worker_unavailable_confidence_floor	core/committee/cio_parse.py:87	0.4	"Worker incomplete → insufficient evidence" — design intent	Confidence cap when worker fails

1.3 LLM Call Parameters (core/committee/agent_io.py, max_debate_rounds in debate.py)

Parameter	File:Line	Current Value	Basis	Impact Scope
LLM_MAX_ATTEMPTS	core/committee/agent_io.py:23	3 (env)	"3 attempts complete within ~14s" — engineering experience	Retry count, affects latency and cost
LLM_BASE_DELAY	core/committee/agent_io.py:24	2.0s (env)	No basis	Retry interval exponential backoff base
LLM_MAX_DELAY	core/committee/agent_io.py:25	20.0s (env)	No basis	Maximum retry wait time
temperature	core/committee/agent_io.py:35	0.2	No basis. CIO uses 0.1	LLM output randomness
max_tool_iterations	core/committee/agent_io.py:75	4	No basis	Agent tool call round limit
max_debate_rounds	core/committee/debate.py:153	1 (default) / 4 (live)	Engineering experience	Cross-challenge rounds, affects LLM call volume

1.4 Dreaming System Parameters (jobs/dreaming.py)

Parameter	File:Line	Current Value	Basis	Impact Scope
LOOKBACK_DAYS	jobs/dreaming.py:64	90 (env)	"Only learn recent behavior" — design intent	Light Sleep ingestion window
WINDOWS	jobs/dreaming.py:65	[7, 30]	"Don't look at 1d (1d barely moves, it's a horizon artifact)" — commit comment	Lookback window days
MIN_RECALL	jobs/dreaming.py:66	3	No basis. LLM falsification prompt says "count≥6"	Minimum sample size threshold
MIN_SCORE	jobs/dreaming.py:67	0.8	"Same as OpenClaw" — clear source but no self-validation	Deep Sleep selection threshold
CAUTION_MIN_BASE_DOWN	jobs/dreaming.py:72	0.15	2026-05-27 ADR 008, has commit and design rationale	Caution mode downside base rate threshold
CAUTION_LIFT_FULL	jobs/dreaming.py:73	0.20	2026-05-27 ADR 008, has commit and design rationale	Caution lift full score threshold
VIX buckets (low/mid/high)	jobs/dreaming.py:227-231	<18 / 18-25 / ≥25	No basis. Standard empirical values	REM Sleep regime label
TNX buckets (low/mid/high)	jobs/dreaming.py:234-238	<4.0 / 4.0-4.5 / ≥4.5	No basis. Standard empirical values	REM Sleep regime label

1.5 Verdict Review Scoring Parameters (jobs/verdict_review.py)

Parameter	File:Line	Current Value	Basis	Impact Scope
K_FLAT	jobs/verdict_review.py:244	1.0	"Don't let the system self-learn the scoring ruler" — clear anti-reward-hacking design	HOLD "no movement" judgment multiplier
FLAT_CEILING_PCT	jobs/verdict_review.py:245	8.0%	"Prevent ruler from going out of control when ATR is abnormal" — design intent	HOLD judgment cap
DEFAULT_DAILY_VOL_PCT	jobs/verdict_review.py:246	2.0%	"Fallback when atr_pct fetch fails" — engineering fault tolerance	Fallback value when ATR fetch fails

1.6 Backtest Reward Function Weights (core/backtest_reward.py)

Parameter	File:Line	Current Value	Basis	Impact Scope
max_drawdown weight	core/backtest_reward.py:56	-0.5	No basis. Formula: reward = annualized - 0.5*max_dd + 0.5*alpha + 0.2*max(0,sharpe-1)	Optuna optimization objective
alpha_vs_yuebao weight	core/backtest_reward.py:57	0.5	No basis	Same as above
sharpe_bonus weight	core/backtest_reward.py:58	0.2	No basis	Same as above
sharpe_bonus_threshold	core/backtest_reward.py:58	1.0	No basis	Same as above
lam_mdd (v2 forward)	core/backtest_reward.py:91	1.0	No basis	per-sample reward MDD penalty
lam_return (v2 forward)	core/backtest_reward.py:92	0.05	"Small weight to prevent double-counting Sharpe" — design intent	per-sample reward return bonus

1.7 Verdict Oracle Accuracy Thresholds (core/backtest_reward.py:123-180)

Parameter	File:Line	Current Value	Basis	Impact Scope
BUY: +5% / -3%	backtest_reward.py:150-154	5.0 / -3.0	No basis	DSPy training oracle metric
ACCUMULATE: +3% / -3%	backtest_reward.py:156-160	3.0 / -3.0	No basis	Same as above
HOLD: ±3% / ±8%	backtest_reward.py:162-166	3.0 / 8.0	No basis	Same as above
TRIM: -3% / +5%	backtest_reward.py:168-172	-3.0 / 5.0	No basis	Same as above
SELL: -5% / +3%	backtest_reward.py:174-178	-5.0 / 3.0	No basis	Same as above

1.8 Market Metrics Parameters (utils/market_metrics.py)

Parameter	File:Line	Current Value	Basis	Impact Scope
LOOKBACK_30D	market_metrics.py:28	30	Trading day standard	return_30d, rebound_off_30d_low calculation
RVOL_WINDOW	market_metrics.py:30	20	No basis	Relative volume average window
RSI period	market_metrics.py:43	14	Wilder 1978 original standard	RSI calculation
ATR period	market_metrics.py:75	14	Standard	ATR% calculation, regime crash determination
MA20 / MA120 / MA250	market_metrics.py:250-252	20 / 120 / 250	Standard technical analysis periods	Regime trend determination, quantile calculation

1.9 Strategy Metrics Parameters (core/strategy_metrics.py)

Parameter	File:Line	Current Value	Basis	Impact Scope
risk_free_annual	strategy_metrics.py:72	0.013 (1.3%)	"Current Yu'ebao rate" — time-sensitive	Sharpe / Sortino calculation
trading_days_per_year	strategy_metrics.py:73	252	Standard	Annualized return / volatility

1.10 Trading Friction Parameters (utils/exchange_fee.py)

Parameter	File:Line	Current Value	Basis	Impact Scope
cn_cable_fee	exchange_fee.py:66	¥150	Bank wire transfer fee. Measured value	Currency exchange friction calculation
cn_commission_rate	exchange_fee.py:67	0.1%	Bank commission rate. Measured value	Same as above
cn_commission_min	exchange_fee.py:68	¥50	Bank minimum commission. Measured value	Same as above
cn_commission_max	exchange_fee.py:69	¥260	Bank maximum commission. Measured value	Same as above
au_inward_fee	exchange_fee.py:70	¥15	Australian inward credit fee. Measured value	Same as above
CommSec tier ladder	exchange_fee.py:71-74	$5/$10/$19.95/0.12%	CommSec official fee schedule	Trading commission
GOLD_OZ_PER_GRAM	utils/gold_price.py:23	31.1035	Physical constant	Gold price unit conversion

1.11 Other Scattered Parameters

Parameter	File:Line	Current Value	Basis	Impact Scope
max_single_invest_cny default	core/portfolio_manager.py:188	¥10,000	No basis	Single investment upper limit fallback
max_workers (committee)	core/runner/session.py	4	"Prevent LLM API rate limiting" — engineering experience	Concurrency level
step_days (walk-forward)	experiments/train_config.py:121	7	No basis	Optuna training decision frequency
Dreaming LLM falsification temperature	jobs/dreaming.py:440	0.1	"Falsification needs stability" — design intent	LLM falsification randomness
pnl_snapshot rendering constants	jobs/pnl_snapshot.py:56-66	W=800, LINE_H=240 etc.	Visual experience	Chart rendering (non-decision)

---

2. Parameter Classification

Type A: Deterministic, Cleanly Calibratable (Pure Arithmetic, No LLM Needed for Fit)

Total 28. These parameters reside in pure mathematical functions and can be directly optimized using historical market data.

#	Parameter	Module	Calibration Method
1	trend_ma_spread_pct (default + 4 per-asset)	regime	Grid search on historical regime accuracy
2	crash_atr_pct_min (default + 3 per-asset)	regime	Grid search on crash detection precision/recall
3	crash_drawdown_30d_pct	regime	Historical drawdown distribution analysis
4	crash_deep_drawdown_30d_pct	regime	Same as above
5	recovery_rebound_pct	regime	Historical rebound distribution
6	recovery_quantile_max	regime	Quantile distribution analysis
7	low_quantile_threshold / high_quantile_threshold	regime	2y price quantile distribution
8	K_FLAT	verdict_review	Grid search on HOLD accuracy vs flat band width
9	FLAT_CEILING_PCT	verdict_review	ATR distribution tail analysis
10	DEFAULT_DAILY_VOL_PCT	verdict_review	ATR distribution fallback position
11	VIX buckets 18/25	dreaming	VIX historical distribution quantiles
12	TNX buckets 4.0/4.5	dreaming	TNX historical distribution quantiles
13	LOOKBACK_DAYS	dreaming	Signal decay curve
14	WINDOWS [7, 30]	dreaming	Forward return convergence window
15	MIN_RECALL	dreaming	Sample size vs stability tradeoff
16	Reward function 4 weights	backtest_reward	Optuna already ran 30 trials
17	Oracle accuracy 10 thresholds	backtest_reward	Historical return distribution per verdict
18	risk_free_annual	strategy_metrics	Yu'ebao real-time interest rate

Type B: LLM-Dependent, Contamination Limited

Total 9. The "optimal value" for these parameters depends on LLM output characteristics (confidence distribution, verdict distribution) and requires LLM involvement for evaluation.

#	Parameter	Module	Contamination Limitation
1	buy_confidence_overdrive (0.95)	committee	Requires LLM confidence distribution + ex-post hit rate. But can be reverse-engineered from verdict_review.jsonl historical data
2	buy_confidence_downgrade_to (0.6)	committee	Same as above
3	alloc_cny_ceiling (100k)	committee	Requires LLM alloc output distribution. Can be statistically analyzed from historical transcripts
4	worker_unavailable_confidence_floor (0.4)	committee	Requires simulating worker failure scenarios
5	MIN_SCORE (0.8)	dreaming	Requires validating insight quality before/after LLM falsification
6	temperature (0.2)	committee	Requires multiple LLM sampling for consistency comparison
7	max_tool_iterations (4)	committee	Requires observing LLM tool call behavior
8	max_debate_rounds (1/4)	committee	Requires LLM debate output quality assessment
9	Dreaming score weights (0.7/0.3)	dreaming	Hit rate and sample weights in the score formula

---

3. Decoupling Difficulty

3.1 Current Sweep Workflow Infrastructure

Existing:

• scripts/backtest_committee.py — Runs historical committee day-by-day, blind to the future ✅
• scripts/backtest_runner.py — Workspace isolation, does not pollute real memory/ ✅
• scripts/run_walk_forward.py — Walk-forward + paper trading simulator ✅
• experiments/train_config.py — Optuna parameter space definition ✅
• scripts/rl_train.py — Optuna training entry point ✅
• scripts/holdout_validate.py — Hold-out validation ✅
• core/backtest_reward.py — Reward function ✅
• core/strategy_metrics.py — Complete strategy metrics ✅
• jobs/verdict_review.py — Verdict post-mortem analysis ✅

Missing:

• Regime thresholds have no external injection point — THRESHOLDS and ASSET_OVERRIDES are module-level dict literals; backtest can only monkey-patch, cannot pass via env/CLI
• Dreaming parameters have no external injection point — MIN_SCORE, MIN_RECALL, WINDOWS, LOOKBACK_DAYS; only LOOKBACK_DAYS has an env override
• Reward function weights have no external injection point — The 4 weights in compute_strategy_reward are hardcoded
• verdict_review K_FLAT/FLAT_CEILING_PCT have no external injection point — Hardcoded constants
• backtest_committee's run_committee call is not parameterized — max_debate_rounds is passed via env INVEST_MAX_DEBATE_ROUNDS, but other parameters are not

3.2 Rough Effort Estimate

Change	Effort	Description
Parameterize Regime THRESHOLDS	2-3 hours	Add env override or CLI arg → _per_asset_thresholds() reads external values → backtest automatically applies. Changes concentrated in core/regime.py
Parameterize Dreaming parameters	1-2 hours	Add env overrides for MIN_SCORE/MIN_RECALL/WINDOWS, following the same pattern as LOOKBACK_DAYS
Parameterize Reward function weights	1 hour	Add kwargs to compute_strategy_reward or read from TrainConfig.reward
Parameterize verdict_review parameters	1 hour	Add env overrides for K_FLAT/FLAT_CEILING_PCT
Sweep runner (batch grid search)	4-6 hours	Extend based on existing rl_train.py + backtest_runner.py, support grid/random search for Type A parameters + result aggregation
Complete pipeline (param sweep → backtest → metrics → comparison)	1-2 days	Chain sweep runner + walk-forward + strategy_metrics + report generation
Offline evaluation pipeline for Type B parameters	2-3 days	Requires extracting LLM output distribution + ex-post hit rate from verdict_review.jsonl, building a "quasi A/B test" framework

Total: Making the sweep workflow usable takes approximately 1-2 days; making both Type A and Type B tunable takes approximately 3-5 days.

---

4. Data Availability

4.1 Data Available for Type A Parameters

Parameter	post-cutoff	Full 2018-2023	Either	Description
trend_ma_spread_pct	✅	✅	✅	Pure market arithmetic. post-cutoff has ~90 days of NDQ.AX + GC=F; full has 2y DB data
crash_atr_pct_min	✅	✅	✅	ATR is purely historical calculation. But crash events are rare (2020 COVID, 2022 rate hikes), full data is better
crash_drawdown_30d_pct	⚠️	✅	⚠️	post-cutoff may have no 30-day 20% drop samples. Full data has 2020.03
crash_deep_drawdown_30d_pct	❌	✅	❌	post-cutoff has no 30% drop. Requires full data
recovery_rebound_pct	⚠️	✅	⚠️	post-cutoff has 2026.04 rebound samples, but few
recovery_quantile_max	✅	✅	✅	Quantile calculation only needs sufficiently long history
low/high_quantile_threshold	✅	✅	✅	Same as above
K_FLAT	✅	✅	✅	HOLD hit rate varies with flat band width, usable with any window containing verdicts
FLAT_CEILING_PCT	✅	✅	✅	ATR tail analysis
VIX/TNX buckets	✅	✅	✅	Macro data is sufficient
LOOKBACK_DAYS	✅	❌	⚠️	post-cutoff has 90 days of verdicts; full 2018-2023 has no verdict_review data
WINDOWS [7,30]	✅	❌	⚠️	Same as above. But full data can be evaluated after generating verdicts via backtest
MIN_RECALL	✅	❌	⚠️	Same as above
Reward weights	✅	✅	✅	Pure arithmetic, only needs P&L curve
Oracle accuracy thresholds	✅	✅	✅	Pure arithmetic, only needs return distribution
risk_free_annual	✅	✅	✅	Yu'ebao interest rate changes slowly

4.2 Data Available for Type B Parameters

Parameter	Contamination Limitation	Workaround
buy_confidence_overdrive	Requires LLM confidence output distribution	Can be reverse-engineered from verdict_review.jsonl historical data — jsonl already records each verdict's confidence + ex-post hit rate. Do confidence bucketing → hit rate curve → find inflection point. No need to re-run LLM
buy_confidence_downgrade_to	Same as above	Same as above. Find the inflection point where "BUY confidence above which hit rate decreases"
alloc_cny_ceiling	Requires LLM alloc output distribution	Can be statistically analyzed from backtest transcripts — .backtest/ directory has historical CIO memos, regex extract alloc_cny distribution
worker_unavailable_confidence_floor	Requires simulating worker failure	Can filter records containing [WORKER_UNAVAILABLE] from verdict_review.jsonl — few in number but exist
MIN_SCORE	Requires evaluating insight quality	Can do retrospective analysis — look at the difference in ex-post hit rate for insights with scores in the 0.7-0.9 range
temperature	Requires multiple sampling	Most contaminated — cannot be calibrated with historical data, requires A/B testing
max_tool_iterations	Requires observing tool call behavior	Can be statistically analyzed from llm_telemetry.jsonl — already records iteration for each tool call
max_debate_rounds	Requires comparing multi-round debate quality	Can be compared via backtest — has INVEST_MAX_DEBATE_ROUNDS env, run 1 vs 3 vs 4 round backtests and compare
Dreaming score weights	Requires insight quality evaluation	Can do retrospective analysis — look at ex-post validation rate of insights under different weights

---

5. Priority Recommendations

Based on a three-dimensional score of "Honestly Calibratable + Significant Impact + Low Decoupling Cost":

Priority 1: trend_ma_spread_pct (default + per-asset)

• Why it's worth tuning: This is the most upstream parameter for regime classification, directly affecting the three-way split of uptrend/downtrend/range_bound. Quant prompt strategy hints, Dreaming regime bucketing, and verdict_review regime tagging all depend on it. The current default of 3% + per-asset overrides (GC=F 5%, NDQ.AX 4%) are entirely empirical, with no quantitative validation.
• What data to use: Full DB history (2018-2023 + post-cutoff), calculate the historical distribution of MA20/MA120 spread for each asset.
• What validation method to use: Grid search [2%, 8%] step 0.5% → For each threshold, calculate the "separation of ex-post 30d return bucketed by regime" (average gain after uptrend vs average loss after downtrend). The threshold with the highest separation is optimal.
• Expected output: Optimal MA spread threshold for each asset + quantified improvement in separation.

Priority 2: K_FLAT (HOLD "no movement" judgment multiplier)

• Why it's worth tuning: HOLD is the committee's most common verdict (>70%), and K_FLAT directly determines the criterion for "HOLD hit". Too loose (large K_FLAT) → inflated HOLD hit rate → Dreaming learns bad HOLDs as good patterns; too tight → all HOLDs judged as "missed opportunity" → excessive conservatism. Current K_FLAT=1.0 is unvalidated.
• What data to use: verdict_review.jsonl (already has ex-post return + ATR for each verdict), full data available.
• What validation method to use: Grid search [0.5, 2.0] step 0.25 → Recalculate HOLD hit rate for each K_FLAT → Plot "HOLD hit rate vs K_FLAT" curve → Find the inflection point (where the curve slope changes abruptly).
• Expected output: Optimal K_FLAT value + quantitative diagnosis of whether the "HOLD judgment standard is too loose/tight".

Priority 3: crash_atr_pct_min (default + per-asset)

• Why it's worth tuning: Crash determination is the system's liability exemption mechanism — verdicts during crashes are not fed into Dreaming learning and do not count towards hit rates. If the crash threshold is too loose (frequent false crash positives), many samples are exempted and Dreaming learns nothing; if too tight, the system continues learning during real crashes as if it were normal market conditions. The current default of 5% with a 3.5% override for gold is empirical.
• What data to use: Full DB history ATR distribution + historical crash event annotations (2020.03 COVID, 2022 rate hikes, 2026.04 tariffs).
• What validation method to use: For each asset, plot the ATR historical distribution → annotate known crash events → find the "minimum ATR among crash events" as the lower bound → select the "minimum threshold that does not miss any known crash".
• Expected output: Crash ATR threshold for each asset + historical crash event recall rate.

Priority 4: buy_confidence_overdrive (BUY→ACCUMULATE downgrade line)

• Why it's worth tuning: This is the only defense against LLM overconfidence. The current value of 0.95 means "only BUY with confidence ≥ 0.95 is downgraded" — but the commit comment says "BUY with >0.95 confidence has a lower historical hit rate than 0.7-0.8", suggesting the threshold should probably be lower.
• What data to use: verdict_review.jsonl (already records each verdict's confidence + ex-post hit rate).
• What validation method to use: Confidence bucketing [0.5, 0.6, ..., 0.95, 1.0] → BUY hit rate per bucket → Plot "hit rate vs confidence" curve → Find the inflection point where "hit rate starts to decline". The inflection point is the optimal threshold.
• Expected output: Optimal downgrade threshold for BUY confidence + quantitative profile of "LLM overconfidence".

Priority 5: Oracle accuracy thresholds (BUY +5%/-3% etc., 10 values)

• Why it's worth tuning: This set of thresholds defines "what counts as a correct verdict direction" and is the label standard for DSPy training. If the BUY +5% threshold is too high (a 5% gain in 30 days is uncommon in actual markets), many correct BUYs are labeled as "neutral" → training signal is diluted.
• What data to use: Forward return distribution (7d/30d) from verdict_review.jsonl.
• What validation method to use: For each verdict, plot the forward return distribution → redefine thresholds using quantiles (e.g., BUY +1 threshold = 60th percentile of return distribution, -1 threshold = 20th percentile).
• Expected output: Oracle thresholds based on actual return distribution + quantified improvement in training signal density.

---

Appendix: Parameters Not Recommended for Priority Tuning

Parameter	Reason
temperature	Most contaminated, requires A/B testing, cannot be done retrospectively
max_debate_rounds	Optuna 30 trials already confirmed it's a placebo (confirmed_effective=False)
cio_confidence_cap	Optuna 30 trials already confirmed it's a placebo
regime_uptrend / regime_atr	Optuna 30 trials already confirmed it's a placebo (weak)
Trading friction parameters	Measured values, determined by bank/broker fee schedules, not calibration targets
risk_free_annual	Yu'ebao real-time interest rate, just look it up
Rendering constants (pnl_snapshot)	Non-decision parameters

---

Appendix: Key File Index

File	Content
core/regime.py	Regime classification Single Source of Truth (THRESHOLDS + ASSET_OVERRIDES)
core/committee/cio_parse.py:84-87	Committee sanity check thresholds (THRESHOLDS build)
core/backtest_reward.py	Reward function + Oracle accuracy
core/strategy_metrics.py	Strategy metric calculation
jobs/dreaming.py:60-73	Dreaming system parameters
jobs/verdict_review.py:244-246	HOLD judgment parameters
utils/market_metrics.py	Technical indicator calculation parameters
utils/exchange_fee.py	Trading friction parameters
experiments/train_config.py	Optuna parameter space (5 hyperparameters)
scripts/backtest_committee.py	Historical backtest entry point
scripts/rl_train.py	Optuna training entry point