Spaces:

InosLihka
/

rhythm_env

Sleeping

App Files Files Community

rhythm_env / docs /architecture.md

InosLihka

Algorithm Distillation: grader v2 with belief_accuracy + SFT pipeline

ece0bbe 12 days ago

preview code

raw

history blame contribute delete

40.4 kB

	# RhythmEnv — Architecture & Training Flow

	Visual deep dive into how the env is structured and how the LLM agent
	learns from it. All examples use concrete values (seed=42, sampled
	profile, real numbers from the reward calculation).

	---

	## 1. System: three-layer separation

	```
	┌────────────────────────────────────────────┐
	│ AGENT (Qwen 2.5-3B + LoRA r=8, 4-bit) │
	│ │
	│ Input: prompt (state + history) │
	│ Output: "3 7 5 DEEP_WORK" │
	│ ↑ ↑ ↑ ↑ │
	│ social\|morn\|work action │
	│ belief\|pref\|pref │
	└─────────────────┬──────────────────────────┘
	│
	│ N=8 completions per prompt
	│ (sampling temp=1.5)
	▼
	┌───────────────────────────────────────────────┐
	│ ORCHESTRATION (TRL GRPOTrainer + Unsloth) │
	│ │
	│ • Picks 1 prompt from dataset (~3000 rows) │
	│ • Generates 8 completions │
	│ • Calls 4 reward functions in parallel │
	│ • Computes group-relative advantages │
	│ • Backprop on LoRA weights only (~30M) │
	│ • KL constraint to base Qwen (β=0.04) │
	└─────────────────┬─────────────────────────────┘
	│
	│ env.reset(seed) → step(action)
	│ (replay-based reward)
	▼
	┌───────────────────────────────────────────────┐
	│ ENVIRONMENT (RhythmEnvironment / FastAPI) │
	│ │
	│ reset(seed=42) → samples profile │
	│ step(action) → updates 5 meters, │
	│ returns observation │
	│ + per-step reward │
	│ │
	│ Lives at: huggingface.co/spaces/... │
	└───────────────────────────────────────────────┘
	```

	The agent never imports env code. They communicate via the OpenEnv
	HTTP/WebSocket contract: `POST /reset`, `POST /step`, `GET /state`.

	---

	## 2. One episode = one week (28 steps)

	```
	ONE EPISODE = 1 WEEK = 28 STEPS

	Day 1 (Mon) Day 2 (Tue) ... Day 7 (Sun)
	┌─────┬─────┬─────┬─────┐ ┌─────┬─────┬─────┬─────┐ ┌─────┬─────┬─────┬─────┐
	│ M │ A │ E │ N │ │ M │ A │ E │ N │ .. │ M │ A │ E │ N │
	│ 0 │ 1 │ 2 │ 3 │ │ 4 │ 5 │ 6 │ 7 │ │ 24 │ 25 │ 26 │ 27 │
	└─────┴─────┴─────┴─────┘ └─────┴─────┴─────┴─────┘ └─────┴─────┴─────┴─────┘
	↑ ↑ ↑
	reset() random event roll (8% per step) done=True
	meters: V=0.7 meditate? deep_work? grader runs
	C=0.7 P=0 S=0.7 sleep? socialize? final_score [0,1]
	Cn=0.5

	At each step the agent picks 1 of 10 actions:

	┌────────────────┬─────────────────┬───────────────┬────────────────┐
	│ PRODUCTIVITY │ RECOVERY │ SOCIAL │ LEISURE │
	├────────────────┼─────────────────┼───────────────┼────────────────┤
	│ DEEP_WORK │ SLEEP │ FAMILY_TIME │ ME_TIME │
	│ ADMIN_WORK │ EXERCISE │ SOCIALIZE │ BINGE_WATCH │
	│ LEARN │ MEDITATE │ │ │
	└────────────────┴─────────────────┴───────────────┴────────────────┘

	Time-of-day multipliers (apply to action effects):
	Morning (M): cognition gains × 1.2 vitality drain × 0.8
	Afternoon (A): cognition gains × 1.0 vitality drain × 1.0
	Evening (E): cognition gains × 0.8 vitality drain × 1.1
	Night (N): cognition gains × 0.6 vitality drain × 1.3
	(sleep BYPASSES these multipliers)

	Critical thresholds:
	Any meter < 0.1 at end of step → -0.30 reward penalty (one per meter)
	Connection decays passively every step (-0.01 to -0.02 per profile)
	```

	---

	## 3. State tree: what the env tracks (and what's hidden)

	```
	RhythmEnvironment instance state
	│
	├── _profile ◄──── HIDDEN from agent ─────────┐
	│ │ │
	│ ├── name: "sampled_42" │
	│ ├── reward_weights: │
	│ │ {V: 0.05, C: 0.05, P: 0.30, │
	│ │ S: 0.20, Cn: 0.40} │
	│ │ │ Used INTERNALLY
	│ ├── social_vitality_multiplier: 1.8 │ to compute reward
	│ ├── morning_cognition_bonus: 1.5 │ and modify action
	│ ├── evening_night_cognition_bonus: None │ effects.
	│ ├── morning_penalty: None │ Belief reward
	│ ├── binge_shame: False │ rewards INFERRING
	│ ├── progress_serenity_bonus: 0.04 │ this profile.
	│ ├── idle_serenity_decay: 0.02 │
	│ ├── vitality_decay_rate: 0.01 │
	│ ├── stress_tolerance: 0.22 │
	│ ├── event_impact_multiplier: 0.7 │
	│ ├── connection_decay_rate: 0.012 │
	│ ├── solo_serenity_bonus: 0.05 │
	│ ├── social_connection_multiplier: 1.4 │
	│ ├── social_serenity_bonus: 0.03 │
	│ ├── work_vitality_recovery: 0.03 │
	│ └── (14 hidden parameters total) │
	│ ──────────────────────────────────────────┘
	│
	│ Reduces to a 3-D BELIEF VECTOR (the inference target):
	│ profile_to_belief_vector(profile) → [0.30, 0.70, 0.50]
	│ │ │ │
	│ │ │ work_pref
	│ │ morning_pref
	│ social_pref
	│
	├── meters ◄────── visible to agent ────────────────────────────┐
	│ ├── _vitality: 0.62 (range 0–1) │
	│ ├── _cognition: 0.51 │
	│ ├── _progress: 0.24 │ Agent
	│ ├── _serenity: 0.71 │ observes
	│ └── _connection: 0.38 │ in prompt
	│ ────────────────────────────────────────────────────────────┘
	│
	├── _timestep: 5
	│
	├── _step_history: list[StepRecord] ◄── visible to agent ──────┐
	│ ├── (last 7 steps) │
	│ │ step 0: deep_work → +0.42 │
	│ │ deltas: V-0.10 C-0.12 P+0.18 S-0.05 Cn+0.00 │
	│ │ anomalies: V+0.00 C+0.00 P+0.06 S+0.00 Cn+0.00 ◄─ profile fingerprint
	│ │ step 1: sleep → +0.18 │
	│ │ deltas: V+0.20 C+0.10 P+0.00 S+0.05 Cn+0.00 │
	│ │ anomalies: V+0.00 C+0.00 P+0.00 S+0.00 Cn+0.00 │
	│ │ ... │
	│ ─────────────────────────────────────────────────────────────┘
	│
	└── _step_rewards: [+0.42, +0.18, +0.31, +0.55, ...]
	└── used by grader to compute adaptation_score (late-half - early-half)
	```

	---

	## 4. What the agent sees (concrete prompt)

	```
	┌───────────────────────────────────── SYSTEM PROMPT ─────────────────────────────────┐
	│ You are a life-management agent helping a person whose preferences are HIDDEN. │
	│ Each step, output ONE LINE in this exact format: │
	│ S M W ACTION_NAME │
	│ S = social pref (0=hates, 9=loves), M = morning, W = work │
	│ ACTION_NAME ∈ {DEEP_WORK, ADMIN_WORK, LEARN, SLEEP, EXERCISE, MEDITATE, │
	│ FAMILY_TIME, SOCIALIZE, ME_TIME, BINGE_WATCH} │
	│ Example: 3 8 7 DEEP_WORK │
	│ Tactics: probe early, exploit late; don't repeat actions; ... │
	└──────────────────────────────────────────────────────────────────────────────────────┘
	┌──────────────────────────────────── USER PROMPT ────────────────────────────────────┐
	│ Step: 5/28 (Tuesday Afternoon) │
	│ Remaining steps: 22 │
	│ │
	│ Meters: │
	│ Vitality: 0.62 │
	│ Cognition: 0.51 │
	│ Progress: 0.24 │
	│ Serenity: 0.71 │
	│ Connection: 0.38 │
	│ │
	│ Recent history (anom = how this person deviated from neutral baseline): │
	│ step 0: deep_work -> reward +0.42 (V-0.10 C-0.12 P+0.18 S-0.05 Cn+0.00) │
	│ [anom V+0.00 C+0.00 P+0.06 S+0.00 Cn+0.00] │
	│ step 1: sleep -> reward +0.18 (V+0.20 C+0.10 P+0.00 S+0.05 Cn+0.00) │
	│ [anom V+0.00 C+0.00 P+0.00 S+0.00 Cn+0.00] │
	│ step 2: socialize -> reward -0.05 (V-0.11 C-0.03 P+0.00 S+0.04 Cn+0.17) │
	│ [anom V-0.05 C+0.00 P+0.00 S+0.00 Cn+0.05] ← strong profile signal │
	│ step 3: meditate -> reward +0.30 (V+0.03 C+0.08 P+0.00 S+0.20 Cn+0.00) │
	│ [anom V+0.00 C+0.00 P+0.00 S+0.05 Cn+0.00] ← solo bonus visible │
	│ step 4: deep_work -> reward +0.18 (V-0.06 C-0.06 P+0.18 S+0.04 Cn+0.00) │
	│ [anom V+0.00 C+0.00 P+0.00 S+0.00 Cn+0.00] │
	│ │
	│ Output your belief, then your action (format: S M W ACTION_NAME): │
	└──────────────────────────────────────────────────────────────────────────────────────┘

	LLM completion (1 of 8 sampled at temp=1.5):
	┌──────────────────────┐
	│ "3 7 5 DEEP_WORK" │
	└──────────┬───────────┘
	▼
	Parsed: belief = [0.33, 0.78, 0.56]
	action = DEEP_WORK
	```

	---

	## 5. The reward stack (4 layers, with concrete values for the example above)

	```
	┌─────────────────────────────────────┐
	│ Completion: "3 7 5 DEEP_WORK" │
	│ for prompt above (seed=42, step 5) │
	└─────────────────┬───────────────────┘
	│
	┌──────────────────────────────┼──────────────────────────────┐
	│ │ │
	▼ ▼ ▼
	┌──────────────┐ ┌──────────────────┐ ┌────────────────────┐
	│ format_valid │ │ action_legal │ │ env_reward │
	├──────────────┤ ├──────────────────┤ ├────────────────────┤
	│ parses? │ │ action ∈ 10 ? │ │ env.replay(seed=42,│
	│ has belief? │ │ │ │ step5, action) │
	│ │ │ DEEP_WORK ✓ │ │ │
	│ ✓ +1.0 │ │ reward 0.0 │ │ deltas: │
	│ × wt 0.05 │ │ × wt 0.05 │ │ V-0.10 C-0.12 │
	│ │ │ │ │ P+0.18 S-0.05 │
	│ → +0.05 │ │ → 0.00 │ │ Cn+0.00 │
	└──────────────┘ └──────────────────┘ │ │
	│ profile_reward = │
	│ sum(deltas × │
	│ prof_weights) │
	│ × 15 │
	│ = (-0.10×0.05 │
	│ + -0.12×0.05 │
	│ + 0.18×0.30 │
	│ + -0.05×0.20 │
	│ + 0×0.40) × 15 │
	│ = +0.65 │
	│ │
	│ + grader_bias: │
	│ 0.5×0.18+0.4×0 │
	│ = +0.09 │
	│ + new-action: +0.07│
	│ + b-act coupling: │
	│ work=0.56 (mid), │
	│ no bonus = 0 │
	│ - cycle pen: 0 │
	│ - rep pen: 0 │
	│ │
	│ env_reward = +0.81 │
	│ × wt 1.5 = +1.21 │
	└────────────────────┘
	│
	│
	▼
	┌─────────────────────────────┐
	│ belief_accuracy │
	├─────────────────────────────┤
	│ true belief (sampled_42) │
	│ = [0.30, 0.70, 0.50] │
	│ agent belief │
	│ = [0.33, 0.78, 0.56] │
	│ │
	│ MAE = (0.03 + 0.08 + 0.06) │
	│ / 3 = 0.057 │
	│ similarity = 0.943 │
	│ │
	│ baseline (constant 0.5): │
	│ MAE = (0.20+0.20+0.00)/3 │
	│ = 0.133 │
	│ similarity = 0.867 │
	│ │
	│ reward = 0.943 - 0.867 │
	│ = +0.076 │
	│ × wt 3.0 = +0.23 │
	└─────────────────────────────┘

	Σ TOTAL REWARD (this completion)
	= 0.05 + 0.00 + 1.21 + 0.23
	= +1.49
	```

	---

	## 6. GRPO update step (8 completions → 1 gradient update)

	```
	ONE TRAINING STEP (one row from the dataset → one gradient update)

	┌─ pick prompt ────────────────────────────────────────────────────────┐
	│ prompt#1247 from dataset: │
	│ state: step 5, seed=42, history=[deep_work, sleep, ...] │
	└─────────────┬────────────────────────────────────────────────────────┘
	│
	▼ generate 8 completions @ temp=1.5
	┌──────────────────────────────────────────────────────────────────────┐
	│ c1: "3 7 5 DEEP_WORK" → reward +1.49 │
	│ c2: "5 5 5 SLEEP" → reward -0.21 (constant belief: -0.03 │
	│ + sleep replay: -0.18) │
	│ c3: "3 6 4 ADMIN_WORK" → reward +1.10 │
	│ c4: "4 7 6 LEARN" → reward +1.32 │
	│ c5: "2 8 3 MEDITATE" → reward +0.45 (rep pen: -0.10 since │
	│ meditate at step 3 too) │
	│ c6: "3 7 7 DEEP_WORK" → reward +1.55 (belief slightly better) │
	│ c7: "5 5 5 EXERCISE" → reward -0.15 │
	│ c8: "4 6 5 FAMILY_TIME" → reward +0.92 │
	│ │
	│ group_mean = +0.81 │
	└─────────────┬────────────────────────────────────────────────────────┘
	│
	▼ advantages = reward - group_mean
	┌──────────────────────────────────────────────────────────────────────┐
	│ ADVANTAGE (this is what GRPO actually backprops on) │
	│ c1: +0.68 ← strongly preferred │
	│ c2: -1.02 ← strongly discouraged (constant belief) │
	│ c3: +0.29 │
	│ c4: +0.51 │
	│ c5: -0.36 │
	│ c6: +0.74 ← most preferred │
	│ c7: -0.96 │
	│ c8: +0.11 │
	│ │
	│ KEY INSIGHT: only the SPREAD matters. If all 8 had the │
	│ same reward, advantages would all be 0 → no gradient. │
	│ This is why iter 1 mode-collapsed: format_valid +1.0 for │
	│ every completion meant zero variance from that layer. │
	└─────────────┬────────────────────────────────────────────────────────┘
	│
	▼ policy loss = -E[ adv × log_prob(completion) ]
	│ + β × KL(policy \|\| base_qwen)
	│
	▼ backprop (only LoRA weights, ~30M params)
	┌──────────────────────────────────────────────────────────────────────┐
	│ Model nudge: │
	│ push: "3 7 5 DEEP_WORK"-like outputs UP │
	│ push: "3 7 7 DEEP_WORK"-like outputs UP │
	│ pull: "5 5 5 *"-like outputs DOWN │
	│ │
	│ KL constraint: prevents the policy from diverging too │
	│ far from base Qwen (avoids gibberish drift). │
	└──────────────────────────────────────────────────────────────────────┘

	Repeat 800-2000 times. Each step ~3-5 sec on A100.
	```

	---

	## 7. The dataset is just starting positions (not labels)

	```
	DATASET (~3000 rows, generated ONCE before training)

	For 200-300 episodes:
	env.reset(seed=N)
	for step in range(28):
	record {
	prompt: [system, user_for_this_state],
	seed: N, ◄─── replay metadata
	step_index: <current step>,
	action_history: <actions taken so far>,
	profile_mode: "continuous",
	}
	env.step(rollout_policy(obs)) ◄─ rollout=heuristic OR random
	(only matters for STATE diversity,
	the agent's training generations
	REPLACE these actions)

	┌──────────────────────────────────────────────────────────────────┐
	│ A row from the dataset: │
	│ { │
	│ prompt: [...full state at step 5 of episode seed=42...] │
	│ seed: 42, │
	│ step_index: 5, │
	│ action_history: ["deep_work", "sleep", "socialize", │
	│ "meditate", "deep_work"], │
	│ profile_mode: "continuous" │
	│ } │
	│ │
	│ NOTE: NO "correct action" or "label" anywhere. │
	│ The reward function reconstructs the env from this metadata │
	│ and scores whatever action the LLM picks. │
	└──────────────────────────────────────────────────────────────────┘

	This is fundamentally different from supervised learning:
	- Supervised: (input, target_output) — model learns to mimic target
	- GRPO: (prompt, replay_metadata) — model learns to maximize reward
	```

	---

	## 8. Final episode grader (only fires at step 28)

	```
	_grade_episode() — runs at done=True

	┌──────────────┐
	│ final_score │
	│ ∈ [0, 1] │
	└──────┬───────┘
	┌──────────────┬─────────────┼──────────────┬──────────────┬──────────────┐
	│ │ │ │ │ │
	▼ ▼ ▼ ▼ ▼ ▼
	┌────────────┐ ┌────────────┐ ┌──────────┐ ┌────────────┐ ┌────────────┐ ┌────────────┐
	│ crash_free │ │ progress │ │ connection│ │ adaptation │ │ efficiency │ │ belief │
	│ × 0.15 │ │ × 0.20 │ │ × 0.10 │ │ × 0.25 │ │ × 0.10 │ │ accuracy │
	│ │ │ │ │ │ │ │ │ │ │ × 0.20 │
	├────────────┤ ├────────────┤ ├──────────┤ ├────────────┤ ├────────────┤ ├────────────┤
	│ 1 - crashes│ │ final P │ │ final Cn │ │ late-half │ │ avg_reward │ │ 1 - MAE │
	│ /total_ck │ │ value │ │ value │ │ mean reward│ │ normalized │ │ vs true │
	│ │ │ │ │ │ │ - early │ │ to [0,1] │ │ profile │
	│ e.g. 0.95 │ │ e.g. 0.42 │ │ e.g. 0.51│ │ e.g. +0.18 │ │ e.g. 0.55 │ │ e.g. 0.80 │
	│ ×0.15=0.14 │ │ ×0.20=0.084│ │ ×0.10=0.05│ │ ×0.25=0.045│ │ ×0.10=0.055│ │ ×0.20=0.16 │
	└────────────┘ └────────────┘ └──────────┘ └────────────┘ └────────────┘ └────────────┘

	Σ = 0.14 + 0.084 + 0.05 + 0.045 + 0.055 + 0.16
	= 0.534 ← final_score (with inference)

	Heuristic / random baselines never call env.record_belief(), so the belief
	component scores 0 for them — by design: the meta-RL skill is INFERENCE,
	and only agents that actually try get credit on this axis.

	Plus a sparse terminal reward (added to step 27's per-step reward):
	terminal_bonus = (final_score - 0.5) × 5 → e.g. (0.534 - 0.5) × 5 = +0.17

	This means: at step 27, agent gets last per-step reward + bonus from grader.
	This is the only direct gradient signal pointing at the actual episode quality.
	```

	---

	## 9. Three eval conditions (post-training)

	```
	inference_eval.py runs ALL THREE

	┌────────────────────────────┬────────────────────────────┬────────────────────────────┐
	│ discrete-3-profiles │ continuous-in-distribution│ continuous-OOD │
	│ (3 reference profiles) │ (was the agent able to │ (does meta-policy │
	│ │ learn the meta-policy?) │ generalize?) │
	├────────────────────────────┼────────────────────────────┼────────────────────────────┤
	│ env.reset(seed=N, │ env.reset(seed=N) │ env.reset(seed=10000+N) │
	│ profile="introvert_ │ ← samples from training │ ← samples from a region │
	│ morning") │ distribution │ never seen in training│
	│ │ │ │
	│ → 3 hardcoded profiles │ → ~10 sampled profiles │ → ~10 sampled profiles │
	│ (from PROFILES list) │ from seeds 100..110 │ from seeds 10000..10010 │
	│ │ │ │
	│ Each strategy plays each │ Each strategy plays each │ Each strategy plays each │
	│ profile 5x = 15 episodes │ seed 1x = 10 episodes │ seed 1x = 10 episodes │
	│ │ │ │
	│ Strategies tested: │ Strategies tested: │ Strategies tested: │
	│ • random │ • random │ • random │
	│ • heuristic │ • heuristic │ • heuristic │
	│ • model (trained Qwen) │ • model (trained Qwen) │ • model (trained Qwen) │
	└────────────────────────────┴────────────────────────────┴────────────────────────────┘

	THE KEY METRIC: trained model's score on continuous-OOD vs heuristic baseline.

	Heuristic baseline (profile-blind hand rules): score 0.580 on OOD.
	Trained meta-RL agent's target: > 0.580 on OOD.

	If the trained agent beats the heuristic on OOD (profiles never seen in
	training), that's direct proof it learned the SKILL of profile inference,
	not just memorized training profiles.
	```

	---

	## 10. Sim → Real mapping (why this POC matters)

	The whole point of training in this env is that the SKILL transfers. The
	mapping from simulated signals to production signals is direct: same
	information, different source.

	```
	┌─────────────────────── SIMULATION (training) ───────┐ ┌──── REAL PRODUCT (deployment) ───────┐
	│ │ │ │
	│ meter: Vitality = 0.62 │──┼─► HRV from watch + sleep score │
	│ meter: Cognition = 0.51 │──┼─► focus app metrics + screen time │
	│ meter: Progress = 0.24 │──┼─► calendar density + task completion│
	│ meter: Serenity = 0.71 │──┼─► HRV variability + voice sentiment │
	│ meter: Connection = 0.38 │──┼─► call/message freq + calendar │
	│ │ │ │
	│ hidden profile (sampled per episode) │──┼─► the real user's actual personality│
	│ └─ never visible to agent │ │ └─ also never visible │
	│ │ │ │
	│ agent emits: "3 7 5 DEEP_WORK" │ │ │
	│ ├─ "3 7 5" = belief about user (hidden internal) │──┼─► agent's internal user model │
	│ └─ "DEEP_WORK" = action choice │──┼─► recommendation to user │
	│ │ │ ("I suggest a focus block now") │
	│ │ │ │
	│ per-step env_reward │──┼─► meter trend + accept/ignore tap │
	│ └─ "did meters improve under profile weights?" │ │ └─ "did user accept and benefit?"│
	│ │ │ │
	│ episode = 1 week (28 steps) │──┼─► rolling weekly window │
	│ │ │ │
	│ 3 eval conditions │ │ USER ARRIVES (cold start) │
	│ ├─ discrete-3 (memorization check) │ │ ├─ Agent has zero info about them │
	│ ├─ continuous-in-dist (training-distrib) │ │ ├─ ~5-10 interactions to converge │
	│ └─ continuous-OOD ◄── KEY ──► │──┼─► on a confident belief vector │
	│ agent must infer profile NEVER seen in │ │ └─ Acts on belief, learns from │
	│ training. THIS is the production scenario. │ │ tap responses │
	└─────────────────────────────────────────────────────┘ └───────────────────────────────────────┘

	SUCCESS METRIC IN BOTH WORLDS: how fast does the agent personalize?

	Sim: belief_accuracy curve over a single episode (does it climb?)
	adaptation_score (late-half mean reward > early-half mean reward?)

	Real: how many interactions until user's accept-rate stabilizes high?
	How quickly does the agent stop suggesting things the user always ignores?

	Both reduce to the same skill: form a model of this person from limited
	observation, then act on it.
	```

	Design choices that explicitly preserve the passive-only constraint:
	- ❌ No "ASK USER" action — agent never quizzes
	- ❌ No profile feature in observation at eval time — agent must infer
	- ❌ No explicit "what's your goal?" prompt — only sensor-equivalent meters
	- ✅ Belief output is INTERNAL (visible in POC only for training reward)
	- ✅ Reward formula uses only quantities computable from passive signals

	If the agent learns to do well on continuous-OOD (profiles never seen in
	training), it has acquired the skill of "figure out a new person from
	observation alone" — exactly the skill the production assistant needs
	when meeting a real user for the first time.

	---

	## 11. Spend & timing (concrete)

	```
	HARDWARE: A100 large (80GB) on HF Jobs at $2.50/hr ($0.0417/min)

	FAST_MODE (200-800 steps):
	dataset gen: ~2 min
	model load: ~3 min
	training: ~10-25 min (depends on steps)
	eval: ~3 min
	plot + upload: ~2 min
	────────────────────────────
	total: 20-35 min ($0.80-1.50 per iter)

	FULL RUN (2000 steps):
	dataset gen: ~3 min
	model load: ~3 min
	training: ~60-90 min
	eval: ~3 min
	plot + upload: ~2 min
	────────────────────────────
	total: 70-100 min ($3-4)

	Iter 1 (200 steps): $0.50 ❌ mode collapse (single action)
	Iter 2 (400 steps): $1.50 ❌ mode collapse (2-cycle)
	Iter 3 (800 steps): $5 ⏳ in flight (control)
	Iter 4 (800 steps): $5 ⏳ in flight (with full fixes)
	Final (2000 steps): $4 ⏳ pending iter 3+4 results
	──────
	~$16 of $30 budget
	```