tasks_tough: add 10 domain-classifier tough scenarios (seed batch)

10 hand-crafted classification scenarios where the verbose hand-written
prompt is 200-300 tokens but the minimum effective prompt is non-obvious.
All use the existing exact_label scorer, no new scorer code needed.
Wired into _ALL_TASKS, train loop, and eval harness.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

server/prompt_golf_environment.py

@@ -48,6 +48,7 @@ try:
     from .target_model import TargetBackend, TargetGeneration, get_target_backend
     from .tasks import TASKS, TaskSpec, get_task, list_task_ids
     from .tasks_v2 import TASKS_V2
+    from .tasks_tough import TASKS_TOUGH
 except ImportError:
     from models import (
         DEFAULT_PROMPT_BUDGET,
@@ -62,9 +63,11 @@ except ImportError:
     from server.target_model import TargetBackend, TargetGeneration, get_target_backend
     from server.tasks import TASKS, TaskSpec, get_task, list_task_ids
     from server.tasks_v2 import TASKS_V2
+    from server.tasks_tough import TASKS_TOUGH
 
-# Merged v1 + v2 task bank. v2 task_ids don't clash with v1 by construction.
-_ALL_TASKS = {**TASKS, **TASKS_V2}
+# Merged v1 + v2 + tough task bank. task_ids don't clash by construction
+# (v2 tasks are uniquely named, tough tasks are prefixed `tough_`).
+_ALL_TASKS = {**TASKS, **TASKS_V2, **TASKS_TOUGH}
 
 
 # Baseline zero-shot scores are (target_id, task_id) -> score. Computed on

server/tasks_tough.py

@@ -0,0 +1,554 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""
+Tough-scenarios task bank for Prompt Golf (v3).
+
+Goal: scenarios where the *original* (verbose, hand-written) prompt that
+naturally steers the target is 150-300 tokens long, but the MINIMUM
+effective prompt is much shorter and non-obvious. The agent's job is to
+find that compressed prompt — i.e. learn which fragments of the verbose
+specification are load-bearing for the target model.
+
+This file is the seed batch (10 scenarios — domain classifiers). The
+remaining 42 will be added in later commits across:
+  - Structured extraction
+  - Format-strict generation
+  - Persona + constraint
+  - Multi-step reasoning
+  - Adversarial / calibration
+
+Why classifiers first: they exercise the existing `exact_label` scorer
+deterministically, so we can validate the whole base→trained CSV
+pipeline before investing in the fuzzier tasks.
+
+Each scenario follows the existing TaskSpec contract from server/tasks.py
+so it merges into _ALL_TASKS without code changes elsewhere.
+"""
+
+from __future__ import annotations
+
+try:
+    from .tasks import TaskSpec
+except ImportError:
+    from server.tasks import TaskSpec
+
+
+TASKS_TOUGH: dict[str, TaskSpec] = {}
+
+
+def _add(task: TaskSpec) -> None:
+    TASKS_TOUGH[task.task_id] = task
+
+
+def list_task_ids_tough() -> list[str]:
+    return list(TASKS_TOUGH.keys())
+
+
+# ============================================================================
+# Domain classifiers (10)
+#
+# All use scorer="exact_label". Expected output is exactly one token from a
+# closed vocabulary (lowercase, hyphenated, no punctuation, no explanation).
+# ============================================================================
+
+_add(TaskSpec(
+    task_id="tough_fallacy_classify",
+    category="classification_tough",
+    description=(
+        "Read the short argument and identify the dominant logical fallacy "
+        "it commits. The target must output exactly one label from this "
+        "closed vocabulary, in lowercase with hyphens, with no punctuation "
+        "and no explanation:\n"
+        "  - ad-hominem (attacking the person, not the argument)\n"
+        "  - straw-man (misrepresenting an opponent's position to refute it)\n"
+        "  - false-dilemma (presenting only two options when more exist)\n"
+        "  - slippery-slope (claiming one event inevitably leads to extreme "
+        "consequences without evidence)\n"
+        "  - appeal-to-authority (citing an irrelevant or unqualified "
+        "authority as proof)\n"
+        "  - circular-reasoning (the conclusion is assumed in the premises)\n"
+        "  - hasty-generalization (drawing a broad conclusion from a small "
+        "or biased sample)\n"
+        "  - red-herring (introducing an irrelevant topic to distract)\n"
+        "If multiple fallacies are present, choose the one most central to "
+        "the argument's structure. Output ONLY the label."
+    ),
+    scorer="exact_label",
+    train_examples=[
+        ("You can't trust Maria's economic analysis — she failed math in "
+         "high school.", "ad-hominem"),
+        ("Either we ban all cars or we accept that cities will be unlivable "
+         "forever.", "false-dilemma"),
+        ("My grandfather smoked his whole life and lived to 95, so smoking "
+         "isn't really dangerous.", "hasty-generalization"),
+    ],
+    test_examples=[
+        ("If we let students redo one exam, soon they'll demand to redo "
+         "every assignment and graduation will be meaningless.", "slippery-slope"),
+        ("Senator Park says climate policy is hurting jobs. He's been "
+         "divorced twice — why would anyone listen to him?", "ad-hominem"),
+        ("Of course the new drug works. It works because it's effective at "
+         "treating the condition.", "circular-reasoning"),
+        ("My opponent wants modest gun-safety reform. So she wants to "
+         "confiscate every firearm in America.", "straw-man"),
+        ("A famous actor endorses this supplement, so it must be "
+         "medically sound.", "appeal-to-authority"),
+        ("You ask about the budget overruns? Let's talk about how much "
+         "the previous administration wasted.", "red-herring"),
+    ],
+    budget_tokens=120,
+    difficulty="hard",
+    tags=["classification", "tough", "reasoning"],
+))
+
+
+_add(TaskSpec(
+    task_id="tough_bias_detect",
+    category="classification_tough",
+    description=(
+        "Identify the cognitive bias most clearly demonstrated by the "
+        "scenario. Output exactly one label from this closed vocabulary "
+        "(lowercase, hyphenated, no punctuation, no explanation):\n"
+        "  - confirmation (seeking/weighing only evidence that supports a "
+        "prior belief)\n"
+        "  - anchoring (over-relying on the first number or fact "
+        "encountered)\n"
+        "  - availability (judging probability by how easily examples come "
+        "to mind)\n"
+        "  - sunk-cost (continuing because of past investment rather than "
+        "future value)\n"
+        "  - survivorship (drawing conclusions from successful cases while "
+        "ignoring failed ones)\n"
+        "  - dunning-kruger (low-skill overconfidence; high-skill "
+        "under-confidence)\n"
+        "  - hindsight (believing past events were predictable after the "
+        "fact)\n"
+        "  - recency (overweighting the most recent data point)\n"
+        "Output ONLY the label."
+    ),
+    scorer="exact_label",
+    train_examples=[
+        ("After watching three plane-crash documentaries, Priya is now "
+         "afraid to fly even though she drives daily.", "availability"),
+        ("The first house Raj saw was listed at $800k. Every other house "
+         "now feels overpriced or like a steal compared to that number.",
+         "anchoring"),
+        ("Studied successful CEOs all dropped out of college, so dropping "
+         "out is the path to success.", "survivorship"),
+    ],
+    test_examples=[
+        ("I've already spent two years on this PhD topic — even though I "
+         "don't believe in it anymore, I have to finish.", "sunk-cost"),
+        ("After his stock dropped 8% yesterday, Arun is sure the whole "
+         "market is collapsing despite a steady year.", "recency"),
+        ("She only reads news outlets that agree with her political views "
+         "and dismisses the rest as biased.", "confirmation"),
+        ("After the company went bankrupt, every analyst said the warning "
+         "signs were obvious all along.", "hindsight"),
+        ("A first-year coder confidently tells the senior team their "
+         "architecture is wrong; she's never shipped to production.",
+         "dunning-kruger"),
+        ("He only studies founders of unicorn startups to figure out how "
+         "to build a unicorn, ignoring the thousands that failed.",
+         "survivorship"),
+    ],
+    budget_tokens=120,
+    difficulty="hard",
+    tags=["classification", "tough", "psychology"],
+))
+
+
+_add(TaskSpec(
+    task_id="tough_rhetorical_device",
+    category="classification_tough",
+    description=(
+        "Identify the dominant rhetorical device used in the sentence. "
+        "Output exactly one label from this closed vocabulary (lowercase, "
+        "no punctuation, no explanation):\n"
+        "  - anaphora (repeating the same word/phrase at the start of "
+        "successive clauses)\n"
+        "  - metaphor (implicit comparison, no 'like' or 'as')\n"
+        "  - hyperbole (deliberate, obvious exaggeration)\n"
+        "  - irony (saying the opposite of what is meant)\n"
+        "  - alliteration (repeated initial consonant sounds)\n"
+        "  - antithesis (juxtaposed contrasting ideas in parallel "
+        "structure)\n"
+        "  - rhetorical-question (a question asked for effect, not an "
+        "answer)\n"
+        "  - chiasmus (reversed grammatical structure: A-B-B-A)\n"
+        "Pick the device most central to the rhetorical effect. Output "
+        "ONLY the label."
+    ),
+    scorer="exact_label",
+    train_examples=[
+        ("We shall fight on the beaches. We shall fight on the landing "
+         "grounds. We shall fight in the fields.", "anaphora"),
+        ("Time is a thief that steals our youth.", "metaphor"),
+        ("Ask not what your country can do for you — ask what you can do "
+         "for your country.", "chiasmus"),
+    ],
+    test_examples=[
+        ("I've told you a million times to clean your room.", "hyperbole"),
+        ("Peter Piper picked a peck of pickled peppers.", "alliteration"),
+        ("It is the best of times, it is the worst of times.", "antithesis"),
+        ("Oh great, another Monday — exactly what I was hoping for.",
+         "irony"),
+        ("Do we really need another committee to study the obvious?",
+         "rhetorical-question"),
+        ("The classroom was a zoo during the substitute's lesson.",
+         "metaphor"),
+    ],
+    budget_tokens=120,
+    difficulty="hard",
+    tags=["classification", "tough", "literature"],
+))
+
+
+_add(TaskSpec(
+    task_id="tough_medical_urgency",
+    category="classification_tough",
+    description=(
+        "Read the short patient complaint and triage it into one of four "
+        "urgency tiers. Output exactly one label, lowercase, no "
+        "punctuation, no explanation. This is a TRIAGE classifier — not "
+        "medical advice — so be conservative on red-flag symptoms.\n"
+        "  - emergency: chest pain, stroke signs, severe bleeding, loss of "
+        "consciousness, anaphylaxis, suicidal intent — call ambulance now.\n"
+        "  - urgent: significant injury, high fever with stiffness, severe "
+        "pain, infection signs, pregnancy complications — same-day care.\n"
+        "  - routine: persistent but stable symptoms, follow-ups, "
+        "medication refills — schedule within 1-2 weeks.\n"
+        "  - self-care: minor cuts, common cold, mild headache, fatigue "
+        "without alarm features — rest, OTC, monitor.\n"
+        "Default to the higher tier when symptoms are ambiguous. Output "
+        "ONLY the label."
+    ),
+    scorer="exact_label",
+    train_examples=[
+        ("Crushing chest pressure radiating to my left arm, sweating, "
+         "started 20 minutes ago.", "emergency"),
+        ("Sore throat and runny nose for two days, no fever, eating "
+         "normally.", "self-care"),
+        ("Rash on forearm that's been spreading slowly for three weeks, "
+         "no pain.", "routine"),
+    ],
+    test_examples=[
+        ("Sudden numbness on one side of my face and slurred speech for "
+         "the last 10 minutes.", "emergency"),
+        ("Deep cut on my hand from a kitchen knife, bleeding has slowed "
+         "but it might need stitches.", "urgent"),
+        ("Ongoing knee stiffness in the mornings for the past month, "
+         "manageable.", "routine"),
+        ("Mild headache after a long day on screens, no other symptoms.",
+         "self-care"),
+        ("High fever 39.5C, stiff neck, and a new pinpoint rash that "
+         "started this evening.", "emergency"),
+        ("Persistent cough for four days, low-grade fever, achy but "
+         "drinking fluids and resting.", "urgent"),
+    ],
+    budget_tokens=140,
+    difficulty="hard",
+    tags=["classification", "tough", "medical"],
+))
+
+
+_add(TaskSpec(
+    task_id="tough_code_smell",
+    category="classification_tough",
+    description=(
+        "Read the short code description and identify the dominant code "
+        "smell. Output exactly one label from this closed vocabulary "
+        "(lowercase, hyphenated, no punctuation, no explanation):\n"
+        "  - long-method (a single function does too many things over too "
+        "many lines)\n"
+        "  - god-class (one class accumulates unrelated responsibilities)\n"
+        "  - duplicate-code (the same logic appears in multiple places)\n"
+        "  - dead-code (unused variables, branches, or functions)\n"
+        "  - magic-number (unexplained literal constants in logic)\n"
+        "  - primitive-obsession (using strings/ints where a small type "
+        "would clarify intent)\n"
+        "  - feature-envy (a method uses another class's data more than "
+        "its own)\n"
+        "  - shotgun-surgery (one logical change requires edits across "
+        "many files)\n"
+        "Output ONLY the label."
+    ),
+    scorer="exact_label",
+    train_examples=[
+        ("`processOrder()` is 600 lines long and handles validation, "
+         "pricing, payment, shipping, email, and audit logging in one "
+         "function.", "long-method"),
+        ("`if total > 4500: applyDiscount(0.07)` — neither number is "
+         "explained.", "magic-number"),
+        ("Adding a new currency requires editing the database schema, "
+         "three services, the UI, and two config files.",
+         "shotgun-surgery"),
+    ],
+    test_examples=[
+        ("`UserManager` handles authentication, profile editing, billing, "
+         "email sending, audit logs, and CSV export.", "god-class"),
+        ("The same 30-line block computing tax appears in CheckoutService, "
+         "InvoiceService, and ReportService.", "duplicate-code"),
+        ("`Order.calculateShipping()` reads 8 fields from `Customer` and "
+         "uses only 1 from its own object.", "feature-envy"),
+        ("There's a private helper `oldFormatLegacy()` that nothing in "
+         "the repo references anymore.", "dead-code"),
+        ("Phone numbers, emails, postal codes, and currency amounts are "
+         "all stored as plain `str` everywhere.", "primitive-obsession"),
+        ("A single function `handleRequest()` parses input, validates, "
+         "queries DB, formats output, logs, and emails — 400 lines.",
+         "long-method"),
+    ],
+    budget_tokens=140,
+    difficulty="hard",
+    tags=["classification", "tough", "software"],
+))
+
+
+_add(TaskSpec(
+    task_id="tough_news_framing",
+    category="classification_tough",
+    description=(
+        "Read the short news headline and identify its dominant framing "
+        "technique. Output exactly one label from this closed vocabulary "
+        "(lowercase, hyphenated, no punctuation, no explanation):\n"
+        "  - episodic (focuses on a single event or individual case)\n"
+        "  - thematic (focuses on broader trends, statistics, or "
+        "context)\n"
+        "  - conflict (frames the story as a clash between sides)\n"
+        "  - human-interest (emotional angle on a person's experience)\n"
+        "  - economic (frames consequences in financial / market terms)\n"
+        "  - morality (frames the story in terms of right vs wrong, "
+        "values)\n"
+        "  - responsibility (assigns blame or credit to a specific "
+        "actor)\n"
+        "Pick the dominant frame even if minor frames are present. Output "
+        "ONLY the label."
+    ),
+    scorer="exact_label",
+    train_examples=[
+        ("Single mother of three struggles to afford groceries as prices "
+         "rise.", "human-interest"),
+        ("National food-insecurity rate hits 12-year high, USDA report "
+         "shows.", "thematic"),
+        ("Senate Democrats and Republicans clash over food-stamp "
+         "spending bill.", "conflict"),
+    ],
+    test_examples=[
+        ("Local bakery owner closes shop after 30 years, blames soaring "
+         "rent.", "episodic"),
+        ("Inflation eats into household budgets as wages stagnate.",
+         "economic"),
+        ("Mayor accused of approving the contract that caused the water "
+         "crisis.", "responsibility"),
+        ("Is it ever right to lie to protect a friend? Readers weigh in.",
+         "morality"),
+        ("Climate-policy fight escalates as governors trade public "
+         "letters.", "conflict"),
+        ("Childhood obesity rates nationwide climbed 4% over the last "
+         "decade.", "thematic"),
+    ],
+    budget_tokens=130,
+    difficulty="hard",
+    tags=["classification", "tough", "media"],
+))
+
+
+_add(TaskSpec(
+    task_id="tough_claim_verifiability",
+    category="classification_tough",
+    description=(
+        "Classify the claim by what kind of statement it is. Output "
+        "exactly one label, lowercase, hyphenated, no punctuation, no "
+        "explanation:\n"
+        "  - verifiable (in principle checkable against publicly "
+        "available facts or measurements)\n"
+        "  - unverifiable (about private mental states, future events, "
+        "or otherwise not externally checkable)\n"
+        "  - value-judgment (expresses a preference, taste, or moral "
+        "evaluation rather than a fact)\n"
+        "  - tautology (true by definition; carries no empirical "
+        "content)\n"
+        "Distinguish carefully: an unverifiable empirical claim is NOT "
+        "the same as a value-judgment. Output ONLY the label."
+    ),
+    scorer="exact_label",
+    train_examples=[
+        ("The Eiffel Tower is 330 meters tall.", "verifiable"),
+        ("Chocolate ice cream is the best dessert ever invented.",
+         "value-judgment"),
+        ("All bachelors are unmarried.", "tautology"),
+    ],
+    test_examples=[
+        ("The president secretly regrets signing the trade deal last "
+         "year.", "unverifiable"),
+        ("Mount Everest is taller than Mount Kilimanjaro.", "verifiable"),
+        ("A triangle has three sides.", "tautology"),
+        ("Modern art is shallow and pretentious.", "value-judgment"),
+        ("Earth's average surface temperature has risen since 1900.",
+         "verifiable"),
+        ("Pluto will be reclassified as a planet again before 2050.",
+         "unverifiable"),
+    ],
+    budget_tokens=120,
+    difficulty="hard",
+    tags=["classification", "tough", "epistemology"],
+))
+
+
+_add(TaskSpec(
+    task_id="tough_argument_strength",
+    category="classification_tough",
+    description=(
+        "Evaluate the short argument and classify its logical status. "
+        "Output exactly one label, lowercase, hyphenated, no punctuation, "
+        "no explanation:\n"
+        "  - sound (valid form AND all premises are true / plausibly "
+        "true)\n"
+        "  - valid-but-unsound (the conclusion follows IF the premises "
+        "are true, but at least one premise is false)\n"
+        "  - invalid (the conclusion does NOT follow from the premises "
+        "even if they were true)\n"
+        "  - fallacious (commits a recognized informal fallacy that "
+        "undermines the inference)\n"
+        "Apply this order of priority: if the argument commits a clear "
+        "informal fallacy, label it `fallacious` over `invalid`. Output "
+        "ONLY the label."
+    ),
+    scorer="exact_label",
+    train_examples=[
+        ("All humans are mortal. Socrates is human. Therefore Socrates is "
+         "mortal.", "sound"),
+        ("All birds can fly. Penguins are birds. Therefore penguins can "
+         "fly.", "valid-but-unsound"),
+        ("Some dogs are brown. My cat is brown. Therefore my cat is a "
+         "dog.", "invalid"),
+    ],
+    test_examples=[
+        ("The new policy must be wrong because the senator proposing it "
+         "had an affair last year.", "fallacious"),
+        ("All squares have four sides. This shape is a square. Therefore "
+         "it has four sides.", "sound"),
+        ("If it rains, the streets get wet. The streets are wet. "
+         "Therefore it rained.", "invalid"),
+        ("Every prime number is odd. Seven is prime. Therefore seven is "
+         "odd.", "valid-but-unsound"),
+        ("Either you support our tax bill or you hate working families.",
+         "fallacious"),
+        ("All mammals are warm-blooded. Whales are mammals. Therefore "
+         "whales are warm-blooded.", "sound"),
+    ],
+    budget_tokens=140,
+    difficulty="hard",
+    tags=["classification", "tough", "logic"],
+))
+
+
+_add(TaskSpec(
+    task_id="tough_emotion_primary",
+    category="classification_tough",
+    description=(
+        "Identify the dominant primary emotion expressed by the speaker, "
+        "using Plutchik's eight basic emotions. Output exactly one label, "
+        "lowercase, no punctuation, no explanation:\n"
+        "  - joy (happiness, delight, contentment)\n"
+        "  - trust (acceptance, confidence in someone/something)\n"
+        "  - fear (apprehension, worry about a threat)\n"
+        "  - surprise (unexpectedness, being caught off guard)\n"
+        "  - sadness (sorrow, loss, dejection)\n"
+        "  - disgust (revulsion, moral or physical aversion)\n"
+        "  - anger (frustration, hostility, indignation)\n"
+        "  - anticipation (expectation, looking forward)\n"
+        "Pick the SINGLE strongest emotion even if blends are present. "
+        "Output ONLY the label."
+    ),
+    scorer="exact_label",
+    train_examples=[
+        ("I can't believe she actually showed up — I had no idea she was "
+         "in town!", "surprise"),
+        ("My team has my back; I know they'll deliver no matter what.",
+         "trust"),
+        ("Everything I worked for these last five years is just gone.",
+         "sadness"),
+    ],
+    test_examples=[
+        ("I'm counting down the days until the trip — only two weeks "
+         "left!", "anticipation"),
+        ("How DARE they reroute my flight without a single email?",
+         "anger"),
+        ("Reading those emails made my skin crawl. I had to stop "
+         "halfway.", "disgust"),
+        ("What if the test results come back bad? I haven't slept in "
+         "days.", "fear"),
+        ("Got the offer, the salary, AND the team I wanted — best week "
+         "ever.", "joy"),
+        ("I keep replaying the call. She just isn't coming back.",
+         "sadness"),
+    ],
+    budget_tokens=130,
+    difficulty="hard",
+    tags=["classification", "tough", "psychology"],
+))
+
+
+_add(TaskSpec(
+    task_id="tough_policy_stance",
+    category="classification_tough",
+    description=(
+        "Classify the speaker's stance on the policy proposal mentioned "
+        "in the quote. Output exactly one label, lowercase, hyphenated, "
+        "no punctuation, no explanation:\n"
+        "  - support (clearly endorses the proposal)\n"
+        "  - oppose (clearly rejects the proposal)\n"
+        "  - neutral (declines to take a side, observes both views, or "
+        "stays purely descriptive)\n"
+        "  - conditional-support (would support IF certain conditions "
+        "were met)\n"
+        "  - conditional-oppose (would oppose UNLESS certain conditions "
+        "were met)\n"
+        "Distinguish carefully: a hedged endorsement that names "
+        "preconditions is conditional-support, not neutral. A statement "
+        "of mixed views without a stance is neutral. Output ONLY the "
+        "label."
+    ),
+    scorer="exact_label",
+    train_examples=[
+        ("I'm fully behind the rent-cap proposal — it'll protect "
+         "vulnerable tenants.", "support"),
+        ("The mining permit is a disaster for the watershed and I will "
+         "vote no.", "oppose"),
+        ("Some economists like the tariff plan, others don't — the "
+         "evidence is genuinely mixed.", "neutral"),
+    ],
+    test_examples=[
+        ("I'd back the carbon-tax bill, but only if the revenue is "
+         "rebated to households.", "conditional-support"),
+        ("I cannot support the surveillance program unless judicial "
+         "review is built in from day one.", "conditional-oppose"),
+        ("The infrastructure package is exactly what this district has "
+         "needed for a decade.", "support"),
+        ("I won't comment on the merits of the bill; that's for the "
+         "committee to weigh.", "neutral"),
+        ("This zoning change will gut the neighborhood — count me as a "
+         "firm no.", "oppose"),
+        ("I'll support the immigration reform if it includes a real "
+         "pathway to citizenship.", "conditional-support"),
+    ],
+    budget_tokens=140,
+    difficulty="hard",
+    tags=["classification", "tough", "politics"],
+))
+
+
+# ============================================================================
+# Module-level helpers
+# ============================================================================
+
+if __name__ == "__main__":
+    print(f"tasks_tough: {len(TASKS_TOUGH)} scenarios")
+    for tid, spec in TASKS_TOUGH.items():
+        print(f"  {tid:36s} {spec.category:24s} budget={spec.budget_tokens}")

training/eval_before_after.py

@@ -126,11 +126,12 @@ def main() -> None:
     from prompt_golf_env.server.prompt_golf_environment import PromptGolfEnvironment
     from prompt_golf_env.server.tasks import TASKS, list_task_ids as list_v1
     from prompt_golf_env.server.tasks_v2 import TASKS_V2, list_task_ids_v2
+    from prompt_golf_env.server.tasks_tough import TASKS_TOUGH, list_task_ids_tough
 
-    _ALL_TASKS = {**TASKS, **TASKS_V2}
+    _ALL_TASKS = {**TASKS, **TASKS_V2, **TASKS_TOUGH}
 
     def list_task_ids():
-        return list_v1() + list_task_ids_v2()
+        return list_v1() + list_task_ids_v2() + list_task_ids_tough()
 
     # Load agent
     model, tok = load_agent(args.agent_model, args.adapter)

training/train_grpo.py

@@ -324,6 +324,7 @@ def main() -> None:
     from prompt_golf_env.server.prompt_golf_environment import PromptGolfEnvironment
     from prompt_golf_env.server.tasks import list_task_ids as list_task_ids_v1
     from prompt_golf_env.server.tasks_v2 import list_task_ids_v2
+    from prompt_golf_env.server.tasks_tough import list_task_ids_tough
 
     # NOTE: we deliberately do NOT import Unsloth here. Unsloth patches
     # Qwen2Attention at import time, which breaks the target model
@@ -343,7 +344,7 @@ def main() -> None:
 
     # ----- env (target loaded lazily on first forward pass) -----
     env = PromptGolfEnvironment()
-    all_tasks = list_task_ids_v1() + list_task_ids_v2()
+    all_tasks = list_task_ids_v1() + list_task_ids_v2() + list_task_ids_tough()
     held_out = {t.strip() for t in args.held_out_tasks.split(",") if t.strip()}
     train_tasks = [t for t in all_tasks if t not in held_out]
     print(f"[setup] tasks total={len(all_tasks)} train={len(train_tasks)} held_out={len(held_out)}", flush=True)