core/scripts/generate_monotonic_cases.py

"""Generate the monotonic-learning cases.jsonl files for dev / public / holdouts.

Protocol shape (per split):

    cycle 0:               probe all probes (baseline)
    cycle 1..cycle_count:  one teach step (rotating domains) + probe all

Layout written:
    evals/monotonic_learning/dev/cases.jsonl
    evals/monotonic_learning/public/v1/cases.jsonl
    evals/monotonic_learning/holdouts/v1/cases.jsonl
"""
from __future__ import annotations

import json
from pathlib import Path
from typing import Sequence


def _probe(cycle: int, domain: str, probe_id: str, prompt: str, terms: list[str]) -> dict:
    return {
        "cycle": cycle,
        "op": "probe",
        "domain": domain,
        "id": probe_id,
        "prompt": prompt,
        "expected_terms": terms,
    }


def _teach(cycle: int, domain: str, prime: list[str], prompt: str) -> dict:
    return {
        "cycle": cycle,
        "op": "teach",
        "domain": domain,
        "prime": prime,
        "prompt": prompt,
    }


def build_split(
    *,
    out_path: Path,
    probes_per_domain: dict[str, list[tuple[str, str, list[str]]]],
    teaching_steps_per_domain: dict[str, list[tuple[list[str], str]]],
    cycle_count: int,
) -> int:
    domains: Sequence[str] = list(probes_per_domain.keys())
    teach_cursor: dict[str, int] = {d: 0 for d in domains}

    rows: list[dict] = []

    # Cycle 0: baseline probes only
    for d in domains:
        for probe_id, prompt, terms in probes_per_domain[d]:
            rows.append(_probe(0, d, probe_id, prompt, terms))

    # Cycles 1..N: one teach (rotating domain) + all probes
    for cycle in range(1, cycle_count + 1):
        teach_domain = domains[(cycle - 1) % len(domains)]
        steps = teaching_steps_per_domain[teach_domain]
        prime, prompt = steps[teach_cursor[teach_domain] % len(steps)]
        teach_cursor[teach_domain] += 1
        rows.append(_teach(cycle, teach_domain, prime, prompt))

        for d in domains:
            for probe_id, prompt_p, terms in probes_per_domain[d]:
                rows.append(_probe(cycle, d, probe_id, prompt_p, terms))

    out_path.parent.mkdir(parents=True, exist_ok=True)
    with out_path.open("w") as f:
        for row in rows:
            f.write(json.dumps(row, ensure_ascii=False) + "\n")

    return len(rows)


# Domain definitions: probes are deterministic queries; teaching steps are
# (prime turns, correction prompt) pairs.

_DOMAIN_TRUTH_PROBES = [
    ("PT-1", "What is truth?",           ["truth"]),
    ("PT-2", "Is truth coherent?",       ["truth"]),
    ("PT-3", "Why does truth matter?",   ["truth"]),
]

_DOMAIN_LIGHT_PROBES = [
    ("PL-1", "What is light?",            ["light"]),
    ("PL-2", "Why does light reveal?",    ["light"]),
    ("PL-3", "Is light revelation?",      ["light"]),
]

_DOMAIN_WISDOM_PROBES = [
    ("PW-1", "What is wisdom?",            ["wisdom"]),
    ("PW-2", "Is wisdom valuable?",        ["wisdom"]),
    ("PW-3", "Compare wisdom and knowledge", ["wisdom", "knowledge"]),
]

_DOMAIN_CREATION_PROBES = [
    ("PC-1", "What is creation?",          ["creation"]),
    ("PC-2", "Why does creation matter?",  ["creation"]),
    ("PC-3", "Is creation ongoing?",       ["creation"]),
]

_DOMAIN_KNOWLEDGE_PROBES = [
    ("PK-1", "What is knowledge?",         ["knowledge"]),
    ("PK-2", "Is knowledge wisdom?",       ["knowledge"]),
    ("PK-3", "Why does knowledge matter?", ["knowledge"]),
]


def _teach_steps_for(domain: str) -> list[tuple[list[str], str]]:
    """Three teaching examples per domain (rotated as cycles advance)."""
    base = f"What is {domain}?"
    return [
        ([base], f"Actually {domain} is more than that."),
        ([base], f"No, {domain} requires deeper understanding."),
        ([base], f"Actually {domain} relates to coherence."),
    ]


def main() -> None:
    root = Path(__file__).resolve().parent.parent / "evals" / "monotonic_learning"

    # Public v1: three domains x three probes, 12 cycles -> 9 + 12*(1+9) = 129 rows
    public_domains = {
        "truth":  _DOMAIN_TRUTH_PROBES,
        "light":  _DOMAIN_LIGHT_PROBES,
        "wisdom": _DOMAIN_WISDOM_PROBES,
    }
    public_teaching = {d: _teach_steps_for(d) for d in public_domains}
    n_public = build_split(
        out_path=root / "public" / "v1" / "cases.jsonl",
        probes_per_domain=public_domains,
        teaching_steps_per_domain=public_teaching,
        cycle_count=12,
    )

    # Dev: two domains x two probes, 10 cycles -> 4 + 10*(1+4) = 54 rows
    dev_domains = {
        "truth":  _DOMAIN_TRUTH_PROBES[:2],
        "light":  _DOMAIN_LIGHT_PROBES[:2],
    }
    dev_teaching = {d: _teach_steps_for(d) for d in dev_domains}
    n_dev = build_split(
        out_path=root / "dev" / "cases.jsonl",
        probes_per_domain=dev_domains,
        teaching_steps_per_domain=dev_teaching,
        cycle_count=10,
    )

    # Holdouts v1: distinct domains (creation, knowledge), three probes each,
    # 12 cycles -> 6 + 12*(1+6) = 90 rows
    holdout_domains = {
        "creation":  _DOMAIN_CREATION_PROBES,
        "knowledge": _DOMAIN_KNOWLEDGE_PROBES,
    }
    holdout_teaching = {d: _teach_steps_for(d) for d in holdout_domains}
    n_holdout = build_split(
        out_path=root / "holdouts" / "v1" / "cases.jsonl",
        probes_per_domain=holdout_domains,
        teaching_steps_per_domain=holdout_teaching,
        cycle_count=12,
    )

    print(f"wrote dev: {n_dev} rows")
    print(f"wrote public/v1: {n_public} rows")
    print(f"wrote holdouts/v1: {n_holdout} rows")


if __name__ == "__main__":
    main()