scripts/build_protocols.py

"""

Silicon Subjects — protocol suite builder.



Writes data/protocols.jsonl: classic experimental protocols from behavioural

economics and psychology, each paired with *human anchor* statistics taken from

the published literature, split along the three axes a simulated population has

to reproduce:



    1. location    — the level of the response (mean)

    2. dispersion  — the spread of the response (sd)

    3. effect      — the direction and magnitude of the treatment contrast



Anchors are literature aggregates, not raw subject-level data. They are meant as

reference points for auditing a simulated population, NOT as exact replications;

always re-check against the cited source before drawing conclusions.



Usage:

    python scripts/build_protocols.py [--out-dir data]

"""

from __future__ import annotations

import argparse
import json
from pathlib import Path
from typing import Any, Dict, List, Optional


def P(  # noqa: N802 - short builder alias

    protocol_id: str,

    name: str,

    domain: str,

    paradigm: str,

    source: str,

    response: Dict[str, Any],

    conditions: List[Dict[str, str]],

    anchors: Dict[str, Dict[str, Any]],

    effect: Optional[Dict[str, Any]] = None,

    shape: Optional[str] = None,

    notes: str = "",

) -> Dict[str, Any]:
    return {
        "protocol_id": protocol_id,
        "name": name,
        "domain": domain,
        "paradigm": paradigm,
        "source": source,
        "response": response,
        "conditions": conditions,
        "anchors": anchors,
        "effect": effect,
        "shape": shape,
        "notes": notes,
    }


NUM = lambda lo, hi, unit: {"type": "numeric", "min": lo, "max": hi, "unit": unit}  # noqa: E731
BIN = lambda lo_label, hi_label: {  # noqa: E731
    "type": "binary",
    "min": 0,
    "max": 1,
    "unit": f"1 = {hi_label}, 0 = {lo_label}",
}


# ---------------------------------------------------------------------------
# 1-8. Social preferences
# ---------------------------------------------------------------------------

DICTATOR = P(
    "dictator_game",
    "Dictator game",
    "social_preferences",
    "one-shot unilateral allocation",
    "Engel (2011) 'Dictator games: a meta study', Experimental Economics 14:583-610",
    NUM(0, 10, "dollars transferred out of a $10 endowment"),
    [
        {
            "condition_id": "single",
            "label": "single decision",
            "prompt": (
                "You are taking part in an economics experiment. You have been given $10. "
                "You may transfer any amount from $0 to $10 to another participant, who "
                "cannot respond and cannot affect your payoff. You keep whatever you do "
                "not transfer. How many dollars do you transfer? Answer with a single number."
            ),
        }
    ],
    {
        "single": {
            "mean": 2.83,
            "sd": 2.60,
            "confidence": "medium",
            "note": "Meta-analytic mean share is ~28% of the stake; individual-level dispersion "
            "varies widely across subject pools, so treat sd as approximate.",
        }
    },
    shape="bimodal — mass at $0 and at half the stake",
    notes="The canonical other-regarding-preference measure. A simulated population that gives "
    "the 'fair' 50% is reproducing a norm, not the human distribution.",
)

ULTIMATUM_OFFER = P(
    "ultimatum_offer",
    "Ultimatum game — proposer offer",
    "social_preferences",
    "one-shot bargaining, proposer role",
    "Oosterbeek, Sloof & van de Kuilen (2004) 'Cultural differences in ultimatum game "
    "experiments', Experimental Economics 7:171-188; Camerer (2003)",
    NUM(0, 10, "dollars offered out of a $10 pie"),
    [
        {
            "condition_id": "proposer",
            "label": "proposer",
            "prompt": (
                "You are the proposer. You and one other participant will split $10. You "
                "propose how much of the $10 the other person receives. If they accept, you "
                "both get the proposed split; if they reject, neither of you gets anything. "
                "How many dollars do you offer? Answer with a single number."
            ),
        }
    ],
    {
        "proposer": {
            "mean": 4.00,
            "sd": 1.30,
            "confidence": "medium-high",
            "note": "Meta-analytic mean offer is ~40% of the pie, with a strong mode at 50%.",
        }
    },
    shape="unimodal with a spike at 50%",
    notes="Offers cluster at the equal split; very low offers are rare because proposers "
    "anticipate rejection.",
)

ULTIMATUM_REJECT = P(
    "ultimatum_reject_low",
    "Ultimatum game — rejection of a low offer",
    "social_preferences",
    "one-shot bargaining, responder role",
    "Camerer (2003) Behavioral Game Theory; Oosterbeek et al. (2004)",
    BIN("accept", "reject"),
    [
        {
            "condition_id": "low_offer",
            "label": "offer of $2 out of $10",
            "prompt": (
                "You are the responder. The proposer offers you $2 out of $10 and keeps $8. "
                "If you accept, you get $2 and they get $8. If you reject, neither of you "
                "gets anything. Answer 1 if you reject the offer, 0 if you accept it."
            ),
        }
    ],
    {
        "low_offer": {
            "mean": 0.50,
            "sd": None,
            "confidence": "medium",
            "note": "Rejection probability for offers at or below 20% of the pie is roughly "
            "40-60%; it rises steeply as the offer falls.",
        }
    },
    notes="Human responders punish unfairness at a material cost to themselves. Populations "
    "that never reject are behaving like pure payoff maximisers.",
)

PUBLIC_GOODS = P(
    "public_goods_oneshot",
    "Public goods game — one shot",
    "social_preferences",
    "linear voluntary contribution mechanism, one-shot",
    "Ledyard (1995) in Kagel & Roth (eds.), Handbook of Experimental Economics; "
    "Chaudhuri (2011) 'Sustaining cooperation in laboratory public goods experiments', "
    "Experimental Economics 14:47-83",
    NUM(0, 10, "tokens contributed out of 10"),
    [
        {
            "condition_id": "oneshot",
            "label": "one-shot, 4 players, MPCR 0.4",
            "prompt": (
                "You are in a group of 4 people. Each person receives 10 tokens. Any token "
                "you put into a common pool is multiplied by 1.6 and then split equally "
                "among the 4 members. Tokens you keep are worth 1 point each to you. This "
                "is a single, one-shot decision with the same 4 people. How many of your 10 "
                "tokens do you contribute to the pool? Answer with a single number."
            ),
        }
    ],
    {
        "oneshot": {
            "mean": 4.50,
            "sd": 2.60,
            "confidence": "medium",
            "note": "One-shot contributions typically run 40-60% of the endowment; "
            "individual-level spread is large.",
        }
    },
    notes="Contributions start well above the Nash prediction of zero. Free-riding is a "
    "minority behaviour, not the modal one.",
)

PUBLIC_GOODS_DECAY = P(
    "public_goods_decay",
    "Public goods game — decay across rounds",
    "social_preferences",
    "repeated voluntary contribution, partners condition",
    "Ledyard (1995); Chaudhuri (2011); Isaac & Walker (1988)",
    NUM(0, 10, "tokens contributed out of 10"),
    [
        {
            "condition_id": "round_1",
            "label": "round 1 of 10",
            "prompt": (
                "You are in a group of 4 people and will play 10 rounds with the same "
                "people. Each round every person receives 10 tokens; tokens placed in a "
                "common pool are multiplied by 1.6 and split equally among the 4 members. "
                "This is round 1. How many of your 10 tokens do you contribute? Answer with "
                "a single number."
            ),
        },
        {
            "condition_id": "round_10",
            "label": "round 10 of 10 (final round)",
            "prompt": (
                "You are in a group of 4 people and have played 9 rounds with the same "
                "people. Each round every person receives 10 tokens; tokens placed in a "
                "common pool are multiplied by 1.6 and split equally among the 4 members. "
                "The average contribution started around 4 tokens and has fallen steadily. "
                "This is round 10, the final round. How many of your 10 tokens do you "
                "contribute? Answer with a single number."
            ),
        },
    ],
    {
        "round_1": {"mean": 4.50, "sd": 2.60, "confidence": "medium", "note": ""},
        "round_10": {"mean": 1.80, "sd": 2.00, "confidence": "medium", "note": ""},
    },
    effect={
        "a": "round_1",
        "b": "round_10",
        "mean_diff": 2.70,
        "cohen_d": 1.10,
        "confidence": "medium-high",
        "source": "Ledyard (1995); Chaudhuri (2011)",
        "note": "The decay of cooperation without punishment or communication is one of the "
        "most robust regularities in experimental economics.",
    },
    notes="A population that holds contributions flat across rounds is missing the entire "
    "dynamics of conditional cooperation.",
)

TRUST_SEND = P(
    "trust_game_send",
    "Trust game — amount sent",
    "social_preferences",
    "one-shot investment game, trustor role",
    "Johnson & Mislin (2011) 'Trust games: a meta-analysis', Journal of Economic "
    "Psychology 32:865-889",
    NUM(0, 10, "dollars sent out of $10 (tripled before transfer)"),
    [
        {
            "condition_id": "trustor",
            "label": "trustor",
            "prompt": (
                "You are given $10 and paired with a stranger. You may send any amount from "
                "$0 to $10. Any amount you send is tripled by the experimenter and given to "
                "the stranger, who then decides how much, if anything, to send back to you. "
                "How many dollars do you send? Answer with a single number."
            ),
        }
    ],
    {
        "trustor": {
            "mean": 5.00,
            "sd": 2.70,
            "confidence": "medium",
            "note": "Meta-analytic mean send is ~50% of the endowment.",
        }
    },
    notes="Sending is positive in nearly every subject pool, but far from the payoff-maximising "
    "100%.",
)

TRUST_RETURN = P(
    "trust_game_return",
    "Trust game — amount returned",
    "social_preferences",
    "one-shot investment game, trustee role",
    "Johnson & Mislin (2011)",
    NUM(0, 18, "dollars returned out of $18 available"),
    [
        {
            "condition_id": "trustee",
            "label": "trustee",
            "prompt": (
                "A stranger was given $10 and sent you $6. The experimenter tripled it, so "
                "you have $18. You may send back any amount from $0 to $18 to the stranger. "
                "How many dollars do you send back? Answer with a single number."
            ),
        }
    ],
    {
        "trustee": {
            "mean": 6.50,
            "sd": 4.00,
            "confidence": "approximate",
            "note": "Meta-analytic 'trustworthiness' is ~0.37 of the amount available; values "
            "here are approximate and sensitive to the amount sent.",
        }
    },
    notes="Reciprocity, not pure self-interest: trustees typically return enough to make "
    "sending worthwhile.",
)

PRISONERS = P(
    "prisoners_dilemma",
    "Prisoner's dilemma — one shot",
    "social_preferences",
    "one-shot 2x2 symmetric game",
    "Sally (1995) 'Conversation and cooperation in social dilemmas', Rationality and "
    "Society 7:58-92",
    BIN("defect", "cooperate"),
    [
        {
            "condition_id": "oneshot",
            "label": "one-shot, anonymous, no communication",
            "prompt": (
                "You and one other person each choose to cooperate or defect, simultaneously "
                "and anonymously. You will never meet again. If both cooperate, each gets $6. "
                "If both defect, each gets $2. If you cooperate and the other defects, you get "
                "$0 and they get $9. If you defect and the other cooperates, you get $9 and "
                "they get $0. Answer 1 to cooperate, 0 to defect."
            ),
        }
    ],
    {
        "oneshot": {
            "mean": 0.47,
            "sd": None,
            "confidence": "medium",
            "note": "Meta-analytic cooperation in the one-shot prisoner's dilemma is ~45-50%; "
            "it collapses toward 0 with experience and rises with communication.",
        }
    },
    notes="Cooperation in the one-shot game is the benchmark case where a purely rational "
    "population and a human population diverge most sharply.",
)


# ---------------------------------------------------------------------------
# 9-14. Heuristics and biases
# ---------------------------------------------------------------------------

ASCH = P(
    "asch_conformity",
    "Asch line-judgement conformity",
    "social_influence",
    "unambiguous perceptual judgement under group pressure",
    "Asch (1951, 1955); Bond & Smith (1996) 'Culture and conformity', Psychological "
    "Bulletin 119:111-137",
    BIN("give the correct answer", "conform to the group's wrong answer"),
    [
        {
            "condition_id": "critical",
            "label": "critical trial, unanimous wrong majority",
            "prompt": (
                "You are judging which of three comparison lines matches a standard line in "
                "length. The correct match is obvious and you are confident about it. Six "
                "other participants answer before you and all six give the same clearly "
                "incorrect answer. It is now your turn, aloud. Answer 1 if you give the "
                "group's incorrect answer, 0 if you give the obviously correct one."
            ),
        }
    ],
    {
        "critical": {
            "mean": 0.32,
            "sd": None,
            "confidence": "medium",
            "note": "~32% of critical trials draw a conforming response in the classic "
            "setting; roughly 75% of participants conform at least once and ~25% never do.",
        }
    },
    notes="The cleanest available test of whether a simulated subject is willing to contradict "
    "an unambiguous fact under social pressure.",
)

ANCHORING = P(
    "anchoring",
    "Anchoring — comparative anchor then estimate",
    "heuristics_biases",
    "anchoring-and-adjustment",
    "Jacowitz & Kahneman (1995) 'Measures of anchoring in estimation tasks', Personality "
    "and Social Psychology Bulletin 21:1161-1166; Tversky & Kahneman (1974)",
    NUM(0, 100, "estimated percentage of African countries in the UN"),
    [
        {
            "condition_id": "high",
            "label": "high anchor (65%)",
            "prompt": (
                "Is the percentage of African countries in the United Nations more or less "
                "than 65%? Now estimate, as accurately as you can, the actual percentage of "
                "African countries in the United Nations. Answer with a single number "
                "between 0 and 100."
            ),
        },
        {
            "condition_id": "low",
            "label": "low anchor (10%)",
            "prompt": (
                "Is the percentage of African countries in the United Nations more or less "
                "than 10%? Now estimate, as accurately as you can, the actual percentage of "
                "African countries in the United Nations. Answer with a single number "
                "between 0 and 100."
            ),
        },
    ],
    {},
    effect={
        "a": "high",
        "b": "low",
        "mean_diff": 20.0,
        "cohen_d": 0.90,
        "confidence": "approximate",
        "source": "Jacowitz & Kahneman (1995)",
        "note": "Absolute estimates are item-specific, so only the contrast is anchored here; "
        "the high-low gap is typically 15-25 percentage points.",
    },
    notes="Deliberately has no level anchor: the point of this protocol is the contrast, not "
    "the absolute estimate. This is the pattern to use whenever only an effect size is "
    "trustworthy.",
)

ASIAN_DISEASE = P(
    "asian_disease",
    "Asian disease problem — risky-choice framing",
    "heuristics_biases",
    "gain vs. loss frame, logically equivalent options",
    "Tversky & Kahneman (1981) 'The framing of decisions and the psychology of choice', "
    "Science 211:453-458",
    BIN("choose the risky option", "choose the sure option"),
    [
        {
            "condition_id": "gain",
            "label": "gain frame (lives saved)",
            "prompt": (
                "Imagine the US is preparing for an unusual disease expected to kill 600 "
                "people. Two programmes are proposed. If programme A is adopted, 200 people "
                "will be saved. If programme B is adopted, there is a one-third probability "
                "that 600 people will be saved and a two-thirds probability that no one will "
                "be saved. Answer 1 for programme A (the sure option), 0 for programme B."
            ),
        },
        {
            "condition_id": "loss",
            "label": "loss frame (lives lost)",
            "prompt": (
                "Imagine the US is preparing for an unusual disease expected to kill 600 "
                "people. Two programmes are proposed. If programme C is adopted, 400 people "
                "will die. If programme D is adopted, there is a one-third probability that "
                "nobody will die and a two-thirds probability that 600 people will die. "
                "Answer 1 for programme C (the sure option), 0 for programme D."
            ),
        },
    ],
    {
        "gain": {"mean": 0.72, "sd": None, "confidence": "high", "note": ""},
        "loss": {"mean": 0.22, "sd": None, "confidence": "high", "note": ""},
    },
    effect={
        "a": "gain",
        "b": "loss",
        "mean_diff": 0.50,
        "cohen_d": 1.00,
        "confidence": "high",
        "source": "Tversky & Kahneman (1981)",
        "note": "Options A/C and B/D are formally identical; only the description differs.",
    },
    notes="One of the most replicated effects in the literature. A population that is "
    "frame-invariant here is not behaving like a human sample.",
)

ENDOWMENT = P(
    "endowment_effect",
    "Endowment effect — WTA vs. WTP",
    "heuristics_biases",
    "buyer vs. seller valuation of the same object",
    "Kahneman, Knetsch & Thaler (1990) 'Experimental tests of the endowment effect and "
    "the Coase theorem', Journal of Political Economy 98:1325-1348",
    NUM(0, 10, "dollars"),
    [
        {
            "condition_id": "wtp",
            "label": "buyer — maximum willingness to pay",
            "prompt": (
                "A coffee mug is for sale in the experiment. You do not own one. What is the "
                "largest amount, from $0 to $10, that you would be willing to pay to buy this "
                "mug? Answer with a single number."
            ),
        },
        {
            "condition_id": "wta",
            "label": "seller — minimum willingness to accept",
            "prompt": (
                "You have just been given a coffee mug and now own it. What is the smallest "
                "amount, from $0 to $10, that you would be willing to accept to sell this "
                "mug? Answer with a single number."
            ),
        },
    ],
    {
        "wtp": {"mean": 3.00, "sd": 1.50, "confidence": "approximate", "note": ""},
        "wta": {"mean": 6.50, "sd": 2.00, "confidence": "approximate", "note": ""},
    },
    effect={
        "a": "wta",
        "b": "wtp",
        "mean_diff": 3.50,
        "cohen_d": 0.80,
        "confidence": "approximate",
        "source": "Kahneman, Knetsch & Thaler (1990); Horowitz & McConnell (2002) meta-analysis",
        "note": "The WTA/WTP ratio is typically 2-3; absolute dollar values depend on the "
        "object and the subject pool.",
    },
    notes="Ownership alone shifts valuation. A population that quotes the same number in both "
    "roles is treating this as a pure valuation problem.",
)

CRT = P(
    "crt_batball",
    "Cognitive reflection — bat and ball",
    "heuristics_biases",
    "intuitive-but-wrong vs. reflective answer",
    "Frederick (2005) 'Cognitive reflection and decision making', Journal of Economic "
    "Perspectives 19:25-42",
    BIN("intuitive wrong answer (10 cents)", "reflective correct answer (5 cents)"),
    [
        {
            "condition_id": "batball",
            "label": "bat and ball",
            "prompt": (
                "A bat and a ball cost $1.10 in total. The bat costs $1.00 more than the "
                "ball. How much does the ball cost? Answer 1 if your answer is 5 cents "
                "(correct), 0 if your answer is 10 cents."
            ),
        }
    ],
    {
        "batball": {
            "mean": 0.40,
            "sd": None,
            "confidence": "approximate",
            "note": "Correct rates on the bat-and-ball item run roughly 30-50% in student "
            "samples and lower in general-population samples.",
        }
    },
    notes="Scored as a binary here so that no free-text grading is needed. If you want the "
    "full 3-item CRT, run it as three binary protocols.",
)

CONJUNCTION = P(
    "conjunction_fallacy",
    "Conjunction fallacy — Linda problem",
    "heuristics_biases",
    "representativeness vs. the conjunction rule",
    "Tversky & Kahneman (1983) 'Extensional versus intuitive reasoning', Psychological "
    "Review 90:293-315",
    BIN("choose the single (logically correct) option", "choose the conjunction"),
    [
        {
            "condition_id": "between",
            "label": "between-subjects, ranking removed",
            "prompt": (
                "Linda is 31 years old, single, outspoken and very bright. She majored in "
                "philosophy. As a student she was deeply concerned with issues of "
                "discrimination and social justice, and she participated in anti-nuclear "
                "demonstrations. Which is more probable? Option 1: Linda is a bank teller. "
                "Option 2: Linda is a bank teller and is active in the feminist movement. "
                "Answer 1 if you choose option 2 (the conjunction), 0 if you choose option 1."
            ),
        }
    ],
    {
        "between": {
            "mean": 0.80,
            "sd": None,
            "confidence": "medium",
            "note": "Roughly 80-90% choose the conjunction in the standard version; rates fall "
            "with transparent wording or within-subjects designs.",
        }
    },
    notes="Choosing the conjunction violates the conjunction rule. This is the classic case "
    "where a well-aligned model looks *worse* than a human sample.",
)


# ---------------------------------------------------------------------------
# 15. Reasoning
# ---------------------------------------------------------------------------

WASON = P(
    "wason_selection",
    "Wason selection task — abstract vs. deontic",
    "reasoning",
    "conditional-reasoning / cheater-detection",
    "Wason (1968); Cosmides & Tooby (1992) 'Cognitive adaptations for social exchange', "
    "Cognition 31:187-276",
    BIN("any other selection", "exactly the two cards that can falsify the rule"),
    [
        {
            "condition_id": "abstract",
            "label": "abstract rule (D / K / 3 / 7)",
            "prompt": (
                "Four cards are on a table, showing D, K, 3 and 7. Every card has a letter on "
                "one side and a number on the other. The rule is: if a card has a D on one "
                "side, then it has a 3 on the other side. You may turn over only the cards "
                "needed to test whether the rule is false. The correct answer is the D card "
                "and the 7 card. Answer 1 if you would turn over exactly those two cards, "
                "0 otherwise."
            ),
        },
        {
            "condition_id": "deontic",
            "label": "social-contract rule (under-age drinking)",
            "prompt": (
                "You are checking whether people are violating the rule 'if a person is "
                "drinking beer, then they must be over 21'. Four people are in front of you: "
                "one drinking beer, one drinking coke, one aged 25, one aged 16. You may check "
                "only the ones needed to find a violation. The correct answer is the beer "
                "drinker and the 16-year-old. Answer 1 if you would check exactly those two "
                "people, 0 otherwise."
            ),
        },
    ],
    {
        "abstract": {"mean": 0.15, "sd": None, "confidence": "medium-high", "note": ""},
        "deontic": {"mean": 0.70, "sd": None, "confidence": "medium", "note": ""},
    },
    effect={
        "a": "deontic",
        "b": "abstract",
        "mean_diff": 0.55,
        "cohen_d": 1.20,
        "confidence": "medium-high",
        "source": "Cosmides & Tooby (1992); Wason (1968)",
        "note": "The content effect is large: the same logical form is easy when framed as "
        "cheater detection and hard when framed abstractly.",
    },
    notes="The correct answers are stated in the prompts, so this measures whether the "
    "population's *selection* matches human selection, not whether it can solve the logic.",
)


# ---------------------------------------------------------------------------
# 16-17. Prosocial behaviour and moral judgement
# ---------------------------------------------------------------------------

BYSTANDER = P(
    "bystander_effect",
    "Bystander effect — alone vs. in a group",
    "social_influence",
    "diffusion of responsibility",
    "Darley & Latané (1968) 'Bystander intervention in emergencies', Journal of Personality "
    "and Social Psychology 8:377-383; Latané & Nida (1981)",
    BIN("do not help", "help"),
    [
        {
            "condition_id": "alone",
            "label": "participant alone",
            "prompt": (
                "You are alone in a room filling out a questionnaire. You hear someone in the "
                "next room fall, cry out in pain and groan. No one else is present. Answer 1 "
                "if you would go to help, 0 if you would not."
            ),
        },
        {
            "condition_id": "group",
            "label": "two passive others present",
            "prompt": (
                "You are in a room with two other people, filling out a questionnaire. You "
                "hear someone in the next room fall, cry out in pain and groan. The other two "
                "people look up, then go back to their questionnaires without reacting. Answer "
                "1 if you would go to help, 0 if you would not."
            ),
        },
    ],
    {
        "alone": {"mean": 0.75, "sd": None, "confidence": "medium", "note": ""},
        "group": {"mean": 0.35, "sd": None, "confidence": "medium", "note": ""},
    },
    effect={
        "a": "alone",
        "b": "group",
        "mean_diff": 0.40,
        "cohen_d": 0.85,
        "confidence": "medium",
        "source": "Darley & Latané (1968); Latané & Nida (1981)",
        "note": "Meta-analytic effect is moderate and smaller than textbook accounts suggest, "
        "but consistently present.",
    },
    notes="A simulated population that helps at the same rate in both conditions shows no "
    "diffusion of responsibility at all.",
)

TROLLEY = P(
    "trolley_dilemma",
    "Trolley problem — switch vs. footbridge",
    "moral_judgment",
    "sacrificial moral dilemmas, means vs. side-effect",
    "Foot (1967); Thomson (1985) 'The trolley problem', Yale Law Journal 94:1395-1415; "
    "large-N replications (e.g. Hauser et al. 2007)",
    BIN("do nothing", "act to save the five"),
    [
        {
            "condition_id": "switch",
            "label": "switch (impersonal)",
            "prompt": (
                "A runaway trolley is heading for five people who will be killed if it "
                "continues. You are standing next to a switch. If you pull it, the trolley "
                "diverts to a side track where it will kill one person instead. Answer 1 if "
                "you pull the switch, 0 if you do not."
            ),
        },
        {
            "condition_id": "footbridge",
            "label": "footbridge (personal)",
            "prompt": (
                "A runaway trolley is heading for five people who will be killed if it "
                "continues. You are on a footbridge above the track next to a large stranger. "
                "The only way to stop the trolley is to push the stranger off the bridge onto "
                "the track; his body will stop the trolley, killing him but saving the five. "
                "Answer 1 if you push the stranger, 0 if you do not."
            ),
        },
    ],
    {
        "switch": {"mean": 0.85, "sd": None, "confidence": "medium-high", "note": ""},
        "footbridge": {"mean": 0.25, "sd": None, "confidence": "medium-high", "note": ""},
    },
    effect={
        "a": "switch",
        "b": "footbridge",
        "mean_diff": 0.60,
        "cohen_d": 1.30,
        "confidence": "medium-high",
        "source": "Thomson (1985); Hauser et al. (2007)",
        "note": "Both dilemmas sacrifice one to save five; the personal/impersonal contrast "
        "drives a very large difference in responding.",
    },
    notes="The two dilemmas are consequentially identical. A population that gives the same "
    "answer to both is doing utilitarian arithmetic, not reproducing human moral judgement.",
)


PROTOCOLS: List[Dict[str, Any]] = [
    DICTATOR,
    ULTIMATUM_OFFER,
    ULTIMATUM_REJECT,
    PUBLIC_GOODS,
    PUBLIC_GOODS_DECAY,
    TRUST_SEND,
    TRUST_RETURN,
    PRISONERS,
    ASCH,
    ANCHORING,
    ASIAN_DISEASE,
    ENDOWMENT,
    CRT,
    CONJUNCTION,
    WASON,
    BYSTANDER,
    TROLLEY,
]


def _validate(p: Dict[str, Any]) -> None:
    ids = {c["condition_id"] for c in p["conditions"]}
    for cid in p["anchors"]:
        assert cid in ids, f"{p['protocol_id']}: anchor for unknown condition {cid}"
    e = p.get("effect")
    if e:
        assert e["a"] in ids and e["b"] in ids, f"{p['protocol_id']}: bad effect condition"
        assert e["mean_diff"] != 0, f"{p['protocol_id']}: zero effect anchor"
    assert p["response"]["min"] < p["response"]["max"]


def main() -> None:
    ap = argparse.ArgumentParser()
    ap.add_argument(
        "--out-dir", default=str(Path(__file__).resolve().parents[1] / "data")
    )
    args = ap.parse_args()

    for p in PROTOCOLS:
        _validate(p)

    out = Path(args.out_dir)
    out.mkdir(parents=True, exist_ok=True)
    with (out / "protocols.jsonl").open("w", encoding="utf-8", newline="\n") as f:
        for p in PROTOCOLS:
            f.write(json.dumps(p, ensure_ascii=False) + "\n")

    n_eff = sum(1 for p in PROTOCOLS if p["effect"])
    n_loc = sum(1 for p in PROTOCOLS if p["anchors"])
    print(f"wrote {len(PROTOCOLS)} protocols -> {out / 'protocols.jsonl'}")
    print(f"  with location/dispersion anchors: {n_loc}")
    print(f"  with a treatment-effect anchor:   {n_eff}")


if __name__ == "__main__":
    main()