car-verify 0.35.0

//! Code World Models — trajectory-validated effect models (Slice 1).
//!
//! Applies *Code World Models for General Game Playing* (arXiv 2510.04542)
//! to CAR. The paper's recipe: have an LLM emit an executable world model
//! (`apply(state, action) -> state'`) instead of acting as the policy, then
//! **unit-test that model against recorded trajectories** and **feed failing
//! cases back** until transition accuracy reaches 1.0. See
//! `docs/proposals/code-world-models.md`.
//!
//! This module is Slice 1 of that proposal: the *verification harness* and
//! the *self-repair loop*, with no change to the live execution path.
//!
//! - [`Transition`] is the paper's unit-test record:
//!   `(state_before, action, state_after)`. The bridge that rebuilds these
//!   from a `car_eventlog` JSONL tail — the supervised signal CAR already
//!   records but never used to validate a model — lives in
//!   `car_ffi_common::cwm` (which already depends on the event log), keeping
//!   this crate dependency-light.
//! - [`score`] / [`score_predictions`] compute **transition accuracy** (the
//!   paper's central metric) over held-out transitions, with structured
//!   per-case failures so a repair agent can act on them.
//! - [`synthesize_cwm`] is the prompt→generate→run→repair loop, generic over
//!   the model (`generate`) and the code runner (`run`) — exactly the
//!   injection pattern `car_builder::build_workflow` uses for inference, so
//!   neither inference nor a code sandbox is a dependency of this crate.
//!
//! Nothing here executes real tools or generated code itself: a CWM is a
//! *predictor* for planning/pre-flight checks (Slices 2–3), never a
//! substitute for running the tool. Execution of candidate code is the
//! caller's responsibility (e.g. a Python sandbox), injected as a closure.

use crate::{apply_action_effects, StaticState};
use car_ir::{build_dag, Action, ActionProposal, ActionType};
use serde::{Deserialize, Serialize};
use serde_json::Value;
use std::collections::HashMap;

/// A state map — the same shape `car_verify::simulate` returns.
pub type State = HashMap<String, Value>;

/// One recorded transition: the paper's `(state_before, action, state_after)`
/// unit-test record. `action` is an arbitrary JSON value (a full action
/// object with `tool`/`parameters`, or a bare id when that's all the log
/// resolves) so a model has whatever the trajectory captured to predict from.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct Transition {
    pub state_before: State,
    pub action: Value,
    pub state_after: State,
}

/// A single mismatch found by scoring — enough for a human or a repair agent
/// to act on (the paper feeds the analogous failing case + stack trace back
/// to the model).
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct Failure {
    /// Index of the failing transition in the scored slice.
    pub index: usize,
    /// The action that was predicted from.
    pub action: Value,
    /// The recorded ground-truth post-state.
    pub expected: State,
    /// What the model predicted (absent when the predictor errored).
    #[serde(skip_serializing_if = "Option::is_none")]
    pub predicted: Option<State>,
    /// The predictor's error/stack trace, when running the model threw rather
    /// than returning a state (the paper's "feed the stack trace back").
    #[serde(skip_serializing_if = "Option::is_none")]
    pub error: Option<String>,
}

/// The result of scoring a model against a set of transitions — the paper's
/// transition-accuracy report.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct ScoreReport {
    /// Transitions scored (the sample size — report it so consumers don't
    /// trust an accuracy of 1.0 computed over three examples).
    pub total: usize,
    /// Transitions the model reproduced exactly.
    pub correct: usize,
    /// Transitions where the predictor errored instead of returning a state.
    pub errored: usize,
    /// `correct / total` (the paper's transition accuracy); `0.0` when empty.
    pub accuracy: f64,
    /// Per-case mismatches (correct cases are omitted).
    pub failures: Vec<Failure>,
}

impl ScoreReport {
    /// True when every scored transition was reproduced exactly (the paper's
    /// refinement target). `false` for an empty set — nothing was proven.
    pub fn is_perfect(&self) -> bool {
        self.total > 0 && self.correct == self.total
    }
}

/// Numeric-aware deep equality. Mirrors the rationale in
/// [`crate::transaction`]'s `values_equal` (JSON that round-trips through a
/// system which floatifies integers must not read as a mismatch), but
/// recurses through arrays and objects so a *nested* `1` vs `1.0` — common in
/// a state map — compares equal too.
fn values_equal(a: &Value, b: &Value) -> bool {
    match (a, b) {
        (Value::Number(x), Value::Number(y)) => match (x.as_f64(), y.as_f64()) {
            (Some(fx), Some(fy)) => fx == fy,
            _ => x == y,
        },
        (Value::Array(xs), Value::Array(ys)) => {
            xs.len() == ys.len() && xs.iter().zip(ys).all(|(x, y)| values_equal(x, y))
        }
        (Value::Object(xs), Value::Object(ys)) => {
            xs.len() == ys.len()
                && xs
                    .iter()
                    .all(|(k, x)| ys.get(k).is_some_and(|y| values_equal(x, y)))
        }
        _ => a == b,
    }
}

/// Two state maps are equal when they have the same keys with numeric-aware
/// equal values.
fn states_equal(a: &State, b: &State) -> bool {
    a.len() == b.len()
        && a.iter()
            .all(|(k, av)| b.get(k).is_some_and(|bv| values_equal(av, bv)))
}

/// Score a predictor against `transitions` — the paper's transition-accuracy
/// metric. `predict(state_before, action)` returns the model's predicted
/// post-state, or `Err(stack_trace)` if running it threw. Generic over the
/// predictor so the same metric serves an in-process model, a sandboxed
/// Python CWM, or a test fake.
pub fn score<P>(transitions: &[Transition], mut predict: P) -> ScoreReport
where
    P: FnMut(&State, &Value) -> Result<State, String>,
{
    let mut correct = 0;
    let mut errored = 0;
    let mut failures = Vec::new();
    for (index, t) in transitions.iter().enumerate() {
        match predict(&t.state_before, &t.action) {
            Ok(predicted) => {
                if states_equal(&predicted, &t.state_after) {
                    correct += 1;
                } else {
                    failures.push(Failure {
                        index,
                        action: t.action.clone(),
                        expected: t.state_after.clone(),
                        predicted: Some(predicted),
                        error: None,
                    });
                }
            }
            Err(e) => {
                errored += 1;
                failures.push(Failure {
                    index,
                    action: t.action.clone(),
                    expected: t.state_after.clone(),
                    predicted: None,
                    error: Some(e),
                });
            }
        }
    }
    let total = transitions.len();
    ScoreReport {
        total,
        correct,
        errored,
        accuracy: if total == 0 {
            0.0
        } else {
            correct as f64 / total as f64
        },
        failures,
    }
}

/// Score *precomputed* predictions against `transitions`, aligned by index.
/// The stateless path for callers that ran the generated code themselves
/// (e.g. in a sandbox) and just want the paper's accuracy metric + structured
/// failures. `predictions[i]` is the model's post-state for `transitions[i]`,
/// or `Err(stack_trace)`. A length mismatch is an error rather than a silent
/// truncation.
pub fn score_predictions(
    transitions: &[Transition],
    predictions: &[Result<State, String>],
) -> Result<ScoreReport, String> {
    if transitions.len() != predictions.len() {
        return Err(format!(
            "transitions ({}) and predictions ({}) differ in length",
            transitions.len(),
            predictions.len()
        ));
    }
    let mut iter = predictions.iter();
    Ok(score(transitions, |_, _| {
        iter.next().expect("length checked above").clone()
    }))
}

/// What `synthesize_cwm` returns. On success `code` holds the model that
/// reached perfect train accuracy and `test_accuracy` its held-out score; on
/// failure it holds the best attempt with its measured accuracy, honest about
/// not reaching 1.0 (the paper's Gin Rummy at ~0.74) — mirrors
/// `car_builder::BuildResult`.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct CwmResult {
    /// The best candidate code produced (`None` if nothing ever parsed/ran).
    pub code: Option<String>,
    /// Whether `code` reached 1.0 transition accuracy on the train split.
    pub perfect: bool,
    /// Best train-split transition accuracy seen.
    pub train_accuracy: f64,
    /// Held-out test-split accuracy — only measured for a perfect model.
    #[serde(skip_serializing_if = "Option::is_none")]
    pub test_accuracy: Option<f64>,
    /// Attempts made (1..=max_attempts).
    pub attempts: u32,
    /// Remaining failures from the best attempt (empty when perfect).
    pub failures: Vec<Failure>,
}

/// Inputs to a synthesis run.
pub struct CwmRequest {
    /// Natural-language rules and/or the tool schema the model compiles into a
    /// world model — the paper's "rule specification".
    pub spec: String,
    /// Transitions to fit against (train split).
    pub train: Vec<Transition>,
    /// Held-out transitions, scored only once a perfect model is found.
    pub test: Vec<Transition>,
    /// Maximum generate→run→repair attempts (clamped to ≥ 1).
    pub max_attempts: u32,
}

/// Build the per-attempt prompt: spec + a sample of transitions to fit + the
/// failing cases from the previous attempt (the repair signal). Kept small
/// and provider-agnostic; the actual model call is injected.
fn build_prompt(spec: &str, train: &[Transition], prior: &[Failure]) -> String {
    let mut p = String::new();
    p.push_str(
        "Write a deterministic world model as a single function `apply(state, action)` \
         returning the next state. Return ONLY the code.\n\n",
    );
    p.push_str("# Rules / tool schema\n");
    p.push_str(spec);
    p.push_str("\n\n# Transitions to reproduce exactly\n");
    // A bounded sample keeps the prompt small; the run loop scores against all.
    for t in train.iter().take(8) {
        if let Ok(line) = serde_json::to_string(t) {
            p.push_str(&line);
            p.push('\n');
        }
    }
    if !prior.is_empty() {
        p.push_str("\n# Your previous model failed these cases — fix them:\n");
        for f in prior.iter().take(8) {
            if let Ok(line) = serde_json::to_string(f) {
                p.push_str(&line);
                p.push('\n');
            }
        }
    }
    p
}

/// Synthesize a Code World Model by the paper's loop: generate code, run it
/// over the train transitions, and on imperfect accuracy feed the failing
/// cases back and retry — up to `max_attempts`, stopping early on a perfect
/// model and then scoring the held-out split.
///
/// `generate(prompt) -> code` is the model call (async, fallible — a transport
/// error is recorded but not treated as a repairable model mistake, matching
/// `build_workflow`). `run(code, state_before, action) -> state_after` runs a
/// candidate model on one transition, returning its prediction or an error
/// string (a stack trace) that becomes repair feedback. Neither inference nor
/// a code sandbox is a dependency of this crate; both are injected.
pub async fn synthesize_cwm<G, GFut, R>(generate: G, mut run: R, req: &CwmRequest) -> CwmResult
where
    G: Fn(String) -> GFut,
    GFut: std::future::Future<Output = Result<String, String>>,
    R: FnMut(&str, &State, &Value) -> Result<State, String>,
{
    let max = req.max_attempts.max(1);
    let mut prior: Vec<Failure> = Vec::new();
    let mut best: Option<(String, ScoreReport)> = None;
    let mut attempts = 0;

    for attempt in 1..=max {
        attempts = attempt;
        let prompt = build_prompt(&req.spec, &req.train, &prior);
        let code = match generate(prompt).await {
            Ok(c) => c,
            // Transport/model error — not something the model can "repair".
            // Re-send the same prompt next attempt rather than a fake hint.
            Err(_) => continue,
        };

        let report = score(&req.train, |s, a| run(&code, s, a));

        if report.is_perfect() {
            let test = score(&req.test, |s, a| run(&code, s, a));
            return CwmResult {
                code: Some(code),
                perfect: true,
                train_accuracy: report.accuracy,
                test_accuracy: Some(test.accuracy),
                attempts,
                failures: Vec::new(),
            };
        }

        prior = report.failures.clone();
        let better = best
            .as_ref()
            .map(|(_, b)| report.accuracy > b.accuracy)
            .unwrap_or(true);
        if better {
            best = Some((code, report));
        }
    }

    match best {
        Some((code, report)) => CwmResult {
            code: Some(code),
            perfect: false,
            train_accuracy: report.accuracy,
            test_accuracy: None,
            attempts,
            failures: report.failures,
        },
        None => CwmResult {
            code: None,
            perfect: false,
            train_accuracy: 0.0,
            test_accuracy: None,
            attempts,
            failures: Vec::new(),
        },
    }
}

// --- Slice 2: predictive simulation ---------------------------------------
//
// `simulate` (car-verify) is static — it propagates each action's *declared*
// `expected_effects`. A verified Code World Model predicts what a tool *actually*
// writes. Slice 2 lets a model contribute those predictions to simulation,
// behind an accuracy gate, while keeping car-verify pure: the model is an
// injected trait object, so executing generated code stays the caller's job
// (the same split as `synthesize_cwm`'s `run` closure). See
// `docs/proposals/code-world-models.md`.

/// A predictive effect model consulted during [`simulate_with_model`]: given an
/// action and the state *before* it runs, predict the state keys it will write.
/// Returning `None` means *abstain* — the simulator falls back to the action's
/// declared `expected_effects`, i.e. the static [`crate::simulate`] behavior.
/// So a model that always abstains reproduces `simulate` exactly.
pub trait EffectModel {
    fn predict(&self, action: &Action, state_before: &State) -> Option<State>;
}

/// One action's predicted post-effect plus the measured transition accuracy of
/// the model that produced it (from [`score`] / [`synthesize_cwm`]).
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct GatedPrediction {
    /// The state keys/values the model predicts this action writes.
    pub effects: State,
    /// The producing model's measured transition accuracy in `[0,1]`.
    pub accuracy: f64,
}

/// An [`EffectModel`] backed by precomputed per-action predictions, used only
/// when the producing model's accuracy clears `min_accuracy` — the paper's
/// "plan against the model only once it's verified" gate. Predictions for
/// actions with no entry, or a sub-threshold one, abstain (fall back to static
/// effects), so an under-accurate model can never *worsen* simulation: at worst
/// it's ignored. Caller computes each prediction by running the generated code;
/// the gate lives here so it's uniform and testable.
pub struct GatedEffectModel {
    /// Keyed by `Action::id`.
    pub predictions: HashMap<String, GatedPrediction>,
    /// Minimum accuracy a prediction must have to be applied.
    pub min_accuracy: f64,
}

impl EffectModel for GatedEffectModel {
    fn predict(&self, action: &Action, _state_before: &State) -> Option<State> {
        self.predictions
            .get(&action.id)
            .filter(|p| p.accuracy >= self.min_accuracy)
            .map(|p| p.effects.clone())
    }
}

/// Simulate a proposal's final state like [`crate::simulate`], but consult
/// `model` for each action's effects in topological order: where the model
/// predicts (`Some`), apply the prediction instead of the action's declared
/// `expected_effects`; where it abstains (`None`), fall back to the static
/// effects. A `StateWrite`'s deterministic key/value is always applied (it is
/// fully known statically — not something a model needs to predict). Passing a
/// model that always abstains yields byte-identical output to `simulate`.
pub fn simulate_with_model(
    proposal: &ActionProposal,
    initial_state: Option<&State>,
    model: &dyn EffectModel,
) -> State {
    let mut state = match initial_state {
        Some(s) => StaticState::from_map(s.clone()),
        None => StaticState::new(),
    };

    for level in build_dag(&proposal.actions) {
        for idx in level {
            let action = &proposal.actions[idx];
            // Predict from the pre-action state (before this action's own write).
            let predicted = model.predict(action, &state.known);

            // Deterministic builtin write always applies.
            if action.action_type == ActionType::StateWrite {
                if let Some(key) = action.parameters.get("key").and_then(|v| v.as_str()) {
                    let value = action
                        .parameters
                        .get("value")
                        .cloned()
                        .unwrap_or(Value::Null);
                    state.set(key, value);
                }
            }

            match predicted {
                Some(delta) => {
                    for (k, v) in delta {
                        state.set(&k, v);
                    }
                }
                // No prediction → identical to `simulate`'s static effects.
                // Re-applies the StateWrite key above harmlessly (same value).
                None => apply_action_effects(action, &mut state),
            }
        }
    }
    state.known
}

#[cfg(test)]
mod tests {
    use super::*;
    use serde_json::json;

    fn st(pairs: &[(&str, Value)]) -> State {
        pairs
            .iter()
            .map(|(k, v)| (k.to_string(), v.clone()))
            .collect()
    }

    fn counter_transitions() -> Vec<Transition> {
        // A tiny world: action {"inc": n} adds n to state["count"].
        vec![
            Transition {
                state_before: st(&[("count", json!(0))]),
                action: json!({ "inc": 1 }),
                state_after: st(&[("count", json!(1))]),
            },
            Transition {
                state_before: st(&[("count", json!(1))]),
                action: json!({ "inc": 2 }),
                state_after: st(&[("count", json!(3))]),
            },
        ]
    }

    // A "model" that correctly implements the counter world.
    fn good_model(s: &State, a: &Value) -> Result<State, String> {
        let cur = s.get("count").and_then(|v| v.as_i64()).unwrap_or(0);
        let inc = a.get("inc").and_then(|v| v.as_i64()).unwrap_or(0);
        Ok(st(&[("count", json!(cur + inc))]))
    }

    #[test]
    fn perfect_model_scores_1() {
        let r = score(&counter_transitions(), good_model);
        assert_eq!(r.total, 2);
        assert_eq!(r.correct, 2);
        assert!(r.is_perfect());
        assert_eq!(r.accuracy, 1.0);
        assert!(r.failures.is_empty());
    }

    #[test]
    fn wrong_model_reports_failures() {
        // Off-by-one: always adds 1 regardless of action.
        let r = score(&counter_transitions(), |s, _| {
            let cur = s.get("count").and_then(|v| v.as_i64()).unwrap_or(0);
            Ok(st(&[("count", json!(cur + 1))]))
        });
        assert_eq!(r.correct, 1); // first case (inc:1) happens to match
        assert_eq!(r.failures.len(), 1);
        assert_eq!(r.failures[0].index, 1);
        assert!(r.failures[0].error.is_none());
        assert_eq!(r.failures[0].predicted, Some(st(&[("count", json!(2))])));
    }

    #[test]
    fn throwing_model_is_counted_as_errored() {
        let r = score(&counter_transitions(), |_, _| Err("boom".to_string()));
        assert_eq!(r.errored, 2);
        assert_eq!(r.correct, 0);
        assert_eq!(r.failures[0].error.as_deref(), Some("boom"));
    }

    #[test]
    fn numeric_equality_int_vs_float() {
        // Recorded post-state has an int; model returns a float — must match.
        let trs = vec![Transition {
            state_before: st(&[]),
            action: json!({}),
            state_after: st(&[("x", json!(2))]),
        }];
        let r = score(&trs, |_, _| Ok(st(&[("x", json!(2.0))])));
        assert!(r.is_perfect());
    }

    #[test]
    fn nested_numeric_equality() {
        let trs = vec![Transition {
            state_before: st(&[]),
            action: json!({}),
            state_after: st(&[("v", json!({ "items": [1, 2] }))]),
        }];
        let r = score(&trs, |_, _| {
            Ok(st(&[("v", json!({ "items": [1.0, 2.0] }))]))
        });
        assert!(r.is_perfect());
    }

    #[test]
    fn score_predictions_length_mismatch_errs() {
        let trs = counter_transitions();
        let preds = vec![Ok(st(&[("count", json!(1))]))]; // too short
        assert!(score_predictions(&trs, &preds).is_err());
    }

    #[test]
    fn score_predictions_matches_closure_path() {
        let trs = counter_transitions();
        let preds: Vec<Result<State, String>> =
            trs.iter().map(|t| Ok(t.state_after.clone())).collect();
        let r = score_predictions(&trs, &preds).unwrap();
        assert!(r.is_perfect());
    }

    #[test]
    fn empty_set_is_not_perfect() {
        let r = score(&[], good_model);
        assert!(!r.is_perfect());
        assert_eq!(r.accuracy, 0.0);
    }

    #[tokio::test]
    async fn synthesize_repairs_then_succeeds() {
        use std::cell::Cell;
        // First attempt returns a broken model, second returns a good one.
        let attempt = Cell::new(0u32);
        let generate = |_prompt: String| {
            let n = attempt.get();
            attempt.set(n + 1);
            async move { Ok(format!("model-v{n}")) }
        };
        // The runner interprets the code tag: v0 is broken, v1+ is correct.
        let run = |code: &str, s: &State, a: &Value| -> Result<State, String> {
            if code == "model-v0" {
                return Ok(st(&[("count", json!(999))])); // wrong
            }
            good_model(s, a)
        };
        let req = CwmRequest {
            spec: "counter world".to_string(),
            train: counter_transitions(),
            test: counter_transitions(),
            max_attempts: 5,
        };
        let res = synthesize_cwm(generate, run, &req).await;
        assert!(res.perfect, "should repair to a perfect model");
        assert_eq!(res.code.as_deref(), Some("model-v1"));
        assert_eq!(res.attempts, 2);
        assert_eq!(res.test_accuracy, Some(1.0));
    }

    #[tokio::test]
    async fn synthesize_returns_best_when_never_perfect() {
        let generate = |_p: String| async { Ok("stuck".to_string()) };
        let run = |_c: &str, s: &State, _a: &Value| -> Result<State, String> {
            // Always returns state unchanged: matches no counter transition.
            Ok(s.clone())
        };
        let req = CwmRequest {
            spec: "counter".to_string(),
            train: counter_transitions(),
            test: counter_transitions(),
            max_attempts: 3,
        };
        let res = synthesize_cwm(generate, run, &req).await;
        assert!(!res.perfect);
        assert_eq!(res.attempts, 3);
        assert!(res.test_accuracy.is_none());
        assert!(!res.failures.is_empty());
    }

    // --- Slice 2: predictive simulation tests ---

    /// A tool-call action `id` declaring `expected_effects` (built from JSON so
    /// it tracks the IR's serde defaults rather than a brittle field list).
    fn tool_action(id: &str, effects: &[(&str, Value)]) -> Action {
        let eff: serde_json::Map<String, Value> = effects
            .iter()
            .map(|(k, v)| (k.to_string(), v.clone()))
            .collect();
        serde_json::from_value(json!({
            "type": "tool_call", "id": id, "tool": "t", "expected_effects": eff
        }))
        .unwrap()
    }

    fn proposal(actions: Vec<Action>) -> ActionProposal {
        serde_json::from_value(json!({ "actions": actions })).unwrap()
    }

    /// A model that always abstains — must reproduce `simulate` exactly.
    struct Abstain;
    impl EffectModel for Abstain {
        fn predict(&self, _a: &Action, _s: &State) -> Option<State> {
            None
        }
    }

    #[test]
    fn abstaining_model_matches_static_simulate() {
        let p = proposal(vec![
            tool_action("a1", &[("x", json!(1))]),
            tool_action("a2", &[("y", json!(2))]),
        ]);
        let with_model = simulate_with_model(&p, None, &Abstain);
        let static_sim = crate::simulate(&p, None);
        assert_eq!(with_model, static_sim);
        assert_eq!(with_model.get("x"), Some(&json!(1)));
        assert_eq!(with_model.get("y"), Some(&json!(2)));
    }

    #[test]
    fn gated_model_overrides_declared_effect_when_accurate() {
        // Declared effect says x=1; the verified model predicts x=42.
        let p = proposal(vec![tool_action("a1", &[("x", json!(1))])]);
        let mut predictions = HashMap::new();
        predictions.insert(
            "a1".to_string(),
            GatedPrediction {
                effects: st(&[("x", json!(42))]),
                accuracy: 0.99,
            },
        );
        let model = GatedEffectModel {
            predictions,
            min_accuracy: 0.9,
        };
        let out = simulate_with_model(&p, None, &model);
        assert_eq!(out.get("x"), Some(&json!(42)), "prediction should win");
    }

    #[test]
    fn gated_model_abstains_below_threshold() {
        let p = proposal(vec![tool_action("a1", &[("x", json!(1))])]);
        let mut predictions = HashMap::new();
        predictions.insert(
            "a1".to_string(),
            GatedPrediction {
                effects: st(&[("x", json!(42))]),
                accuracy: 0.50, // below gate
            },
        );
        let model = GatedEffectModel {
            predictions,
            min_accuracy: 0.9,
        };
        let out = simulate_with_model(&p, None, &model);
        // Falls back to the declared effect — an under-accurate model never
        // worsens simulation.
        assert_eq!(out.get("x"), Some(&json!(1)));
    }

    #[test]
    fn gated_model_abstains_for_unknown_action() {
        let p = proposal(vec![tool_action("a1", &[("x", json!(1))])]);
        let model = GatedEffectModel {
            predictions: HashMap::new(), // no entry for a1
            min_accuracy: 0.9,
        };
        let out = simulate_with_model(&p, None, &model);
        assert_eq!(out.get("x"), Some(&json!(1)));
    }

    #[test]
    fn statewrite_deterministic_value_always_applies() {
        // A StateWrite is fully known statically; the model abstaining must
        // still produce the write.
        let sw: Action = serde_json::from_value(json!({
            "type": "state_write", "id": "w",
            "parameters": { "key": "k", "value": "v" }
        }))
        .unwrap();
        let out = simulate_with_model(&proposal(vec![sw]), None, &Abstain);
        assert_eq!(out.get("k"), Some(&json!("v")));
    }
}