Struct Policy 

pub struct Policy {

    pub auto_apply_enabled: bool,
    pub auto_apply: Vec<AutoApplyGrant>,
    pub deny: Vec<String>,
    pub severity_floors: BTreeMap<String, Severity>,
    pub telemetry: TelemetryMode,
    pub discover_objective: DiscoverObjective,
    pub outcome_evalset: Option<OutcomeEvalset>,
    pub evidence_attribution: EvidenceAttribution,
    pub cadence: Cadence,
    pub skills: SkillAuthoring,
    pub min_evidence: u32,
    pub plans: PlanAuthoring,
    pub premise_drift: bool,
    pub premise_drift_open_all: bool,
    pub near_duplicate: NearDuplicateMode,
    pub plan_replay: Option<PlanReplayPolicy>,
}

The parsed host policy. Everything default-closed — the two fields whose closed state is not the zero value (skills, min_evidence) say so in their own Default.

Fields

auto_apply_enabled: bool

Master opt-in (same posture as allow_destructive_ops: default off). Auto-apply never fires unless this is true AND a grant matches.

auto_apply: Vec<AutoApplyGrant>

Auto-apply grants (default: none).

deny: Vec<String>

Analyzer families the host disables entirely.

severity_floors: BTreeMap<String, Severity>

Per-analyzer severity floors (family → floor); combined with the file’s floors by taking the stricter of the two.

telemetry: TelemetryMode

discover_objective: DiscoverObjective

The DISCOVER scoring rule (default: the review-queue objective).

outcome_evalset: Option<OutcomeEvalset>

Measure every applicable LLM-authored proposal against this evalset after apply (default: none — authored lessons carry no metric).

evidence_attribution: EvidenceAttribution

Whether an Observation names its observer in the evidence bundle (default: named).

cadence: Cadence

When a pass is due (default: whenever it is called).

skills: SkillAuthoring

Skill authoring by the LLM proposer (default: on, two steps minimum).

min_evidence: u32

The fewest distinct evidence grains an LLM draft must cite to be offered as a change rather than an advisory finding (default 1 — a single instance may become a rule). An independent audit of 88 governed decisions found that 15 of 28 approvals had generalised one instance into standing policy; 2 is the setting that audit argues for. A draft under the threshold is still stored and still reviewable — it simply carries nothing a reviewer could apply.

plans: PlanAuthoring

Plan authoring by the LLM proposer (default: on, two steps minimum).

premise_drift: bool

The Verify gate’s second question (default: on). An applied recommendation cites the grains it was derived from; when one of them is later superseded by a DIFFERENT value, or retracted, the premise the reviewer approved no longer holds. A lesson that outlives its premise is measured harm: on PAST-Bench a rule encoding the old regime’s flag cost the governed arm 0.32 on the migration family it was learned in (crates/areev-bench/PERSIST.md). With this on, the gate records drifted and proposes the revert; a value-identical supersession (consolidation) is not drift.

premise_drift_open_all: bool

Whether EVERY cited grain must have moved before the engine withdraws an open recommendation (#317), or any one is enough.

true (the default, "all") because a finding derived from six grains of which one changed is weakened, not baseless — deciding that is a reviewer’s job. false is "any", for hosts that want the stricter sweep.

near_duplicate: NearDuplicateMode

What to do with an authored lesson that near-duplicates a live one on the same entity (default flag: queue it, marked).

plan_replay: Option<PlanReplayPolicy>

The pre-apply rehearsal gate on plan revisions (default none).

Implementations

impl Policy

pub fn from_json(s: &str) -> Result<Self>

Parse a policy JSON string. Unknown keys are rejected (fail-closed).

pub fn denies(&self, family: &str) -> bool

Is this analyzer family denied by the host?

pub fn severity_floor(&self, family: &str) -> Option<Severity>

The host severity floor for a family, if any.

pub fn grants_auto_apply( &self, family: &str, target_class: &str, severity: Severity, ) -> bool

Does a grant permit auto-applying this family to target_class at severity? Only the memory class is ever eligible.

query was eligible until definition rewrites became executable (issue #28). A grain edit changes one remembered value; a saved-query or template rewrite changes what EVERY future context contains — the blast radius is every turn from now on, not one fact. So a definition rewrite always requires a human APPROVE + APPLY with BECAUSE, and the class is excluded here by name, exactly as code/evalset are.

Trait Implementations

impl Clone for Policy

fn clone(&self) -> Self

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for Policy

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for Policy

fn default() -> Self

Returns the “default value” for a type.

impl<'de> Deserialize<'de> for Policy

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Serialize for Policy

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.