harn-stdlib 0.10.201

Embedded Harn standard library source catalog
Documentation
import {
  DecisionDefinition,
  DecisionEvidence,
  DecisionNonMatch,
  DecisionPolicy,
} from "std/decision"
import { SelectiveRiskOptions, SelectiveRiskReport } from "std/eval/selective_risk"
import { LlmCallOptions } from "std/llm/options"
import { BooleanQuestion, EvaluationOutcome, EvaluationPolicy } from "std/predicate"
import { ToolCatalogEntry } from "std/tools"

/** Reviewed executable coordinates over the ordinary tool catalog. */
pub type LearnedCapability = {
  id: string,
  meaning: string,
  tool: ToolCatalogEntry,
  preflight_tool?: ToolCatalogEntry,
  package_digest: string,
  project_digest: string,
  verification_digest: string,
  minimum_probability: float,
}

pub type LearnedRequest = {request: string}

pub type LearnedExecutionEvidence = {kind: "not_run"} \
  | {kind: "failed", receipt: string, workspace_digest: string} \
  | {kind: "passed", receipt: string, workspace_digest: string}

/** One end-to-end observation, including every committed tool call's outcome. */
pub type LearnedTrial = {
  id: string,
  expected: string,
  selected: string,
  confidence: float,
  abstained: bool,
  execution: LearnedExecutionEvidence,
}

/** Assessment is advisory; commitment begins after argument and preflight checks. */
pub type LearnedObservation = {
  id: string,
  expected: string,
  assessment: LearnedAssessment,
  binding?: EvaluationOutcome,
  disposition: {kind: "abstained", reason: string} \
    | {kind: "committed", capability_id: string, decision_receipt: string, binding_receipt: string},
  execution: LearnedExecutionEvidence,
}

pub type LearnedPreflight = {kind: "ready"} | {kind: "declined", reason: string}

pub type LearnedBindingState = {
  request: string,
  capability_id: string,
  capability_meaning: string,
  input_schema_json: string,
  arguments_json: string,
}

pub type LearnedBindingRequest = {
  site_id: string,
  state: LearnedBindingState,
  questions: {faithful: BooleanQuestion},
  policy: EvaluationPolicy,
}

pub type LearnedIdentity = {
  definition_digest: string,
  capabilities_digest: string,
  evaluator_revision: string,
  binding_digest: string,
  runtime_digest: string,
  grader_digest: string,
  population: string,
}

pub type LearnedQualification = {
  contract: "harn.learned_qualification.v1",
  identity: LearnedIdentity,
  trials_digest: string,
  report: SelectiveRiskReport,
  executed: int,
  failed: int,
  negatives: int,
}

pub type LearnedAssessment = {
  kind: "selected",
  capability: LearnedCapability,
  evidence: DecisionEvidence,
} \
  | {kind: "unqualified", reason: string} \
  | DecisionNonMatch

pub type LearnedDispatcherOptions = {
  id: string,
  instructions: string,
  capabilities: list<LearnedCapability>,
  policy: DecisionPolicy,
  evaluator_revision: string,
  binding: LlmCallOptions,
  runtime_digest: string,
  grader_digest: string,
  population: string,
}

pub type LearnedDispatcher = {
  definition: DecisionDefinition,
  identity: LearnedIdentity,
  binding: LlmCallOptions,
  probe: fn(HarnessLlm, LearnedRequest) -> LearnedAssessment,
  assess: fn(HarnessLlm, LearnedRequest, LearnedQualification?) -> LearnedAssessment,
}

pub type LearnedQualificationOptions = {risk: SelectiveRiskOptions}