openpond 0.2.46 → 0.2.48

package/dist/chunks/{chunk-KE5RH6JA.js → chunk-WITSXYEW.js}

@@ -1,5 +1,5 @@
 import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
-import{n as z,o as V,p as G,q}from"./chunk-XLDW5NN3.js";import{a as A,f as v,g as k,h as p,i as f}from"./chunk-3M53YXEI.js";import{a}from"./chunk-FTQLQJQV.js";import{Buffer as Y}from"node:buffer";import{readdirSync as H,readFileSync as Q,statSync as F}from"node:fs";import b from"node:path";var K=500,Z=8*1024*1024,W=8*1024*1024;async function me(e,t){if(!t.sandboxRuntime&&!t.runtimeId)return{sandbox:await X(e,await e.create(t.sandbox,{async:!0}))};let n=t.runtimeId?void 0:await e.createSandboxRuntime(t.sandboxRuntime),r=t.runtimeId??n.id,o=await e.createSandboxRuntimeSandbox(r,t.sandbox);return{sandbox:await X(e,o.sandbox),runtime:o.runtime??n}}a(me,"createSandboxFromPlan");async function X(e,t){if(t.state==="running"||t.state==="stopped")return t;if(t.state==="error")throw new Error(`sandbox create failed: ${t.id}
+import{n as z,o as V,p as G,q}from"./chunk-N3DEA6PG.js";import{a as A,f as v,g as k,h as p,i as f}from"./chunk-3M53YXEI.js";import{a}from"./chunk-FTQLQJQV.js";import{Buffer as Y}from"node:buffer";import{readdirSync as H,readFileSync as Q,statSync as F}from"node:fs";import b from"node:path";var K=500,Z=8*1024*1024,W=8*1024*1024;async function me(e,t){if(!t.sandboxRuntime&&!t.runtimeId)return{sandbox:await X(e,await e.create(t.sandbox,{async:!0}))};let n=t.runtimeId?void 0:await e.createSandboxRuntime(t.sandboxRuntime),r=t.runtimeId??n.id,o=await e.createSandboxRuntimeSandbox(r,t.sandbox);return{sandbox:await X(e,o.sandbox),runtime:o.runtime??n}}a(me,"createSandboxFromPlan");async function X(e,t){if(t.state==="running"||t.state==="stopped")return t;if(t.state==="error")throw new Error(`sandbox create failed: ${t.id}
 ${t.logs.join(`
 `)}`);let n=12*6e4,r=3e3,o=Date.now()+n,i=t;for(;Date.now()<o;){if(await new Promise(s=>setTimeout(s,r)),i=await e.get(t.id),i.state==="running"||i.state==="stopped")return i;if(i.state==="error")throw new Error(`sandbox create failed: ${i.id}
 ${i.logs.join(`

package/dist/chunks/chunk-WV2APYWG.js

@@ -0,0 +1,2 @@
+import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
+import{C as Qe,Ga as Q,Ha as nt,Ia as at,Ja as rt,Ka as st,La as it,Ma as X,R as Xe,T as Ye,V as he,W as ge,X as Re,a as N,b as le,c as We,j as de,k as Ge,l as ue,m as fe,oa as et,qa as tt,ta as ve,ua as we,v as $,y as Ze,z as pe}from"./chunk-PJM2G7SG.js";import{h as Ue}from"./chunk-SDEDAHV6.js";import{b as v,c as R,d as _,e as C,f as g,h as J,i as j,j as me,l as Ke,m as o,n as V,o as Z}from"./chunk-NNF6KCHD.js";import{b as t}from"./chunk-ES2ISFJC.js";import{a as m}from"./chunk-FTQLQJQV.js";var Me=t.object({adapterId:v,placement:t.enum(["local","remote","colocated","provider_native"]),runtimeVersion:t.string().trim().min(1).max(200),capabilityReceipt:R}).strict(),je=t.object({maxTurns:t.number().int().positive().max(1e5),timeoutMs:t.number().int().positive().max(864e5),maxOutputBytes:t.number().int().positive().max(25e7),maximumSpendUsd:t.number().nonnegative().nullable()}).strict(),ot=t.object({schemaVersion:t.literal("openpond.runManifest.v1"),id:v,harnessRelease:g,tasksetRelease:g,model:Z,runtimeTarget:Me,limits:je,approval:t.object({id:v,contentHash:R}).strict().nullable(),createdAt:_,metadata:C}).strict(),K=ot.extend({contentHash:R}).strict(),Pe=t.object({schemaVersion:t.literal("openpond.attemptReceipt.v1"),id:v,runManifest:g,taskId:v,seed:t.string().trim().min(1).max(500),terminal:t.boolean(),failureClass:me.nullable(),outputHash:R.nullable(),traceHash:R,artifactRefs:t.array(j).max(1e5),graderEvidenceRefs:t.array(j).max(1e4),startedAt:_,completedAt:_,latencyMs:t.number().int().nonnegative(),costUsd:t.number().nonnegative().nullable(),legacyAttemptRef:v.nullable().default(null),metadata:C}).strict(),be=Pe.extend({contentHash:R}).strict(),ct=t.object({schemaVersion:t.literal("openpond.evaluationResult.v1"),id:v,runManifest:g,harnessRelease:g,tasksetRelease:g,model:Z,receiptRefs:t.array(g).min(1).max(1e6),attemptCount:t.number().int().positive(),rewardEligibleCount:t.number().int().nonnegative(),terminalCount:t.number().int().nonnegative(),meanScore:t.number().min(0).max(1).nullable(),failureCounts:t.record(me,t.number().int().nonnegative()),authoredMetric:ue.optional(),metadata:C}).strict(),cn=ct.extend({contentHash:R}).strict(),mn=t.object({schemaVersion:t.literal("openpond.harnessCompatibility.v1"),id:v,baseHarnessRelease:g,candidateHarnessRelease:g,tasksetRelease:g,environmentHash:R,toolContractHash:R,policyHash:R,graderInterfaceHash:R,metadata:C}).strict(),ln=mn.extend({contentHash:R}).strict();function Ra(e){let n=ot.parse(e);return K.parse({...n,contentHash:o(n)})}m(Ra,"createRunManifest");function va(e){let n=Pe.parse(e);return be.parse({...n,contentHash:o(n)})}m(va,"createAttemptReceipt");function D(e){let n=be.safeParse(e);if(!n.success)return!1;let{contentHash:a,...s}=n.data;return o(Pe.parse(s))===a}m(D,"verifyAttemptReceipt");function mt(e){if(!e.receipts.length)throw new Error("An evaluation requires at least one attempt receipt.");for(let r of e.receipts){if(r.runManifest.id!==e.manifest.id||r.runManifest.contentHash!==e.manifest.contentHash)throw new Error(`Attempt receipt ${r.id} belongs to a different Run Manifest.`);if(!D(r))throw new Error(`Attempt receipt ${r.id} has an invalid content hash.`)}if(e.authoredMetric){fe(e.authoredMetric);let r=e.authoredMetric;if(r.runManifest.id!==e.manifest.id||r.runManifest.contentHash!==e.manifest.contentHash||o(r.tasksetRelease)!==o(e.manifest.tasksetRelease)||o(r.receiptRefs)!==o(e.receipts.map(({id:i,contentHash:u})=>({id:i,contentHash:u}))))throw new Error("Authored metric does not describe this evaluation's exact manifest and receipts.")}let n=e.receipts.flatMap(r=>typeof r.metadata.score=="number"&&Number.isFinite(r.metadata.score)?[r.metadata.score]:[]),a=Object.fromEntries(me.options.map(r=>[r,e.receipts.filter(i=>i.failureClass===r).length])),s=ct.parse({schemaVersion:"openpond.evaluationResult.v1",id:e.id,runManifest:{id:e.manifest.id,contentHash:e.manifest.contentHash},harnessRelease:e.manifest.harnessRelease,tasksetRelease:e.manifest.tasksetRelease,model:e.manifest.model,receiptRefs:e.receipts.map(r=>({id:r.id,contentHash:r.contentHash})),attemptCount:e.receipts.length,rewardEligibleCount:e.receipts.filter(r=>r.metadata.rewardEligible===!0&&r.failureClass!=="infrastructure_failure").length,terminalCount:e.receipts.filter(r=>r.terminal).length,meanScore:n.length?n.reduce((r,i)=>r+i,0)/n.length:null,failureCounts:a,...e.authoredMetric?{authoredMetric:e.authoredMetric}:{},metadata:e.metadata??{}});return cn.parse({...s,contentHash:o(s)})}m(mt,"aggregateEvaluationReceipts");function Te(e,n,a){if(e.tasksetRelease.contentHash!==n.tasksetRelease.contentHash)throw new Error("Evaluation runs use different Taskset Releases.");if(e.harnessRelease.contentHash===n.harnessRelease.contentHash)return;if(!a)throw new Error("Evaluation runs use different Harness Releases without a compatibility receipt.");let s=ln.parse(a);if(V(s,"Harness compatibility receipt"),s.baseHarnessRelease.contentHash!==e.harnessRelease.contentHash||s.candidateHarnessRelease.contentHash!==n.harnessRelease.contentHash||s.tasksetRelease.contentHash!==e.tasksetRelease.contentHash)throw new Error("Harness compatibility receipt does not match the compared runs.")}m(Te,"assertComparableRunManifests");var lt=t.object({schemaVersion:t.literal("openpond.humanResultView.v1"),executionId:t.string().min(1),executionHash:t.string().regex(/^[a-f0-9]{64}$/),selection:t.array($).max(1e4),coverage:t.object({pending:t.number().int().nonnegative(),submitted:t.number().int().nonnegative(),accepted:t.number().int().nonnegative(),rejected:t.number().int().nonnegative(),cancelled:t.number().int().nonnegative()}).strict(),results:t.array(t.object({review:$,decision:$.nullable(),attemptIds:t.array(t.string().min(1)).min(1).max(2),mode:t.enum(["individual","pairwise"]),origin:t.enum(["hosted_retained","recorded_evidence","owner_attested_local","native_local"]),status:t.enum(["pending","submitted","accepted","rejected","cancelled"]),grader:$,answers:t.array(Ze).max(100)}).strict()).max(1e4)}).strict(),xe=lt.extend({contentHash:t.string().regex(/^[a-f0-9]{64}$/)}).strict();function dt(e){let n=e.reviews.map(r=>{let i=pe.parse(r),{contentHash:u,...l}=i;if(o(l)!==u||!i.evidence.attempts.some(d=>d.experimentId===e.executionId))throw new N("human_result_selection_mismatch",409);return i});if(new Set(n.map(r=>r.id)).size!==n.length)throw new N("human_result_duplicate_review",422);let a=n.map(r=>{let i=r.decisions.at(-1);return{review:{id:r.id,revision:r.revision,contentHash:r.contentHash},decision:i?{id:i.id,revision:i.revision,contentHash:i.contentHash}:null,attemptIds:r.evidence.attempts.map(u=>u.attemptId),mode:r.form.mode,origin:r.evidence.attempts.some(u=>u.localSource?.kind==="owner_attested_local")?"owner_attested_local":r.evidence.attempts.some(u=>u.localSource?.kind==="native_local")?"native_local":r.evidence.attempts.some(u=>u.recordedSource)?"recorded_evidence":"hosted_retained",status:r.status==="accepted"?"accepted":r.status==="rejected"?"rejected":r.status==="cancelled"?"cancelled":r.status==="submitted"?"submitted":"pending",grader:r.evidence.grader,answers:r.status==="accepted"&&i?.outcome==="accept"?i.answers:[]}}),s={pending:0,submitted:0,accepted:0,rejected:0,cancelled:0};for(let r of a)s[r.status]++;return xe.parse(Qe(lt.parse({schemaVersion:"openpond.humanResultView.v1",executionId:e.executionId,executionHash:e.executionHash,selection:a.map(r=>r.review),coverage:s,results:a})))}m(dt,"createHumanResultView");var dn=t.discriminatedUnion("kind",[t.object({kind:t.literal("model"),model:Z,configurationHash:R}).strict(),t.object({kind:t.literal("fixture")}).strict()]),un=t.discriminatedUnion("kind",[t.object({kind:t.literal("taskset"),environmentRelease:g,verifierSetRelease:g,policyHash:R}).strict(),t.object({kind:t.literal("harness"),harnessRelease:g}).strict()]),fn=t.object({receiptId:v,taskId:v,seed:t.string().trim().min(1).max(500),fixtureId:v.nullable()}).strict(),ut=t.object({schemaVersion:t.literal("openpond.tasksetRunManifest.v1"),id:v,tasksetRelease:g,packageHash:R,execution:un,profileEvaluation:we.optional(),policy:dn,gradingRole:t.literal("evaluation"),metricPolicy:de,population:t.array(fn).min(1).max(1e5),runtimeTarget:Me,limits:je,createdAt:_,metadata:C}).strict().superRefine((e,n)=>{new Set(e.population.map(a=>a.receiptId)).size!==e.population.length&&n.addIssue({code:"custom",path:["population"],message:"Run receipt identities must be unique."}),e.population.some(a=>a.fixtureId!==null!=(e.policy.kind==="fixture"))&&n.addIssue({code:"custom",path:["population"],message:"Only fixture runs require a fixture identity for every member."}),e.profileEvaluation&&(e.execution.kind!=="harness"||e.execution.harnessRelease.id!==e.profileEvaluation.harnessRelease.id||e.execution.harnessRelease.contentHash!==e.profileEvaluation.harnessRelease.contentHash)&&n.addIssue({code:"custom",path:["profileEvaluation"],message:"Profile evaluation must execute its exact bound Harness release."})}),z=ut.safeExtend({contentHash:R});function Ta(e){let n=ut.parse(e);return z.parse({...n,contentHash:o(n)})}m(Ta,"createTasksetRunManifest");function pn(e){return e.metrics??{schemaVersion:"openpond.tasksetMetricPolicy.v1",primaryMetric:"mean_score",aggregation:"mean_score",missingReward:"exclude",customAggregator:null}}m(pn,"tasksetRunMetricPolicy");function Oe(e,n){if(z.parse(e),V(e,"Taskset run manifest"),Re(n),e.tasksetRelease.id!==n.id||e.tasksetRelease.contentHash!==n.contentHash)throw new Error("Taskset run differs from its pinned release.");if(o(e.metricPolicy)!==o(pn(n)))throw new Error("Taskset run metric differs from its pinned release.");let a=new Set(n.tasks.map(s=>s.id));if(e.population.some(s=>!a.has(s.taskId)))throw new Error("Taskset run population contains an unknown task.");if(e.execution.kind==="taskset"&&(!n.environmentRelease||!n.verifierSetRelease||o(e.execution.environmentRelease)!==o(n.environmentRelease)||o(e.execution.verifierSetRelease)!==o(n.verifierSetRelease)||e.execution.policyHash!==o(n.policy)))throw new Error("Taskset-owned execution differs from its pinned release.")}m(Oe,"assertTasksetRunRelease");function Oa(e,n,a){Oe(e,n);let s=ve.parse(a),r=e.profileEvaluation;if(!r||e.execution.kind!=="harness"||e.policy.kind!=="model")throw new Error("Profile evaluation requires fresh model execution through its bound Harness.");if(r.catalogHash!==o(s))throw new Error("Profile evaluation catalog differs from its pinned source.");let i=s.definitions.find(c=>c.id===r.definitionId);if(!i||r.definitionHash!==o(i)||o(r.target)!==o(i.target))throw new Error("Profile evaluation definition differs from its pinned source.");if(i.tasksetRelease.id!==n.id||i.tasksetRelease.contentHash!==n.contentHash)throw new Error("Profile evaluation Taskset differs from its pinned definition.");if(i.split==="train")throw new Error("Profile evaluation cannot execute training cases.");let u=new Map(n.tasks.map(c=>[c.id,c]));if(i.taskIds.some(c=>u.get(c)?.split!==i.split))throw new Error("Profile evaluation case is absent from its declared Taskset split.");if(r.externalDatasetBinding){let c=et(r.externalDatasetBinding);if(r.definitionId!==tt(c)||c.dataset.id!==n.id||c.dataset.revision!==n.revision||c.dataset.contentHash!==n.contentHash||c.packageHash!==e.packageHash||c.profileId!==r.profileId||o(c.target)!==o(r.target)||c.split!==i.split||c.environmentHash!==o(n.environment)||c.population.length!==e.population.length||c.population.some((f,p)=>o(f)!==o({taskId:e.population[p].taskId,seed:e.population[p].seed,fixtureId:e.population[p].fixtureId})))throw new Error("Profile external Dataset execution differs from its exact admitted recipe.");return}let l=new Set(i.taskIds.flatMap(c=>i.seeds.map(f=>JSON.stringify([c,f])))),d=new Set(e.population.map(({taskId:c,seed:f})=>JSON.stringify([c,f])));if(e.population.some(c=>c.fixtureId!==null)||d.size!==l.size||e.population.length!==l.size||[...d].some(c=>!l.has(c)))throw new Error("Profile evaluation population differs from its declared cases and seeds.")}m(Oa,"assertProfileEvaluationRunAdmission");function Ve(e,n){z.parse(e),V(e,"Taskset run manifest");let a=new Map(n.map(s=>[s.id,s]));if(a.size!==n.length||n.length!==e.population.length)throw new Error("Run receipts differ from the complete admitted population.");return e.population.map(s=>{let r=a.get(s.receiptId);if(!r||!D(r)||r.runManifest.id!==e.id||r.runManifest.contentHash!==e.contentHash||r.taskId!==s.taskId||r.seed!==s.seed)throw new Error("Run receipt differs from its admitted member or manifest.");if(r.metadata.gradingRole!=="evaluation")throw new Error("Taskset run requires evaluation-role evidence.");if(!r.terminal&&!["infrastructure_failure","timeout","cancelled"].includes(r.failureClass??""))throw new Error("Taskset run member has no terminal accounting outcome.");return r})}m(Ve,"orderTasksetRunReceipts");var ft=t.object({schemaVersion:t.literal("openpond.profileEvaluationComparison.v1"),id:v,tasksetRelease:g,populationHash:R,environmentHash:R,metricPolicyHash:R,limitsHash:R,members:t.array(t.object({runManifest:g,metricResultHash:R,source:we,policy:z.shape.policy,score:t.number().min(0).max(1).nullable()}).strict()).min(2).max(100),createdAt:_}).strict(),hn=ft.extend({contentHash:R}).strict();function La(e){if(e.members.length<2)throw new Error("A comparison requires at least two runs.");let n=e.members.map(({manifest:d,result:c})=>{if(z.parse(d),V(d,"Taskset run manifest"),fe(c),!d.profileEvaluation||d.execution.kind!=="harness")throw new Error("Profile comparison requires released Profile evaluation runs.");if(c.runManifest.id!==d.id||c.runManifest.contentHash!==d.contentHash||c.tasksetRelease.id!==d.tasksetRelease.id||c.tasksetRelease.contentHash!==d.tasksetRelease.contentHash||c.policyHash!==o(d.metricPolicy))throw new Error("Profile comparison metric differs from its run manifest.");return{manifest:d,result:c,source:d.profileEvaluation}}),a=n[0],s=o(a.manifest.population.map(({taskId:d,seed:c,fixtureId:f})=>({taskId:d,seed:c,fixtureId:f}))),r=o(a.manifest.metricPolicy),i=o(a.manifest.limits),u=new Set;for(let d of n){if(u.has(d.manifest.id))throw new Error("Profile comparison repeats a run.");if(u.add(d.manifest.id),d.manifest.tasksetRelease.id!==a.manifest.tasksetRelease.id||d.manifest.tasksetRelease.contentHash!==a.manifest.tasksetRelease.contentHash||o(d.manifest.population.map(({taskId:c,seed:f,fixtureId:p})=>({taskId:c,seed:f,fixtureId:p})))!==s||d.source.environmentHash!==a.source.environmentHash||o(d.manifest.metricPolicy)!==r||o(d.manifest.limits)!==i||d.manifest.runtimeTarget.adapterId!==a.manifest.runtimeTarget.adapterId||d.manifest.runtimeTarget.placement!==a.manifest.runtimeTarget.placement||d.manifest.runtimeTarget.runtimeVersion!==a.manifest.runtimeTarget.runtimeVersion)throw new Error("Profile comparison runs differ in frozen tasks, environment, grading, limits, or runtime.")}let l=ft.parse({schemaVersion:"openpond.profileEvaluationComparison.v1",id:e.id,tasksetRelease:a.manifest.tasksetRelease,populationHash:s,environmentHash:a.source.environmentHash,metricPolicyHash:r,limitsHash:i,members:n.map(({manifest:d,result:c,source:f})=>({runManifest:{id:d.id,contentHash:d.contentHash},metricResultHash:c.contentHash,source:f,policy:d.policy,score:c.value})),createdAt:e.createdAt});return hn.parse({...l,contentHash:o(l)})}m(La,"createProfileEvaluationComparison");var pt=t.object({schemaVersion:t.literal("openpond.profileEvaluationSuiteRun.v1"),id:v,suiteId:v,suiteHash:R,catalogHash:R,profileId:v,sourceRevision:t.string().trim().min(1).max(240),harnessRelease:g,members:t.array(t.object({definitionId:v,runManifest:g,runHash:R,passed:t.boolean(),score:t.number().min(0).max(1).nullable()}).strict()).min(1).max(1e3),passed:t.boolean(),createdAt:_,completedAt:_}).strict(),gn=pt.extend({contentHash:R}).strict();function Za(e){let n=ve.parse(e.catalog),a=n.suites.find(l=>l.id===e.suiteId);if(!a)throw new Error(`Profile evaluation suite ${e.suiteId} is absent from its catalog.`);if(e.members.length!==a.definitionIds.length)throw new Error("Profile evaluation suite has an incomplete run population.");let s=e.members.map((l,d)=>{let c=a.definitionIds[d],f=n.definitions.find(w=>w.id===c);if(!f||l.definitionId!==c)throw new Error("Profile evaluation suite members differ from declared definition order.");let p=z.parse(l.manifest);V(p,"Profile evaluation suite member manifest");let h=p.profileEvaluation;if(!h||p.execution.kind!=="harness"||h.definitionId!==c||h.definitionHash!==o(f)||o(h.target)!==o(f.target)||h.catalogHash!==o(n)||p.tasksetRelease.id!==f.tasksetRelease.id||p.tasksetRelease.contentHash!==f.tasksetRelease.contentHash)throw new Error("Profile evaluation suite member differs from its released definition.");return{...l,manifest:p,source:h}}),r=s[0].source,i=s[0].manifest.policy;if(i.kind!=="model")throw new Error("Profile evaluation suite requires model runs.");for(let l of s){if(l.source.profileId!==r.profileId||l.source.sourceRevision!==r.sourceRevision||l.source.harnessRelease.id!==r.harnessRelease.id||l.source.harnessRelease.contentHash!==r.harnessRelease.contentHash)throw new Error("Profile evaluation suite members use different Profile releases.");let d=l.manifest.policy;if(d.kind!=="model"||o(d)!==o(i))throw new Error("Profile evaluation suite members use different model configurations.")}let u=pt.parse({schemaVersion:"openpond.profileEvaluationSuiteRun.v1",id:e.id,suiteId:a.id,suiteHash:o(a),catalogHash:o(n),profileId:r.profileId,sourceRevision:r.sourceRevision,harnessRelease:r.harnessRelease,members:s.map(l=>({definitionId:l.definitionId,runManifest:{id:l.manifest.id,contentHash:l.manifest.contentHash},runHash:l.runHash,passed:l.passed,score:l.score})),passed:s.every(l=>l.passed),createdAt:e.createdAt,completedAt:e.completedAt});return gn.parse({...u,contentHash:o(u)})}m(Za,"createProfileEvaluationSuiteRun");var Rn="openpond.connectedEvidence.v1",vn="connected-evidence-1",Se={files:100,sourceBytes:32*1024*1024,decodedBytes:64*1024*1024,lineBytes:64*1024*1024,events:1e5,sessions:1e3,boundaries:5e3,evaluatorBytes:2*1024*1024},P=t.string().min(1).max(500),Y=t.string().regex(/^[a-f0-9]{64}$/),wn=t.string().datetime().nullable(),He=t.number().int().nonnegative().nullable(),bn=t.enum(["native_chat","native_work","codex","claude_code","hermes","openclaw","opencode","grok_build","pi","oh_my_pi"]),xn=t.object({invocationId:P,inputTokens:He,outputTokens:He,cachedInputTokens:He,totalTokens:He,provenance:t.enum(["billed_invocation","reported_invocation","reported_cumulative"])}).strict(),ht=t.object({id:P,sequence:t.number().int().nonnegative(),occurredAt:wn,kind:t.enum(["message","tool_call","tool_result","compaction","artifact","terminal","usage","unknown"]),role:t.enum(["user","assistant","system","tool"]).nullable(),callId:P.nullable(),parentId:P.nullable(),content:t.json(),usage:xn.nullable()}).strict(),Sn=t.object({id:P,familyKey:P,projection:t.enum(["turn","conversation"]),requestEventId:P,start:t.number().int().nonnegative(),end:t.number().int().positive(),inputHash:Y,outputHash:Y.nullable(),revisionHash:Y,terminal:t.enum(["completed","failed","cancelled","unknown"]),coverage:t.object({answer:t.boolean(),context:t.enum(["retained","compacted","unknown"]),process:t.enum(["retained","partial","absent"]),artifacts:t.enum(["retained","references_only","absent"]),replay:t.literal("unavailable"),gaps:t.array(t.string().max(500)).max(100)}).strict()}).strict(),tr=t.object({schemaVersion:t.literal(Rn),normalizerVersion:t.literal(vn),origin:bn,sessionId:P,branchId:P.nullable(),parentSessionId:P.nullable(),exporterVersion:t.string().max(200).nullable(),acquisition:t.object({machineId:P,sourceInstanceId:P}).strict().optional(),sourceFiles:t.array(t.object({path:P,contentHash:Y,sizeBytes:t.number().int().nonnegative()}).strict()).max(Se.files),events:t.array(ht).max(Se.events),unmappedEvents:t.array(ht).max(Se.events),boundaries:t.array(Sn).max(Se.boundaries),warnings:t.array(t.string().max(500)).max(100),contentHash:Y}).strict();function nr(e){return e.kind!=="message"||e.role!=="assistant"||e.content===null?!1:typeof e.content=="string"?e.content.trim().length>0:Array.isArray(e.content)?e.content.length>0:typeof e.content!="object"||!("text"in e.content)||typeof e.content.text=="string"&&e.content.text.trim().length>0}m(nr,"hasRecordedConnectedAnswer");async function lr(e,n){return e.signal?.throwIfAborted(),Oe(e.manifest,e.taskset),gt(e.taskset,e.source),Rt({...e,policy:e.manifest.metricPolicy,receipts:Ve(e.manifest,e.receipts)},n)}m(lr,"aggregateTasksetRunReceipts");function gt(e,n){Re(e),e.metrics?.customAggregator&&vt(e.metrics.customAggregator,n)}m(gt,"assertTasksetMetricSource");async function dr(e,n){if(gt(e.taskset,e.source),e.taskset.id!==e.manifest.tasksetRelease.id||e.taskset.contentHash!==e.manifest.tasksetRelease.contentHash)throw new Error("Metric Taskset differs from the Run Manifest's pinned release.");let a=new Set(e.taskset.tasks.map(r=>r.id));if(e.receipts.some(r=>!a.has(r.taskId)))throw new Error("Metric receipt names a task outside the pinned Taskset.");let s=e.taskset.metrics?await Rt({...e,policy:e.taskset.metrics},n):void 0;return mt({id:e.id,manifest:e.manifest,receipts:e.receipts,metadata:e.metadata,authoredMetric:s})}m(dr,"aggregateTasksetEvaluationReceipts");async function Rt(e,n){e.signal?.throwIfAborted();let a=de.parse(e.policy),s=e.manifest.schemaVersion==="openpond.tasksetRunManifest.v1"?z.parse(e.manifest):K.parse(e.manifest);if(V(s,"Metric Run Manifest"),s.schemaVersion==="openpond.tasksetRunManifest.v1"){if(o(a)!==o(s.metricPolicy))throw new Error("Metric policy differs from the admitted run.");if(Ve(s,e.receipts).some((w,A)=>w.id!==e.receipts[A]?.id))throw new Error("Metric receipts must follow the admitted population order.")}if(!e.receipts.length||e.receipts.length>1e6)throw new Error("Metric execution requires 1 to 1,000,000 attempt receipts.");let r=new Set,i=[],u=[],l=0;for(let h of e.receipts){if(e.signal?.throwIfAborted(),!D(h))throw new Error(`Metric receipt ${h.id} has an invalid content hash.`);if(h.runManifest.id!==s.id||h.runManifest.contentHash!==s.contentHash)throw new Error(`Metric receipt ${h.id} belongs to a different Run Manifest.`);if(r.has(h.id))throw new Error(`Metric receipt ${h.id} is duplicated.`);if(r.add(h.id),!h.terminal||["infrastructure_failure","timeout","cancelled"].includes(h.failureClass??""))continue;let w=h.metadata.score;if(typeof w=="number"&&(!Number.isFinite(w)||w<0||w>1))throw new Error(`Metric receipt ${h.id} has an invalid score.`);let A=h.metadata.rewardEligible===!0&&typeof w=="number"&&(a.aggregation!=="pass_rate"||typeof h.metadata.passed=="boolean");if(!A&&(l+=1,a.missingReward==="exclude"))continue;let B=a.aggregation==="weighted_mean"&&Object.hasOwn(a.taskWeights,h.taskId)?a.taskWeights[h.taskId]:void 0;if(a.aggregation==="weighted_mean"&&B===void 0)throw new Error(`Metric policy has no weight for task ${h.taskId}.`);i.push(A?a.aggregation==="pass_rate"?Number(h.metadata.passed):w:0),u.push(B??1)}let d=null,c=e.receipts.map(({id:h,contentHash:w})=>({id:h,contentHash:w})),f=i.length;if(a.customAggregator){if(vt(a.customAggregator,e.source),!n)throw new Error("Custom metrics require an isolated executor.");if(i.length){let h=await n({source:e.source,module:a.customAggregator.module,exportName:a.customAggregator.exportName,scores:i,timeoutMs:a.customAggregator.timeoutMs,signal:e.signal});if(typeof h!="number"||!Number.isFinite(h)||h<0||h>1)throw new Error("Custom metric must return a finite number between 0 and 1.");d=h}}else if(i.length){let h=u.reduce((w,A)=>w+A,0);d=i.reduce((w,A,B)=>w+A*(u[B]/h),0),d=Math.max(0,Math.min(1,d))}e.signal?.throwIfAborted();let p=Ge.parse({schemaVersion:"openpond.tasksetMetricResult.v1",runManifest:{id:s.id,contentHash:s.contentHash},tasksetRelease:s.tasksetRelease,policy:a,policyHash:o(a),receiptRefs:c,includedCount:f,missingRewardCount:l,excludedCount:c.length-f,value:d});return ue.parse({...p,contentHash:o(p)})}m(Rt,"executeTasksetMetric");function vt(e,n){if(n===void 0||new TextEncoder().encode(n).byteLength>524288||Ke(n)!==e.contentHash)throw new Error("Metric module bytes do not match the pinned content hash or exceed the source limit.")}m(vt,"assertMetricSource");function wt(e,n){let a=pe.parse(e),{contentHash:s,...r}=a,i=a.decisions.at(-1);if(s!==o(r)||s!==n.reviewHash||a.status!=="accepted"||!i||i.outcome!=="accept"||i.contentHash!==n.decisionHash)throw new N("human_accepted_revision_unavailable",409);let u=i.submissions.map(l=>{let d=a.submissions.find(c=>c.id===l.id&&c.contentHash===l.contentHash);if(!d||d.evidenceHash!==o(a.evidence)||d.formHash!==o(a.form))throw new N("human_accepted_lineage_invalid",409);return{id:d.id,revision:d.revision,contentHash:d.contentHash}});return{schemaVersion:"openpond.acceptedHumanEvidence.v1",scope:a.scope,projectId:a.projectId,review:{id:a.id,revision:a.revision,contentHash:s},sourceOrigins:a.evidence.attempts.map(l=>l.localSource?l.localSource.kind:l.recordedSource?"recorded_evidence":"hosted_retained"),decision:{id:i.id,revision:i.revision,contentHash:i.contentHash},submissions:u,evidence:a.evidence,answers:i.answers,...a.execution?.result?{execution:{id:a.execution.executionId,result:a.execution.result,configurationHash:a.execution.configurationHash}}:{},...i.taskProposalHash?{taskProposal:a.submissions.find(l=>l.taskProposal?.contentHash===i.taskProposalHash).taskProposal}:{}}}m(wt,"acceptedHumanEvidence");var _e=t.string().regex(/^[a-f0-9]{64}$/),ee=t.string().min(1).max(240),ye={feedbackKey:ee,criterionId:ee},Hn=t.discriminatedUnion("kind",[t.object({...ye,kind:t.literal("normalized_score"),minimum:t.number().finite(),maximum:t.number().finite(),passThreshold:t.number().min(0).max(1).optional()}).strict(),t.object({...ye,kind:t.literal("boolean")}).strict(),t.object({...ye,kind:t.literal("category")}).strict(),t.object({...ye,kind:t.literal("preference_score"),tieScore:t.number().min(0).max(1).nullable()}).strict()]),bt=t.object({execution:g,resultHash:_e,reviews:t.array(t.object({review:$,decision:$}).strict()).min(1).max(1e4)}).strict(),Ht=t.object({rules:t.array(Hn).min(1).max(100),baseline:bt,candidate:bt}).strict().superRefine((e,n)=>{new Set(e.rules.map(a=>a.feedbackKey)).size!==e.rules.length&&n.addIssue({code:"custom",message:"Select each Human feedback key once."})}),xt=t.object({manifest:at,result:rt}).strict(),St=t.object({id:ee,manifestHash:_e,resultHash:_e}).strict(),yt=t.object({schemaVersion:t.literal("openpond.humanComparisonProjection.v1"),teamId:ee,projectId:ee,selections:Ht,original:t.object({baseline:St,candidate:St}).strict(),humanResults:t.object({baseline:xe,candidate:xe}).strict(),baseline:xt,candidate:xt}).strict(),_t=yt.extend({contentHash:_e}).strict();function y(e){throw new N(e,409)}m(y,"fail");async function Cr(e){let n=Ht.parse(e.selections);await e.authorize();let a={baseline:X(e.baseline),candidate:X(e.candidate)},s=new Set([n.baseline.execution.id,n.candidate.execution.id]);s.size!==2&&y("human_comparison_distinct_executions_required");let r=0;async function i(f){let p=a[f],h=n[f],w=p.manifest.lineage?.execution;p.manifest.lineage?.humanProjection&&y("human_comparison_recursive_projection"),(!w||o(w)!==o(h.execution)||p.manifest.teamId!==e.teamId)&&y("human_comparison_execution_changed");let A=[],B=new Map;for(let x of h.reviews){let b=await e.readReview(x.review);(b.scope!==e.teamId||b.projectId!==e.projectId||b.kind!=="grade"||b.revision!==x.review.revision||b.id!==x.review.id)&&y("human_comparison_review_scope_changed");let T=wt(b,{reviewHash:x.review.contentHash,decisionHash:x.decision.contentHash});r+=new TextEncoder().encode(JSON.stringify({answers:T.answers,form:b.form})).byteLength,r>4194304&&y("human_comparison_projection_too_large"),o(T.decision)!==o(x.decision)&&y("human_comparison_decision_changed"),b.evidence.attempts.some(q=>!s.has(q.experimentId))&&y("human_comparison_foreign_attempt");let E=b.evidence.attempts.filter(q=>q.experimentId===w.id);E.length!==1&&y("human_comparison_attempt_ambiguous");let H=E[0],M=await e.caseIdentity(H),I=p.result.cases.find(q=>Q(q.identity)===Q(M));(!I||M.caseId!==H.taskId||!I.traceRef||I.traceRef.contentHash!==H.trace?.contentHash||I.output===null)&&y("human_comparison_receipt_changed");let O=Q(M)+":"+b.evidence.grader.id;B.has(O)&&y("human_comparison_duplicate_assessment"),B.set(O,{review:b,accepted:T,attempt:H}),A.push(b)}let Ae=dt({executionId:w.id,executionHash:w.contentHash,reviews:A});(Ae.contentHash!==h.resultHash||Ae.coverage.accepted!==A.length)&&y("human_comparison_result_changed");let Fe=new Map(n.rules.map(x=>[x.feedbackKey,x])),en=p.manifest.evaluators.map(x=>{let b=Fe.get(x.feedbackKey);if(!b)return x;let T=A.filter(H=>H.evidence.grader.id===x.release.id&&String(H.evidence.grader.revision)===String(x.release.revision)&&H.evidence.grader.contentHash===x.release.contentHash).map(H=>H.form);(!T.length||new Set(T.map(o)).size!==1)&&y("human_comparison_form_changed");let E=T[0].criteria.find(H=>H.id===b.criterionId);return E||y("human_comparison_criterion_unavailable"),(b.kind==="normalized_score"&&(E.kind!=="score"||b.minimum!==E.minimum||b.maximum!==E.maximum||b.maximum<=b.minimum)||b.kind==="boolean"&&E.kind!=="boolean"||b.kind==="category"&&E.kind!=="category"||b.kind==="preference_score"&&(E.kind!=="preference"||T[0].mode!=="pairwise"))&&y("human_comparison_rule_changed"),{...x,output:b.kind==="boolean"?"boolean":b.kind==="category"?"category":"score",categories:E.kind==="category"?E.options:[],configurationHash:o({original:x.configurationHash,rule:b,formHash:o(T[0])})}});n.rules.some(x=>!p.manifest.evaluators.some(b=>b.feedbackKey===x.feedbackKey))&&y("human_comparison_grader_unavailable");let{contentHash:tn,...nn}=p.manifest,oe={originalManifest:{id:p.manifest.id,contentHash:p.manifest.contentHash},selectionHash:o(h),rulesHash:o(n.rules)},ce=st({...nn,id:nt(oe.originalManifest,oe.selectionHash,oe.rulesHash),lineage:{...p.manifest.lineage,humanProjection:oe},evaluators:en}),an=p.result.cases.map(x=>{let b=x.feedback.map(E=>{let H=Fe.get(E.feedbackKey);if(!H)return E;let M=B.get(Q(x.identity)+":"+E.evaluator.id);(!M||M.review.evidence.grader.contentHash!==E.evaluator.contentHash)&&y("human_comparison_assessment_missing");let I=M.accepted.answers.find(q=>q.criterionId===H.criterionId),O=null;if(I&&!I.abstain&&(H.kind==="normalized_score"&&typeof I.value=="number"&&(O=(I.value-H.minimum)/(H.maximum-H.minimum)),H.kind==="boolean"&&typeof I.value=="boolean"&&(O=I.value),H.kind==="category"&&typeof I.value=="string"&&(O=I.value),H.kind==="preference_score")){let q=M.review.evidence.attempts.findIndex(Je=>Je.experimentId===w.id&&Je.attemptId===M.attempt.attemptId);O=I.value==="tie"?H.tieScore:I.value==="A"||I.value==="B"?I.value===(q===0?"A":"B")?1:0:null}return{...E,status:O===null?"unavailable":"scored",value:O,passed:O===null?null:H.kind==="boolean"?O:H.kind==="normalized_score"&&H.passThreshold!==void 0?Number(O)>=H.passThreshold:null,reasoning:I?.note??null,evidenceRefs:[{id:M.accepted.review.id,contentHash:M.accepted.review.contentHash,mediaType:"application/vnd.openpond.human-review+json",sizeBytes:null},{id:M.accepted.decision.id,contentHash:M.accepted.decision.contentHash,mediaType:"application/vnd.openpond.human-decision+json",sizeBytes:null}]}}),T=x.output!==null&&b.every(E=>E.status==="scored");return{...x,feedback:b,status:T?"completed":x.status,error:T?null:x.error}}),{contentHash:rn,...sn}=p.result,on=it({...sn,manifest:{id:ce.id,contentHash:ce.contentHash},cases:an},ce);for(let x of A){let b=await e.readReview({id:x.id,revision:x.revision,contentHash:x.contentHash});(b.contentHash!==x.contentHash||b.status!=="accepted")&&y("human_comparison_review_changed")}return await e.authorize(),{evidence:{manifest:ce,result:on},humanResult:Ae}}m(i,"lane");let u=await i("baseline"),l=await i("candidate");for(let f of[...n.baseline.reviews,...n.candidate.reviews]){let p=await e.readReview(f.review);(p.contentHash!==f.review.contentHash||p.status!=="accepted"||p.decisions.at(-1)?.contentHash!==f.decision.contentHash)&&y("human_comparison_review_changed")}await e.authorize();let d={baseline:{id:a.baseline.manifest.id,manifestHash:a.baseline.manifest.contentHash,resultHash:a.baseline.result.contentHash},candidate:{id:a.candidate.manifest.id,manifestHash:a.candidate.manifest.contentHash,resultHash:a.candidate.result.contentHash}},c=yt.parse({schemaVersion:"openpond.humanComparisonProjection.v1",teamId:e.teamId,projectId:e.projectId,selections:n,original:d,humanResults:{baseline:u.humanResult,candidate:l.humanResult},baseline:u.evidence,candidate:l.evidence});return new TextEncoder().encode(JSON.stringify(c)).byteLength>8388608&&y("human_comparison_projection_too_large"),_t.parse({...c,contentHash:o(c)})}m(Cr,"projectAcceptedHumanComparison");function Ir(e,n){let a=_t.parse(e),{contentHash:s,...r}=a;s!==o(r)&&y("human_comparison_projection_invalid");for(let i of["baseline","candidate"]){let u=X(a[i]),l=u.manifest.lineage,d=l?.humanProjection,c=a.original[i],f=a.selections[i];(!l||!d||d.originalManifest.id!==c.id||d.originalManifest.contentHash!==c.manifestHash||d.selectionHash!==o(f)||d.rulesHash!==o(a.selections.rules)||o(l.execution)!==o(f.execution)||u.manifest.teamId!==a.teamId||a.humanResults[i].contentHash!==f.resultHash)&&y("human_comparison_projection_lineage_changed");let{contentHash:p,...h}=a.humanResults[i];if((o(h)!==p||h.executionId!==f.execution.id||h.executionHash!==f.execution.contentHash||o(h.selection)!==o(f.reviews.map(w=>w.review)))&&y("human_comparison_human_result_changed"),n){let w=X(n[i]);(w.manifest.lineage?.humanProjection||w.manifest.id!==c.id||w.manifest.contentHash!==c.manifestHash||w.result.contentHash!==c.resultHash||o(w.manifest.lineage?.execution)!==o(f.execution))&&y("human_comparison_original_changed")}}return a}m(Ir,"verifyHumanComparisonProjection");var Et=t.enum(["scored","unscorable"]),ze=t.enum(["policy","environment","collector","verifier","provider","host","user","scheduler"]),yn=t.enum(["completed","policy_failure","incomplete_output","task_deadline","environment_failure","collector_failure","verifier_failure","provider_failure","host_failure","infrastructure_timeout","cancelled"]),te=t.object({schemaVersion:t.literal("openpond.environmentRelease.v1"),id:v,revision:t.number().int().positive(),contract:Xe,actionSchemaRef:J.nullable(),observationSchemaRef:J.nullable(),stateSchemaRef:J.nullable(),artifactCollection:t.object({maxArtifacts:t.number().int().positive().max(1e5),maxTotalBytes:t.number().int().positive().max(1e10)}).strict(),adapterConformanceHashes:t.record(v,R),metadata:C}).strict(),Ct=te.extend({contentHash:R}).strict(),_n=t.enum(["collected","missing","skipped","failed"]),kt=t.enum(["not_requested","passed","failed"]),kn=t.object({requiredOutputPath:t.string().trim().min(1).max(2e3).nullable(),collectedPath:t.string().trim().min(1).max(4e3).nullable(),declaredMediaType:t.string().trim().min(1).max(200).nullable(),detectedMediaType:t.string().trim().min(1).max(200).nullable(),artifact:j.nullable(),status:_n,parseStatus:kt,schemaStatus:kt,errorCode:v.nullable(),failureOwner:ze.nullable(),evidenceRefs:t.array(j).max(1e4),metadata:C}).strict(),L=t.object({schemaVersion:t.literal("openpond.artifactManifest.v1"),id:v,attemptRef:g,entries:t.array(kn).max(1e5),createdAt:_,metadata:C}).strict(),ne=L.extend({contentHash:R}).strict(),U=t.object({schemaVersion:t.literal("openpond.verifierSetRelease.v1"),id:v,revision:t.number().int().positive(),graders:t.array(Ye).min(1).max(1e3),isolation:t.object({processBoundary:t.enum(["same_process","isolated_process","container"]),networkPolicy:t.literal("none"),defaultTimeoutMs:t.number().int().positive().max(3e5)}).strict(),calibrationReceiptRefs:t.array(g).max(1e4),metadata:C}).strict(),It=U.extend({contentHash:R}).strict(),F=t.object({verifierId:v,verifierVersion:t.string().trim().min(1).max(100),status:Et,rawScore:t.number().finite().nullable(),normalizedScore:t.number().min(0).max(1).nullable(),weight:t.number().nonnegative().max(1e3),passed:t.boolean(),hardGate:t.boolean(),rewardEligible:t.boolean(),rewardContribution:t.number().min(0).max(1).nullable(),failureOwner:ze.nullable(),feedback:t.array(t.string().max(2e4)).max(1e3),visibleEvidenceRefs:t.array(j).max(1e4),privilegedEvidenceRefs:t.array(j).max(1e4),metadata:C}).strict().superRefine((e,n)=>{e.status==="scored"&&e.normalizedScore===null&&n.addIssue({code:"custom",message:"A scored component requires a normalized score.",path:["normalizedScore"]}),e.status==="unscorable"&&(e.normalizedScore!==null||e.rewardEligible)&&n.addIssue({code:"custom",message:"An unscorable component cannot contribute reward.",path:["status"]})}),At=t.object({schemaVersion:t.literal("openpond.rewardReceipt.v1"),id:v,attemptRef:g,verifierSetRef:g,artifactManifestRef:g,status:Et,reward:t.number().min(0).max(1).nullable(),learningEligible:t.boolean(),passed:t.boolean(),outcomeClass:yn,failureOwner:ze.nullable(),components:t.array(F).min(1).max(1e4),visibleEvidenceRefs:t.array(j).max(1e5),privilegedEvidenceRefs:t.array(j).max(1e5),supersedes:g.nullable(),createdAt:_,metadata:C}).strict();function Mt(e,n){e.status==="scored"&&(e.reward===null||!e.learningEligible)&&n.addIssue({code:"custom",message:"A scored receipt requires a reward and is learning-eligible.",path:["status"]}),e.status==="unscorable"&&(e.reward!==null||e.learningEligible)&&n.addIssue({code:"custom",message:"An unscorable receipt has null reward and is not learning-eligible.",path:["status"]})}m(Mt,"validateRewardReceipt");var qe=At.superRefine(Mt),Be=At.extend({contentHash:R}).strict().superRefine(Mt);function Nr(e){let n=te.parse(e);return Ct.parse({...n,contentHash:o(n)})}m(Nr,"createEnvironmentRelease");function $r(e){let n=L.parse(e);return ne.parse({...n,contentHash:o(n)})}m($r,"createArtifactManifest");function Dr(e){let n=U.parse(e);return It.parse({...n,contentHash:o(n)})}m(Dr,"createVerifierSetRelease");function Lr(e){if(ae(e.environment,te,"Environment Release"),ae(e.verifierSet,U,"Verifier Set Release"),o(e.taskset.environment)!==o(e.environment.contract))throw new Error("Environment Release does not match the Taskset execution contract.");if(o(e.taskset.graders)!==o(e.verifierSet.graders))throw new Error("Verifier Set Release does not match the Taskset graders.");let{contentHash:n,...a}=e.taskset,s=he.parse({...a,environmentRelease:{id:e.environment.id,contentHash:e.environment.contentHash},verifierSetRelease:{id:e.verifierSet.id,contentHash:e.verifierSet.contentHash}});return ge.parse({...s,contentHash:o(s)})}m(Lr,"bindTasksetExecutionReleases");function Fr(e){switch(e.failureClass){case null:return{outcomeClass:"completed",failureOwner:null};case"policy_failure":return{outcomeClass:"policy_failure",failureOwner:"policy"};case"environment_failure":return{outcomeClass:"environment_failure",failureOwner:"environment"};case"grader_failure":return{outcomeClass:"verifier_failure",failureOwner:"verifier"};case"infrastructure_failure":return{outcomeClass:"host_failure",failureOwner:"host"};case"cancelled":return{outcomeClass:"cancelled",failureOwner:"user"};case"timeout":return e.timeoutKind==="task_deadline"?{outcomeClass:"task_deadline",failureOwner:"policy"}:{outcomeClass:"infrastructure_timeout",failureOwner:"host"}}}m(Fr,"classifyAttemptOutcome");function Jr(e){if(ae(e.verifierSet,U,"Verifier Set Release"),ae(e.artifactManifest,L,"Artifact Manifest"),e.artifactManifest.attemptRef.id!==e.attemptRef.id||e.artifactManifest.attemptRef.contentHash!==e.attemptRef.contentHash)throw new Error("Artifact Manifest does not belong to the Reward Receipt Attempt.");let n=En(e.outcomeClass),a=e.components.map(p=>{let h=F.parse(p),w=n&&h.status==="scored"&&h.normalizedScore!==null&&h.rewardEligible;return F.parse({...h,rewardEligible:w,rewardContribution:w?h.normalizedScore:null})}),s=a.filter(p=>p.rewardEligible&&p.normalizedScore!==null),r=s.length>0?"scored":"unscorable",i=s.some(p=>p.hardGate&&!p.passed),u=s.reduce((p,h)=>p+h.weight,0),l=u>0?s.reduce((p,h)=>p+h.normalizedScore*h.weight,0)/u:0,d=r==="scored"?i?0:l:null,c=r==="scored"&&s.every(p=>p.passed),f=qe.parse({schemaVersion:"openpond.rewardReceipt.v1",id:e.id,attemptRef:e.attemptRef,verifierSetRef:{id:e.verifierSet.id,contentHash:e.verifierSet.contentHash},artifactManifestRef:{id:e.artifactManifest.id,contentHash:e.artifactManifest.contentHash},status:r,reward:d,learningEligible:r==="scored",passed:c,outcomeClass:e.outcomeClass,failureOwner:e.failureOwner,components:a,visibleEvidenceRefs:jt([...e.visibleEvidenceRefs??[],...a.flatMap(p=>p.visibleEvidenceRefs)]),privilegedEvidenceRefs:jt([...e.privilegedEvidenceRefs??[],...a.flatMap(p=>p.privilegedEvidenceRefs)]),supersedes:e.supersedes??null,createdAt:e.createdAt,metadata:e.metadata??{}});return Be.parse({...f,contentHash:o(f)})}m(Jr,"createRewardReceipt");function Kr(e){return Ne(e,te)}m(Kr,"verifyEnvironmentRelease");function Ur(e){return Ne(e,L)}m(Ur,"verifyArtifactManifest");function Wr(e){return Ne(e,U)}m(Wr,"verifyVerifierSetRelease");function Gr(e){let n=Be.safeParse(e);if(!n.success)return!1;let{contentHash:a,...s}=n.data;return o(qe.parse(s))===a}m(Gr,"verifyRewardReceipt");function En(e){return e==="completed"||e==="policy_failure"||e==="incomplete_output"||e==="task_deadline"}m(En,"isScorableOutcome");function jt(e){return[...new Map(e.map(n=>[`${n.id}:${n.contentHash}`,n])).values()]}m(jt,"uniqueArtifactRefs");function ae(e,n,a){let{contentHash:s,...r}=e,i=o(n.parse(r));if(s!==i)throw new Error(`${a} content hash mismatch.`)}m(ae,"verifyContentHash");function Ne(e,n){try{return ae(e,n,"Receipt"),!0}catch{return!1}}m(Ne,"hasValidContentHash");var Cn=t.enum(["calibration","training_reward","validation","frozen_eval"]),as=t.enum(["human","model","deterministic","fixture"]),In=t.enum(["production","synthetic_smoke"]),An=t.enum(["markdown","text","code","json","image"]),Mn=t.object({id:v,label:t.string().trim().min(1).max(500),instruction:t.string().trim().min(1).max(1e4),weight:t.number().nonnegative().max(1e3)}).strict(),jn=t.object({source:t.enum(["attempt_output","artifact"]),path:t.string().trim().min(1).max(2e3),renderer:An}).strict(),Pn=t.object({showTaskPrompt:t.boolean(),randomizeCandidateOrder:t.boolean(),hideModelIdentity:t.boolean(),parts:t.array(jn).min(1).max(20)}).strict().superRefine((e,n)=>{e.parts.some(a=>a.renderer==="markdown"||a.renderer==="text")||n.addIssue({code:"custom",path:["parts"],message:"A preference presentation must expose text or Markdown content."})}),Tn=t.object({strategy:t.literal("randomized_blinded_v1"),maxAssignmentsPerCandidate:t.number().int().positive().max(1e4)}).strict(),On=t.object({algorithm:t.literal("mean_pairwise_win_fraction_v1"),quorum:t.number().int().positive().max(100),rejectAllThreshold:t.number().min(.5).max(1)}).strict(),Vn=t.object({algorithm:t.literal("pairwise_win_fraction_v1"),verifierId:v,verifierVersion:t.string().trim().min(1).max(100),weight:t.number().positive().max(1e3)}).strict(),zn=t.object({minimumSamples:t.number().int().positive().max(1e5),minimumOrderAgreement:t.number().min(0).max(1),minimumTieAgreement:t.number().min(0).max(1),minimumOrderSwapAgreement:t.number().min(0).max(1)}).strict(),De=t.object({schemaVersion:t.literal("openpond.preferenceComparisonRelease.v1"),id:v,revision:t.number().int().positive(),tasksetRelease:g,candidateCount:t.number().int().min(2).max(4),resultMode:t.literal("ordered_tie_groups"),allowTies:t.boolean(),allowRejectAll:t.boolean(),presentation:Pn,rubricRef:j,criteria:t.array(Mn).max(20),assignment:Tn,aggregation:On,rewardProjection:Vn,calibration:zn,metadata:C}).strict().superRefine((e,n)=>{let a=e.criteria.map(s=>s.id);new Set(a).size!==a.length&&n.addIssue({code:"custom",path:["criteria"],message:"Preference criterion IDs must be unique."})}),Tt=De.extend({contentHash:R}).strict(),qn=t.object({id:v,contentHash:R}).strict(),Bn=t.object({tasksetRelease:g,harnessReleases:t.array(g).min(1).max(4),runManifestRefs:t.array(g).min(1).max(4),harnessCompatibilityReceiptRefs:t.array(g).max(6),environmentRelease:g,verifierSetRelease:g,toolContractHash:R,policyHash:R}).strict(),Nn=t.object({attemptRef:g,runManifestRef:g,artifactManifestRef:g,visibleArtifactIds:t.array(v).max(1e5)}).strict().superRefine((e,n)=>{new Set(e.visibleArtifactIds).size!==e.visibleArtifactIds.length&&n.addIssue({code:"custom",path:["visibleArtifactIds"],message:"Visible artifact IDs must be unique per candidate."})}),Ot=t.object({schemaVersion:t.literal("openpond.comparisonAssignment.v1"),id:v,comparisonRelease:g,taskRef:qn,lineage:Bn,candidates:t.array(Nn).min(2).max(4),presentedCandidateOrder:t.array(v).min(2).max(4),purpose:Cn,createdAt:_,metadata:C}).strict().superRefine((e,n)=>{let a=e.candidates.map(s=>s.attemptRef.id);new Set(a).size!==a.length&&n.addIssue({code:"custom",path:["candidates"],message:"An assignment cannot contain the same attempt twice."}),(new Set(e.presentedCandidateOrder).size!==e.presentedCandidateOrder.length||e.presentedCandidateOrder.length!==a.length||e.presentedCandidateOrder.some(s=>!a.includes(s)))&&n.addIssue({code:"custom",path:["presentedCandidateOrder"],message:"Presented candidate order must contain every assignment candidate exactly once."})}),Vt=Ot,zt=Ot.extend({contentHash:R}).strict().superRefine((e,n)=>{let a=e.candidates.map(s=>s.attemptRef.id);new Set(a).size!==a.length&&n.addIssue({code:"custom",path:["candidates"],message:"An assignment cannot contain the same attempt twice."})}),qt=t.array(t.array(v).min(1).max(4)).max(4),$n=t.discriminatedUnion("kind",[t.object({kind:t.enum(["human","model","deterministic"]),releaseRef:g}).strict(),t.object({kind:t.literal("fixture"),releaseRef:g,fixtureRelease:g,labelSource:t.literal("synthetic"),qualificationEligibility:t.literal("smoke_only")}).strict()]),Bt=t.object({schemaVersion:t.literal("openpond.preferenceReceipt.v1"),id:v,assignmentRef:g,reviewer:$n,order:qt,rejectAll:t.boolean(),criterionScores:t.record(v,t.record(v,t.number().finite().min(0).max(1))),feedbackArtifactRef:j.nullable(),startedAt:_,completedAt:_,metadata:C}).strict().superRefine((e,n)=>{e.rejectAll&&e.order.length!==0&&n.addIssue({code:"custom",path:["order"],message:"A reject-all preference receipt cannot contain an order."}),!e.rejectAll&&e.order.length===0&&n.addIssue({code:"custom",path:["order"],message:"A non-reject preference receipt requires an ordered result."}),Date.parse(e.completedAt)<Date.parse(e.startedAt)&&n.addIssue({code:"custom",path:["completedAt"],message:"A preference receipt cannot complete before it starts."})}),Nt=Bt,$t=Bt.extend({contentHash:R}).strict(),Dt=t.object({schemaVersion:t.literal("openpond.preferenceAggregationReceipt.v1"),id:v,assignmentRef:g,comparisonRelease:g,receiptRefs:t.array(g).min(1).max(100),order:qt,rejectAll:t.boolean(),pairwiseWinFractions:t.record(v,t.number().min(0).max(1)),quorum:t.number().int().positive().max(100),createdAt:_,metadata:C}).strict().superRefine((e,n)=>{e.rejectAll&&e.order.length!==0&&n.addIssue({code:"custom",path:["order"],message:"A reject-all aggregate cannot contain an order."})}),Dn=Dt,Ln=Dt.extend({contentHash:R}).strict(),Lt=t.object({schemaVersion:t.literal("openpond.preferenceCalibrationReport.v1"),id:v,comparisonRelease:g,automatedReviewer:g,humanReceiptRefs:t.array(g).min(1).max(1e5),modelReceiptRefs:t.array(g).min(1).max(1e5),sampleCount:t.number().int().positive().max(1e5),orderAgreement:t.number().min(0).max(1),tieAgreement:t.number().min(0).max(1),orderSwapAgreement:t.number().min(0).max(1),passed:t.boolean(),createdAt:_,metadata:C}).strict(),Ft=Lt,Jt=Lt.extend({contentHash:R}).strict();function rs(e){let n=De.parse(e);return Tt.parse({...n,contentHash:o(n)})}m(rs,"createPreferenceComparisonRelease");function Fn(e){return se(e,De,Tt)}m(Fn,"verifyPreferenceComparisonRelease");function ss(e){let n=G(e.comparisonRelease),a=sa(e.taskset);if(!k(n.tasksetRelease,S(a)))throw new Error("Preference Comparison Release does not belong to the supplied Taskset Release.");if(e.candidates.length!==n.candidateCount)throw new Error("Comparison assignment candidate count does not match the immutable comparison release.");let s=n.presentation.parts.some(f=>f.source==="artifact"),r=e.candidates.map(f=>Qn(f,a,s)),i=r[0];for(let f of r.slice(1)){if(f.attempt.taskId!==i.attempt.taskId)throw new Error("Preference comparison candidates must belong to the same Taskset task.");let p=aa(i.runManifest,f.runManifest,e.harnessCompatibilityReceipts??[]);p&&k(p.baseHarnessRelease,f.runManifest.harnessRelease)?Te(f.runManifest,i.runManifest,p):Te(i.runManifest,f.runManifest,p)}let u=a.tasks.find(f=>f.id===i.attempt.taskId);if(!u)throw new Error("Comparison candidates reference a task absent from the supplied Taskset Release.");if(!a.environmentRelease||!a.verifierSetRelease)throw new Error("Preference comparison requires a Taskset bound to immutable Environment and Verifier Set Releases.");let l=r.map(f=>f.attempt.id),d=[...e.presentedCandidateOrder??l],c=Vt.parse({schemaVersion:"openpond.comparisonAssignment.v1",id:e.id,comparisonRelease:S(n),taskRef:{id:u.id,contentHash:o({taskId:u.id,tasksetRelease:S(a)})},lineage:{tasksetRelease:S(a),harnessReleases:$e(r.map(f=>f.runManifest.harnessRelease)),runManifestRefs:$e(r.map(f=>S(f.runManifest))),harnessCompatibilityReceiptRefs:$e((e.harnessCompatibilityReceipts??[]).map(S)),environmentRelease:a.environmentRelease,verifierSetRelease:a.verifierSetRelease,toolContractHash:o(a.tools),policyHash:o(a.policy)},candidates:r.map(f=>({attemptRef:S(f.attempt),runManifestRef:S(f.runManifest),artifactManifestRef:S(f.artifactManifest),visibleArtifactIds:[...f.visibleArtifactIds]})),presentedCandidateOrder:d,purpose:e.purpose,createdAt:e.createdAt,metadata:e.metadata??{}});return zt.parse({...c,contentHash:o(c)})}m(ss,"createComparisonAssignment");function Jn(e){return se(e,Vt,zt)}m(Jn,"verifyComparisonAssignment");function Le(e,n){let a=ie(e),s=G(n);if(!k(a.comparisonRelease,S(s)))throw new Error("Comparison assignment does not reference the supplied comparison release.");if(a.candidates.length!==s.candidateCount)throw new Error("Comparison assignment does not satisfy its release candidate count.")}m(Le,"validateComparisonAssignment");function Kn(e){let n=ie(e.assignment),a=G(e.comparisonRelease);Le(n,a);let s=Nt.parse({schemaVersion:"openpond.preferenceReceipt.v1",id:e.id,assignmentRef:S(n),reviewer:e.reviewer,order:e.order.map(r=>[...r]),rejectAll:e.rejectAll,criterionScores:e.criterionScores??{},feedbackArtifactRef:e.feedbackArtifactRef??null,startedAt:e.startedAt,completedAt:e.completedAt,metadata:e.metadata??{}});return W({order:s.order,rejectAll:s.rejectAll},n,a),Xn(s.criterionScores,n,a),$t.parse({...s,contentHash:o(s)})}m(Kn,"createPreferenceReceipt");function is(e){let n=e.assignment.candidates.map(l=>l.attemptRef.id),a=Object.keys(e.ratings);if(a.length!==n.length||n.some(l=>!e.ratings[l])||a.some(l=>!n.includes(l)))throw new Error("Synthetic fixture ratings must label every assignment candidate exactly once.");let s=["love","like","reject"],r=n.every(l=>e.ratings[l]==="reject"),i=r?[]:s.map(l=>n.filter(d=>e.ratings[d]===l)).filter(l=>l.length>0),u={love:1,like:.5,reject:0};return Kn({id:e.id,assignment:e.assignment,comparisonRelease:e.comparisonRelease,reviewer:{kind:"fixture",releaseRef:e.labelerRelease,fixtureRelease:e.fixtureRelease,labelSource:"synthetic",qualificationEligibility:"smoke_only"},order:i,rejectAll:r,criterionScores:Object.fromEntries(n.map(l=>[l,Object.fromEntries(e.comparisonRelease.criteria.map(d=>[d.id,u[e.ratings[l]]]))])),startedAt:e.startedAt,completedAt:e.completedAt,metadata:e.metadata??{}})}m(is,"createSyntheticFixturePreferenceReceipt");function Un(e){return se(e,Nt,$t)}m(Un,"verifyPreferenceReceipt");function Wn(e){return se(e,Dn,Ln)}m(Wn,"verifyPreferenceAggregationReceipt");function os(e){let n=G(e.comparisonRelease);if(e.pairs.length<n.calibration.minimumSamples)throw new Error(`Preference calibration requires at least ${n.calibration.minimumSamples} human/model pairs.`);let a=e.pairs.map(c=>ta(c,n)),s=a[0].model.reviewer.releaseRef;if(a.some(c=>!k(c.model.reviewer.releaseRef,s)))throw new Error("Preference calibration requires model receipts from one immutable automated reviewer release.");let r=re(a.map(c=>Pt(c.human,c.model,c.assignment))),i=re(a.map(c=>ea(c.human,c.model,c.assignment))),u=re(a.map(c=>c.swappedModel?Pt(c.model,c.swappedModel,c.assignment):1)),l=r>=n.calibration.minimumOrderAgreement&&i>=n.calibration.minimumTieAgreement&&u>=n.calibration.minimumOrderSwapAgreement,d=Ft.parse({schemaVersion:"openpond.preferenceCalibrationReport.v1",id:e.id,comparisonRelease:S(n),automatedReviewer:s,humanReceiptRefs:a.map(c=>S(c.human)),modelReceiptRefs:a.map(c=>S(c.model)),sampleCount:a.length,orderAgreement:r,tieAgreement:i,orderSwapAgreement:u,passed:l,createdAt:e.createdAt,metadata:e.metadata??{}});return Jt.parse({...d,contentHash:o(d)})}m(os,"createPreferenceCalibrationReport");function Gn(e){return se(e,Ft,Jt)}m(Gn,"verifyPreferenceCalibrationReport");function Zn(e){let n=ie(e.assignment),a=G(e.comparisonRelease);if(Le(n,a),e.result.schemaVersion==="openpond.preferenceReceipt.v1"){let r=ke(e.result);if(!k(r.assignmentRef,S(n)))throw new Error("Preference receipt does not belong to the supplied assignment.");return W(r,n,a),Yn(r,n)}let s=ra(e.result);if(!k(s.assignmentRef,S(n))||!k(s.comparisonRelease,S(a)))throw new Error("Preference aggregate does not belong to the supplied assignment and release.");return W(s,n,a),s.rejectAll?Ut(n):s.pairwiseWinFractions}m(Zn,"projectPairwiseWinFraction");function cs(e){let n=ie(e.assignment),a=G(e.comparisonRelease),s=Zn({assignment:n,comparisonRelease:a,result:e.result}),r=new Map((e.candidates??[]).map(f=>[f.attemptRef.id,f])),i=e.result.schemaVersion==="openpond.preferenceReceipt.v1"?e.result:null,u=i?.reviewer.kind==="model",l=i?.reviewer.kind==="fixture",d=In.parse(e.rewardScope??"production"),c=l?d==="synthetic_smoke":!u||na(i,a,e.calibrationReport??null);return Object.fromEntries(n.candidates.map(f=>{let p=r.get(f.attemptRef.id),h=p?.eligible??!0,w=s[f.attemptRef.id],A=h&&c?F.parse({verifierId:a.rewardProjection.verifierId,verifierVersion:a.rewardProjection.verifierVersion,status:"scored",rawScore:w,normalizedScore:w,weight:a.rewardProjection.weight,passed:w>0,hardGate:!1,rewardEligible:!0,rewardContribution:w,failureOwner:null,feedback:[],visibleEvidenceRefs:[],privilegedEvidenceRefs:[],metadata:{comparisonAssignment:S(n),preferenceResult:S(e.result),...e.calibrationReport?{calibrationReport:S(e.calibrationReport)}:{}}}):F.parse({verifierId:a.rewardProjection.verifierId,verifierVersion:a.rewardProjection.verifierVersion,status:"unscorable",rawScore:null,normalizedScore:null,weight:a.rewardProjection.weight,passed:!1,hardGate:!1,rewardEligible:!1,rewardContribution:null,failureOwner:p?.failureOwner??"verifier",feedback:[h?l?"Synthetic fixture preference evidence is admitted only to synthetic-smoke reward scope.":"Automated preference reviewer is uncalibrated or inconsistent.":"Candidate is not eligible for comparative preference scoring."],visibleEvidenceRefs:[],privilegedEvidenceRefs:[],metadata:{comparisonAssignment:S(n),preferenceResult:S(e.result),...e.calibrationReport?{calibrationReport:S(e.calibrationReport)}:{}}});return[f.attemptRef.id,A]}))}m(cs,"createPreferenceRewardComponents");function Qn(e,n,a){let s=be.parse(e.attempt);if(!D(s))throw new Error("Comparison candidate Attempt Receipt has an invalid content hash.");let r=ne.parse(e.artifactManifest);if(!ia(r))throw new Error("Comparison candidate Artifact Manifest has an invalid content hash.");let i=K.parse(e.runManifest);if(!oa(i))throw new Error("Comparison candidate Run Manifest has an invalid content hash.");if(!k(s.runManifest,S(i)))throw new Error("Comparison candidate Attempt Receipt does not belong to its supplied Run Manifest.");if(!k(i.tasksetRelease,S(n)))throw new Error("Comparison candidate Run Manifest does not belong to the supplied Taskset Release.");if(!k(r.attemptRef,S(s)))throw new Error("Comparison candidate Artifact Manifest does not belong to its Attempt Receipt.");if(!s.terminal||s.failureClass!==null)throw new Error("Only completed, non-failed attempts can be assigned for comparative review.");let u=[...e.visibleArtifactIds],l=new Set(u),d=new Set(r.entries.filter(c=>c.status==="collected"&&c.artifact!==null).map(c=>c.artifact.id));if(a&&!u.length)throw new Error("Comparison presentation requires each candidate to expose a reviewable artifact.");if(u.some(c=>!d.has(c)))throw new Error("Comparison candidate exposes an artifact that is missing, uncollected, or unreviewable.");if(l.size!==u.length)throw new Error("Comparison candidate visible artifacts must be unique.");return{attempt:s,artifactManifest:r,runManifest:i,visibleArtifactIds:u}}m(Qn,"validateComparisonCandidate");function W(e,n,a){let s=n.candidates.map(i=>i.attemptRef.id);if(e.rejectAll){if(!a.allowRejectAll)throw new Error("This comparison release does not allow reject-all results.");if(e.order.length)throw new Error("Reject-all results cannot include ordered candidates.");return}let r=e.order.flat();if(new Set(r).size!==r.length||r.length!==s.length||r.some(i=>!s.includes(i)))throw new Error("Preference result must rank every assignment candidate exactly once.");if(!a.allowTies&&e.order.some(i=>i.length>1))throw new Error("This comparison release does not allow tied candidates.")}m(W,"validatePreferenceResult");function Xn(e,n,a){let s=new Set(n.candidates.map(i=>i.attemptRef.id)),r=new Set(a.criteria.map(i=>i.id));for(let[i,u]of Object.entries(e)){if(!s.has(i))throw new Error("Criterion scores include a candidate absent from the assignment.");for(let l of Object.keys(u))if(!r.has(l))throw new Error("Criterion scores include a criterion absent from the comparison release.")}}m(Xn,"validateCriterionScores");function Yn(e,n){if(e.rejectAll)return Ut(n);let a=n.candidates.map(r=>r.attemptRef.id),s=new Map(e.order.flatMap((r,i)=>r.map(u=>[u,i])));return Object.fromEntries(a.map(r=>{let i=s.get(r);if(i===void 0)throw new Error("Preference result is missing a comparison candidate.");let u=0;for(let l of a){if(l===r)continue;let d=s.get(l);if(d===void 0)throw new Error("Preference result is missing a comparison candidate.");u+=i<d?1:i===d?.5:0}return[r,u/(a.length-1)]}))}m(Yn,"pairwiseScoresForResult");function Pt(e,n,a){let s=a.candidates.map(i=>i.attemptRef.id),r=Kt(s);return r.length?re(r.map(([i,u])=>Ee(e,i,u)===Ee(n,i,u)?1:0)):1}m(Pt,"pairwiseAgreement");function ea(e,n,a){let s=Kt(a.candidates.map(r=>r.attemptRef.id));return s.length?re(s.map(([r,i])=>Ee(e,r,i)===0==(Ee(n,r,i)===0)?1:0)):1}m(ea,"tieAgreementScore");function Ee(e,n,a){if(e.rejectAll)return 0;let s=new Map(e.order.flatMap((u,l)=>u.map(d=>[d,l]))),r=s.get(n),i=s.get(a);if(r===void 0||i===void 0)throw new Error("Preference result does not cover all comparison candidates.");return r<i?1:r===i?0:-1}m(Ee,"pairRelation");function Kt(e){let n=[];for(let a=0;a<e.length;a+=1)for(let s=a+1;s<e.length;s+=1)n.push([e[a],e[s]]);return n}m(Kt,"pairsFor");function ta(e,n){let a=ie(e.assignment);if(Le(a,n),a.purpose==="frozen_eval")throw new Error("Frozen-evaluation comparisons cannot calibrate an automated preference reviewer.");let s=ke(e.human),r=ke(e.model);if(s.reviewer.kind!=="human"||r.reviewer.kind!=="model")throw new Error("Preference calibration requires one human and one automated model receipt per assignment.");if(!k(s.assignmentRef,S(a))||!k(r.assignmentRef,S(a)))throw new Error("Preference calibration receipts must belong to their supplied assignment.");W(s,a,n),W(r,a,n);let i=e.swappedModel?ke(e.swappedModel):null;if(i){if(i.reviewer.kind!=="model"||!k(i.assignmentRef,S(a)))throw new Error("Order-swap calibration receipts must be model receipts for the same assignment.");if(!k(i.reviewer.releaseRef,r.reviewer.releaseRef))throw new Error("Order-swap calibration must use the same immutable automated reviewer release.");W(i,a,n)}return{assignment:a,human:s,model:r,swappedModel:i}}m(ta,"validateCalibrationPair");function na(e,n,a){return!a||!Gn(a)?!1:a.passed&&k(a.comparisonRelease,S(n))&&k(a.automatedReviewer,e.reviewer.releaseRef)}m(na,"isCalibratedAutomatedReviewer");function Ut(e){return Object.fromEntries(e.candidates.map(n=>[n.attemptRef.id,0]))}m(Ut,"zeroScores");function re(e){return e.length?e.reduce((n,a)=>n+a,0)/e.length:0}m(re,"mean");function S(e){return{id:e.id,contentHash:e.contentHash}}m(S,"ref");function k(e,n){return e.id===n.id&&e.contentHash===n.contentHash}m(k,"sameRef");function $e(e){return[...new Map(e.map(n=>[`${n.id}:${n.contentHash}`,n])).values()]}m($e,"uniqueRefs");function aa(e,n,a){if(!k(e.harnessRelease,n.harnessRelease))return a.find(s=>k(s.tasksetRelease,e.tasksetRelease)&&(k(s.baseHarnessRelease,e.harnessRelease)&&k(s.candidateHarnessRelease,n.harnessRelease)||k(s.baseHarnessRelease,n.harnessRelease)&&k(s.candidateHarnessRelease,e.harnessRelease)))}m(aa,"compatibilityForRunManifests");function se(e,n,a){let s=a.safeParse(e);if(!s.success||!s.data)return!1;let{contentHash:r,...i}=s.data;try{return o(n.parse(i))===r}catch{return!1}}m(se,"verifyHashed");function G(e){if(!Fn(e))throw new Error("Preference Comparison Release has an invalid content hash.");return e}m(G,"verifyComparisonRelease");function ie(e){if(!Jn(e))throw new Error("Comparison Assignment has an invalid content hash.");return e}m(ie,"verifyAssignment");function ke(e){if(!Un(e))throw new Error("Preference Receipt has an invalid content hash.");return e}m(ke,"verifyReceipt");function ra(e){if(!Wn(e))throw new Error("Preference Aggregation Receipt has an invalid content hash.");return e}m(ra,"verifyAggregate");function sa(e){let{contentHash:n,...a}=ge.parse(e);if(o(he.parse(a))!==n)throw new Error("Taskset Release has an invalid content hash.");return e}m(sa,"verifyTasksetRelease");function ia(e){let n=ne.safeParse(e);if(!n.success)return!1;let{contentHash:a,...s}=n.data;return o(L.parse(s))===a}m(ia,"verifyArtifactManifest");function oa(e){let n=K.safeParse(e);if(!n.success)return!1;let{contentHash:a,...s}=n.data;return o(s)===a}m(oa,"verifyRunManifest");var ca=t.object({profileId:v,sourceRevision:t.string().trim().min(1).max(240),harnessRelease:g}).strict(),ma=t.object({kind:t.enum(["run","suite","comparison"]),id:v,contentHash:R}).strict(),Wt=t.object({schemaVersion:t.literal("openpond.profileEvaluationReport.v1"),id:v,createdAt:_,testedSources:t.array(ca).min(1).max(100),tasksetReleases:t.array(g).min(1).max(100),metricPolicyHashes:t.array(R).min(1).max(100),modelConfigurationHashes:t.array(R).min(1).max(100),evidence:t.array(ma).min(1).max(100),summary:t.object({totalChecks:t.number().int().nonnegative(),passedChecks:t.number().int().nonnegative(),failedChecks:t.number().int().nonnegative(),unscoredChecks:t.number().int().nonnegative(),meanScore:t.number().min(0).max(1).nullable()}).strict()}).strict().superRefine((e,n)=>{let{summary:a}=e;a.passedChecks+a.failedChecks+a.unscoredChecks!==a.totalChecks&&n.addIssue({code:"custom",path:["summary"],message:"Report check counts must add up."}),new Set(e.evidence.map(({kind:s,id:r})=>`${s}:${r}`)).size!==e.evidence.length&&n.addIssue({code:"custom",path:["evidence"],message:"Report evidence must be unique."})}),Gt=Wt.safeExtend({contentHash:R});function fs(e){let n=Wt.parse(e);return Gt.parse({...n,contentHash:o(n)})}m(fs,"createProfileEvaluationReport");function ps(e){let n=Gt.parse(e),{contentHash:a,...s}=n;if(o(s)!==a)throw new Error("Profile evaluation report content hash differs from its source bytes.");return n}m(ps,"assertProfileEvaluationReport");var la=t.string().regex(/^(0|-?[1-9][0-9]*)$/).max(20).refine(e=>{if(e.length>20||!/^(0|-?[1-9][0-9]*)$/.test(e))return!1;let n=BigInt(e);return n>=-(1n<<63n)&&n<1n<<63n},"SQL integers must fit signed 64-bit storage."),Qt=t.union([t.null(),t.string().max(262144),t.number().finite(),t.object({integer:la}).strict(),t.object({blobBase64:t.string().max(349528).regex(/^(?:[A-Za-z0-9+/]{4})*(?:[A-Za-z0-9+/]{2}==|[A-Za-z0-9+/]{3}=)?$/)}).strict()]),Zt=t.string().min(1).max(128).refine(e=>!e.includes("\0")),da=t.object({tables:t.array(t.object({name:Zt,columns:t.array(t.object({name:Zt,type:t.enum(["INTEGER","REAL","TEXT","BLOB"])}).strict()).min(1).max(128),rows:t.array(t.array(Qt).max(128)).max(4096)}).strict()).max(64)}).strict().superRefine((e,n)=>{let a=new Set;for(let[s,r]of e.tables.entries()){let i=r.name.toLowerCase();(a.has(i)||i.startsWith("sqlite_"))&&n.addIssue({code:"custom",path:["tables",s,"name"],message:"Table names must be unique and outside SQLite's reserved namespace."}),a.add(i),new Set(r.columns.map(u=>u.name.toLowerCase())).size!==r.columns.length&&n.addIssue({code:"custom",path:["tables",s,"columns"],message:"Column names must be unique."}),r.rows.some(u=>u.length!==r.columns.length)&&n.addIssue({code:"custom",path:["tables",s,"rows"],message:"Each row must match the declared columns."})}}),Ce=t.object({schemaVersion:t.literal("openpond.sqlExecution.v1"),snapshot:da,sql:t.string().min(1).max(65536).refine(e=>!e.includes("\0")),maxRows:t.number().int().min(1).max(4096),maxResultBytes:t.number().int().min(256).max(262144)}).strict(),ws=t.discriminatedUnion("status",[t.object({status:t.literal("completed"),columns:t.array(t.string()).max(128),rows:t.array(t.array(Qt).max(128)).max(4096)}).strict(),t.object({status:t.literal("rejected"),code:t.enum(["invalid_query","multiple_statements","result_too_large","memory_limit"])}).strict()]);function bs(e){le(e,1048576);let n=Ce.parse(e);if(new TextEncoder().encode(n.sql).byteLength>65536)throw new Error("sql_source_too_large");return n}m(bs,"assertSqlExecutionRequest");var Ie=t.object({scope:t.enum(["input","initialState","state","arguments"]),path:t.array(t.string().min(1).max(256)).max(16)}).strict(),Xt={id:t.string().min(1).max(128),operation:t.enum(["step","collect"]),toolName:t.string().min(1).max(500).optional(),timeoutMs:t.number().int().min(1).max(2e4)},Yt=t.discriminatedUnion("kind",[t.object({...Xt,kind:t.literal("sqlite.v1"),sql:Ie,snapshot:Ie,maxRows:Ce.shape.maxRows,maxResultBytes:Ce.shape.maxResultBytes}).strict(),t.object({...Xt,kind:t.literal("javascript.v1"),source:Ie,cases:Ie,exportName:t.string().regex(/^[A-Za-z_$][A-Za-z0-9_$]*$/).max(128),maxCases:t.number().int().min(1).max(32),maxResultBytes:t.number().int().min(256).max(262144)}).strict()]).superRefine((e,n)=>{e.operation==="step"!=(e.toolName!==void 0)&&n.addIssue({code:"custom",path:["toolName"],message:"Only step services require a declared tool name."}),e.operation==="collect"&&(e.kind==="sqlite.v1"?[e.sql,e.snapshot]:[e.source,e.cases]).some(s=>s.scope==="arguments")&&n.addIssue({code:"custom",message:"Collection has no tool arguments."})});var ua=t.object({schemaVersion:t.literal("openpond.javascriptEnvironment.v1"),id:t.string().trim().min(1).max(500),revision:t.number().int().positive(),module:J.refine(e=>e.visibility==="host_private","Environment code belongs to the execution owner."),tools:t.array(Ue).min(1).max(200),maxSteps:t.number().int().min(1).max(1e3),maxStateBytes:t.number().int().min(1).max(1048576),maxObservationBytes:t.number().int().min(1).max(262144),operationTimeoutMs:t.number().int().min(1).max(3e4),executionServices:t.array(Yt).min(1).max(16).optional()}).strict().superRefine((e,n)=>{new Set(e.tools.map(s=>s.name)).size!==e.tools.length&&n.addIssue({code:"custom",path:["tools"],message:"Environment tool names must be unique."}),e.tools.forEach((s,r)=>{We(s.inputSchema).valid||n.addIssue({code:"custom",path:["tools",r,"inputSchema"],message:"Environment tools require a supported JSON schema."}),s.inputSchemaHash!==o(s.inputSchema)&&n.addIssue({code:"custom",path:["tools",r,"inputSchemaHash"],message:"Tool schema identity differs from its declared schema."})});let a=e.executionServices??[];new Set(a.map(s=>s.id)).size!==a.length&&n.addIssue({code:"custom",path:["executionServices"],message:"Execution service identities must be unique."}),a.forEach((s,r)=>{let i=e.tools.find(u=>u.name===s.toolName);s.operation==="step"&&!i&&n.addIssue({code:"custom",path:["executionServices",r,"toolName"],message:"Execution services must target a declared tool."}),s.timeoutMs>=Math.min(e.operationTimeoutMs,i?.timeoutMs??e.operationTimeoutMs)&&n.addIssue({code:"custom",path:["executionServices",r,"timeoutMs"],message:"Service budgets must leave time for the authored controller."})})}),fa=ua.safeExtend({contentHash:t.string().regex(/^[a-f0-9]{64}$/)}),As=t.enum(["create","reset","step","collect","destroy"]),Ms=t.object({state:t.record(t.string(),t.unknown()),observation:t.record(t.string(),t.unknown())}).strict();function js(e){le(e,4194304);let n=fa.parse(e),{contentHash:a,...s}=n;if(o(s)!==a)throw new Error("environment_definition_hash_mismatch");return structuredClone(n)}m(js,"assertJavaScriptEnvironmentDefinition");export{Rn as a,vn as b,Se as c,bn as d,ht as e,Sn as f,tr as g,nr as h,K as i,Pe as j,be as k,cn as l,Ra as m,va as n,D as o,mt as p,fn as q,z as r,Ta as s,pn as t,Oa as u,Ve as v,lr as w,gt as x,dr as y,xe as z,dt as A,Ht as B,_t as C,Cr as D,Ir as E,Et as F,ze as G,yn as H,Ct as I,kn as J,ne as K,It as L,F as M,Be as N,Nr as O,$r as P,Dr as Q,Lr as R,Fr as S,Jr as T,Kr as U,Ur as V,Wr as W,Gr as X,Tt as Y,zt as Z,$t as _,Jt as $,rs as aa,ss as ba,Jn as ca,Le as da,Kn as ea,is as fa,Un as ga,os as ha,cs as ia,hn as ja,La as ka,Gt as la,fs as ma,ps as na,gn as oa,Za as pa,ws as qa,bs as ra,Yt as sa,fa as ta,As as ua,Ms as va,js as wa};