@zanii/blackbox 0.21.0 → 0.21.1

package/dist/analysis/index.d.ts

@@ -103,4 +103,4 @@ export declare function verifyRegression(events: Ev[]): Finding[];
 /** Stage 2 R4, judged at close: a run labelled (or claimed) a success that its own record contradicts:
  * a check's last result failed, or the last tool call failed with nothing succeeding after it. */
 export declare function falseSuccess(events: Ev[], now: number): Finding[];
-export { type Growth, growth, isoWeek, type Lessons, lessons, POWER_KINDS, promptChanged, SELF_CHANGE_KINDS, selfChanges, systemPromptHash, toolsHash, } from "./learning.ts";
+export { type Growth, growth, isoWeek, type Lessons, lessons, POWER_KINDS, promptChanged, SELF_CHANGE_KINDS, selfChanges, successRange, systemPromptHash, toolsHash, } from "./learning.ts";

package/dist/analysis/index.js

@@ -586,4 +586,4 @@ export function falseSuccess(events, now) {
             : "The agent claimed it was done, but its last tool call failed and nothing succeeded after it.");
     return [];
 }
-export { growth, isoWeek, lessons, POWER_KINDS, promptChanged, SELF_CHANGE_KINDS, selfChanges, systemPromptHash, toolsHash, } from "./learning.js";
+export { growth, isoWeek, lessons, POWER_KINDS, promptChanged, SELF_CHANGE_KINDS, selfChanges, successRange, systemPromptHash, toolsHash, } from "./learning.js";

package/dist/analysis/learning.d.ts

@@ -86,6 +86,10 @@ export interface Growth {
         spend_per_run_micro_usd: number;
         human_steps: number;
         declared: boolean;
+        success_low_ppm: number;
+        success_high_ppm: number;
+        vs_previous: "better" | "worse" | "unclear" | null;
+        certificate: string | null;
     }>;
     versions_total: number;
     learned: Array<{
@@ -98,9 +102,15 @@ export interface Growth {
         corrected: number;
         unknown: number;
         checked: boolean;
+        certificate: string | null;
     }>;
     hints: string[];
 }
+/**
+ * spec/fleet.md §9: the 95% Wilson interval of `s` successes in `n` runs, in ppm (low rounded down,
+ * high up). The same IEEE operations in the same order in both SDKs, so the digits agree.
+ */
+export declare function successRange(s: number, n: number): [number, number];
 /**
  * spec/fleet.md §9: how one agent (its label) grows, from its closed sessions' summaries, newest
  * first. Autonomy per ISO week (a run done alone is a success with no human step, shown next to the

package/dist/analysis/learning.js

@@ -219,6 +219,24 @@ export function isoWeek(ts) {
     return `${year}-W${String(week).padStart(2, "0")}`;
 }
 const ppm = (n, d) => (d ? Math.floor((n * 1_000_000) / d) : 0);
+/**
+ * spec/fleet.md §9: the 95% Wilson interval of `s` successes in `n` runs, in ppm (low rounded down,
+ * high up). The same IEEE operations in the same order in both SDKs, so the digits agree.
+ */
+export function successRange(s, n) {
+    if (!n)
+        return [0, 1_000_000];
+    const z = 1.96;
+    const z2 = z * z;
+    const p = s / n;
+    const denom = 1 + z2 / n;
+    const centre = (p + z2 / (2 * n)) / denom;
+    const half = (z * Math.sqrt((p * (1 - p)) / n + z2 / (4 * n * n))) / denom;
+    return [
+        Math.max(0, Math.floor((centre - half) * 1_000_000)),
+        Math.min(1_000_000, Math.ceil((centre + half) * 1_000_000)),
+    ];
+}
 const VERSIONS = 20;
 const LEARNED = 50;
 /**
@@ -269,6 +287,10 @@ export function growth(summaries, label, options = {}) {
         .map((ss) => {
         const v = ss[0]?.version;
         const starts = ss.map((s) => s.started_at).sort();
+        const wins = ss.filter((s) => s.outcome === "success").length;
+        const [low, high] = successRange(wins, ss.length);
+        // the newest run's certificate first: ss is newest first
+        const cert = ss.flatMap((s) => (s.self_changes ?? []).map((c) => c.certificate)).find(Boolean) ?? null;
         return {
             id: v.id,
             prompt: v.prompt,
@@ -278,16 +300,27 @@ export function growth(summaries, label, options = {}) {
             sessions: ss.length,
             first_seen: starts[0],
             last_seen: starts[starts.length - 1],
-            success_ppm: ppm(ss.filter((s) => s.outcome === "success").length, ss.length),
+            success_ppm: ppm(wins, ss.length),
             spend_per_run_micro_usd: Math.floor(ss.reduce((n, s) => n + s.spend_micro_usd, 0) / ss.length),
             human_steps: ss.reduce((n, s) => n + (s.human_steps ?? 0), 0),
             declared: ss.some((s) => (s.self_changes?.length ?? 0) > 0),
+            success_low_ppm: low,
+            success_high_ppm: high,
+            vs_previous: null,
+            certificate: cert,
         };
     })
         .sort((a, b) => a.first_seen < b.first_seen ? -1 : a.first_seen > b.first_seen ? 1 : a.id < b.id ? -1 : 1);
     for (let i = 1; i < all.length; i++) {
         const [p, v] = [all[i - 1], all[i]];
         v.changed = ["prompt", "tools", "model"].filter((k) => p[k] !== v[k]);
+        // only a difference outside both ranges counts: two runs each tell nothing
+        v.vs_previous =
+            v.success_low_ppm > p.success_high_ppm
+                ? "better"
+                : v.success_high_ppm < p.success_low_ppm
+                    ? "worse"
+                    : "unclear";
     }
     const decided = new Map();
     for (const s of runs)
@@ -312,6 +345,7 @@ export function growth(summaries, label, options = {}) {
                 corrected,
                 unknown,
                 checked: found.length > 0 && unknown === 0,
+                certificate: c.certificate ?? null,
             });
         }
     const hints = [];

package/dist/fleet/index.d.ts

@@ -53,9 +53,14 @@ export interface SelfChangeEntry {
     kind: string;
     approval_id: string | null;
     taught_by: string[];
+    /** A Zanii Evolve certificate's digest behind the change, when the agent names one */
+    certificate?: string;
 }
 /** An approval id: what a lesson may point at, and nothing else (spec/fleet.md §9). */
 export declare const APPROVAL_ID: RegExp;
+/** spec/fleet.md §9: a Zanii Evolve certificate's digest (`sha256:` + 64 hex, or the 64 hex its
+ * certificate carries), as `sha256:<hex>`; null for anything else. */
+export declare function certificateOf(v: unknown): string | null;
 /** spec/fleet.md §9: the approval ids a self-change names as what taught it, at most 20. */
 export declare function taughtBy(v: unknown): string[];
 export declare function sessionSummary(lines: readonly string[], bodies: Bodies, prices: LoadedPrices): Summary;

package/dist/fleet/index.js

@@ -12,6 +12,14 @@ import { callsOf } from "../reconcile/record.js";
 import { canonical } from "../reconcile/shared.js";
 /** An approval id: what a lesson may point at, and nothing else (spec/fleet.md §9). */
 export const APPROVAL_ID = /^apr_[0-9a-f]{18}$/;
+/** spec/fleet.md §9: a Zanii Evolve certificate's digest (`sha256:` + 64 hex, or the 64 hex its
+ * certificate carries), as `sha256:<hex>`; null for anything else. */
+export function certificateOf(v) {
+    if (typeof v !== "string")
+        return null;
+    const m = /^(?:sha256:)?([0-9a-f]{64})$/.exec(v);
+    return m ? `sha256:${m[1]}` : null;
+}
 /** spec/fleet.md §9: the approval ids a self-change names as what taught it, at most 20. */
 export function taughtBy(v) {
     if (!Array.isArray(v))
@@ -127,6 +135,9 @@ export function sessionSummary(lines, bodies, prices) {
             kind: typeof d.kind === "string" && d.kind ? d.kind : "other",
             approval_id: typeof d.approval_id === "string" ? d.approval_id : null,
             taught_by: taughtBy(d.taught_by),
+            ...(certificateOf(d.certificate)
+                ? { certificate: certificateOf(d.certificate) }
+                : {}),
         };
     });
     const version = versionOf(events);

package/dist/index.d.ts

@@ -2,7 +2,7 @@ export { A2A_STREAMING, type AgentCardReport, a2aMethod, a2aRequestMeta, a2aResp
 export { causality, memoryXray } from "./agents/index.ts";
 export { type DetectorOptions, detectors, fingerprint } from "./analysis/detectors.ts";
 export { FAULTS, type Fault, faultOf } from "./analysis/faults.ts";
-export { type Accuracy, type AnalyzeOptions, type AnswerClaims, analyze, answerClaims, answerCredential, answerRisk, answersOf, answerText, type ClaimEntry, type ClaimVerdict, type CriterionResult, canonicalNumber, checkAgainstPacks, checkFlightPlan, checkpointReplay, detectorAccuracy, dualWitness, Evidence, eventsOf, type Fact, type Finding as AnalysisFinding, factsIn, findingKey, type GroundedClaim, groundingFindings, groundingRequest, growth, HOLD_TRIGGERS, identifiersIn, isoWeek, judgeFindings, judgeRequest, type Label as AccuracyLabel, type Landing, landing, lessons, loadReferencePacks, modelGroundingFindings, normaliseValue, OUTCOMES, type Outcome, type OutcomeRollup, packsFor, parseGroundingAnswer, parseJudgeAnswer, promptChanged, type ReferenceFact, type ReferencePack, type ReferenceVerdict, referenceContext, rollup, type Severity, schemaProblem, secretSent, selfChanges, systemPromptHash, toolDrifted, toolHallucinations, toolsHash, userDownvotes, verifyAnswerCredential, type WasteReason, type WasteReport, waste, wilsonLow, } from "./analysis/index.ts";
+export { type Accuracy, type AnalyzeOptions, type AnswerClaims, analyze, answerClaims, answerCredential, answerRisk, answersOf, answerText, type ClaimEntry, type ClaimVerdict, type CriterionResult, canonicalNumber, checkAgainstPacks, checkFlightPlan, checkpointReplay, detectorAccuracy, dualWitness, Evidence, eventsOf, type Fact, type Finding as AnalysisFinding, factsIn, findingKey, type GroundedClaim, groundingFindings, groundingRequest, growth, HOLD_TRIGGERS, identifiersIn, isoWeek, judgeFindings, judgeRequest, type Label as AccuracyLabel, type Landing, landing, lessons, loadReferencePacks, modelGroundingFindings, normaliseValue, OUTCOMES, type Outcome, type OutcomeRollup, packsFor, parseGroundingAnswer, parseJudgeAnswer, promptChanged, type ReferenceFact, type ReferencePack, type ReferenceVerdict, referenceContext, rollup, type Severity, schemaProblem, secretSent, selfChanges, successRange, systemPromptHash, toolDrifted, toolHallucinations, toolsHash, userDownvotes, verifyAnswerCredential, type WasteReason, type WasteReport, waste, wilsonLow, } from "./analysis/index.ts";
 export { injectionFindings, injectionRule } from "./analysis/injection.ts";
 export { MITRE_ATLAS, OWASP_ASI, TAXONOMY, type Taxonomy, taxonomyOf, } from "./analysis/taxonomy.ts";
 export { type ApprovalFinding, approvalConfirmation, approvalFindings, argsHash, verifyConfirmation, } from "./approvals/index.ts";
@@ -30,7 +30,7 @@ export { openSegments, type SealedSegment } from "./encryption/index.ts";
 export { type EvidencePack, type EvidencePackOptions, evidencePack, renderEvidencePack, } from "./evidence/index.ts";
 export { buildEquivocationProof, compareViews, cosignLedgerSth, verifyEquivocationProof, verifyLedgerCosignature, } from "./federation/index.ts";
 export { buildOwnerAttestation, buildPayment, CURRENCIES, FTA_WALL_ID, financeFindings, formatMoney, ftaPayload, parseMoney, paymentsOf, screenCounterparty, screeningPayload, verifyOwnerAttestation, verifyOwnerBinding, } from "./finance/index.ts";
-export { APPROVAL_ID, aggregateFindings, airworthiness, type Collision, collisions, didFromSeed, type FindingGroup, incidentReport, logbook, type Summary, sessionSummary, signatureOf, signatures, signCertificate, taughtBy, verifyCertificate, } from "./fleet/index.ts";
+export { APPROVAL_ID, aggregateFindings, airworthiness, type Collision, certificateOf, collisions, didFromSeed, type FindingGroup, incidentReport, logbook, type Summary, sessionSummary, signatureOf, signatures, signCertificate, taughtBy, verifyCertificate, } from "./fleet/index.ts";
 export { copilotContext, copilotPrompt, type Insights, insights, parseCopilotAnswer, } from "./fleet/insights.ts";
 export { type AppealItem, type AppealPack, appealPack, decisionPayload, GOV_DECISION_KINDS, renderAppealPack, ruleConsistency, signSubjectClaim, subjectTag, verifyDecision, verifyFactors, verifySubjectClaim, } from "./gov/index.ts";
 export { accessPayload, breakGlass, checkEpisodeTag, clinicianConfirmation, type EpisodeItem, episodeSeed, episodesPack, episodeTag, healthFindings, protocolTrail, recommendationHash, recommendationPayload, verifyCommitment, verifyHealthSignature, } from "./health/index.ts";

package/dist/index.js

@@ -4,7 +4,7 @@ export { A2A_STREAMING, a2aMethod, a2aRequestMeta, a2aResponseMeta, a2aRestMetho
 export { causality, memoryXray } from "./agents/index.js";
 export { detectors, fingerprint } from "./analysis/detectors.js";
 export { FAULTS, faultOf } from "./analysis/faults.js";
-export { analyze, answerClaims, answerCredential, answerRisk, answersOf, answerText, canonicalNumber, checkAgainstPacks, checkFlightPlan, checkpointReplay, detectorAccuracy, dualWitness, Evidence, eventsOf, factsIn, findingKey, groundingFindings, groundingRequest, growth, HOLD_TRIGGERS, identifiersIn, isoWeek, judgeFindings, judgeRequest, landing, lessons, loadReferencePacks, modelGroundingFindings, normaliseValue, OUTCOMES, packsFor, parseGroundingAnswer, parseJudgeAnswer, promptChanged, referenceContext, rollup, schemaProblem, secretSent, selfChanges, systemPromptHash, toolDrifted, toolHallucinations, toolsHash, userDownvotes, verifyAnswerCredential, waste, wilsonLow, } from "./analysis/index.js";
+export { analyze, answerClaims, answerCredential, answerRisk, answersOf, answerText, canonicalNumber, checkAgainstPacks, checkFlightPlan, checkpointReplay, detectorAccuracy, dualWitness, Evidence, eventsOf, factsIn, findingKey, groundingFindings, groundingRequest, growth, HOLD_TRIGGERS, identifiersIn, isoWeek, judgeFindings, judgeRequest, landing, lessons, loadReferencePacks, modelGroundingFindings, normaliseValue, OUTCOMES, packsFor, parseGroundingAnswer, parseJudgeAnswer, promptChanged, referenceContext, rollup, schemaProblem, secretSent, selfChanges, successRange, systemPromptHash, toolDrifted, toolHallucinations, toolsHash, userDownvotes, verifyAnswerCredential, waste, wilsonLow, } from "./analysis/index.js";
 export { injectionFindings, injectionRule } from "./analysis/injection.js";
 export { MITRE_ATLAS, OWASP_ASI, TAXONOMY, taxonomyOf, } from "./analysis/taxonomy.js";
 export { approvalConfirmation, approvalFindings, argsHash, verifyConfirmation, } from "./approvals/index.js";
@@ -32,7 +32,7 @@ export { openSegments } from "./encryption/index.js";
 export { evidencePack, renderEvidencePack, } from "./evidence/index.js";
 export { buildEquivocationProof, compareViews, cosignLedgerSth, verifyEquivocationProof, verifyLedgerCosignature, } from "./federation/index.js";
 export { buildOwnerAttestation, buildPayment, CURRENCIES, FTA_WALL_ID, financeFindings, formatMoney, ftaPayload, parseMoney, paymentsOf, screenCounterparty, screeningPayload, verifyOwnerAttestation, verifyOwnerBinding, } from "./finance/index.js";
-export { APPROVAL_ID, aggregateFindings, airworthiness, collisions, didFromSeed, incidentReport, logbook, sessionSummary, signatureOf, signatures, signCertificate, taughtBy, verifyCertificate, } from "./fleet/index.js";
+export { APPROVAL_ID, aggregateFindings, airworthiness, certificateOf, collisions, didFromSeed, incidentReport, logbook, sessionSummary, signatureOf, signatures, signCertificate, taughtBy, verifyCertificate, } from "./fleet/index.js";
 export { copilotContext, copilotPrompt, insights, parseCopilotAnswer, } from "./fleet/insights.js";
 export { appealPack, decisionPayload, GOV_DECISION_KINDS, renderAppealPack, ruleConsistency, signSubjectClaim, subjectTag, verifyDecision, verifyFactors, verifySubjectClaim, } from "./gov/index.js";
 export { accessPayload, breakGlass, checkEpisodeTag, clinicianConfirmation, episodeSeed, episodesPack, episodeTag, healthFindings, protocolTrail, recommendationHash, recommendationPayload, verifyCommitment, verifyHealthSignature, } from "./health/index.js";

package/dist/session/index.d.ts

@@ -390,6 +390,8 @@ export declare class BlackboxSession {
         approvalId?: string;
         /** The approvals (ids) that taught it this, and nothing else: never content (fleet.md §9). */
         taughtBy?: readonly string[];
+        /** A Zanii Evolve certificate's digest behind the change (spec/fleet.md §9). */
+        certificate?: string;
     }): void;
     /** spec/findings.md §13.1: what in an answer (the latest, or the one at `seq`) would hold it,
      * for an app that streams the answer by itself: `{seq, risks, hold}`. `hold` says whether the

package/dist/session/index.js

@@ -10,7 +10,7 @@ import { SELF_CHANGE_KINDS } from "../analysis/learning.js";
 import { argsHash } from "../approvals/index.js";
 import { attest, isReadOnly, parseCommand } from "../attest/index.js";
 import { buildPayment, ftaPayload, screeningPayload, } from "../finance/index.js";
-import { taughtBy } from "../fleet/index.js";
+import { certificateOf, taughtBy } from "../fleet/index.js";
 import { decisionPayload } from "../gov/index.js";
 import { accessPayload, checkEpisodeTag, recommendationPayload, verifyHealthSignature, } from "../health/index.js";
 import { memoryEntry } from "../memory/index.js";
@@ -635,6 +635,9 @@ export class BlackboxSession {
             ...(options.reason ? { reason: options.reason.slice(0, 500) } : {}),
             ...(options.approvalId ? { approval_id: options.approvalId.slice(0, 128) } : {}),
             ...(taught.length ? { taught_by: taught } : {}),
+            ...(certificateOf(options.certificate)
+                ? { certificate: certificateOf(options.certificate) }
+                : {}),
         });
     }
     /** spec/findings.md §13.1: what in an answer (the latest, or the one at `seq`) would hold it,

package/dist/version.d.ts

@@ -1 +1 @@
-export declare const VERSION = "0.21.0";
+export declare const VERSION = "0.21.1";

package/dist/version.js

@@ -1,2 +1,2 @@
 // The SDK version, on its own so the CLI can print it without loading the whole SDK.
-export const VERSION = "0.21.0";
+export const VERSION = "0.21.1";

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@zanii/blackbox",
   "license": "Apache-2.0",
-  "version": "0.21.0",
+  "version": "0.21.1",
   "description": "The flight recorder for AI agents: sessions, a zero-loss spool, framework hooks, approvals, and offline verification of the gateway's hash-chained record.",
   "keywords": [
     "ai-agents",