@tangle-network/agent-runtime 0.285.0 → 0.286.0

package/dist/testing.d.ts

@@ -1,5 +1,5 @@
-import { Ac as SuperviseTestOptions, Mc as superviseWithTestBrain, Tn as driverAgent, Zt as RunGraphTestOptions, tn as runGraphWithTestBrain, ul as SupervisorAgentTestDeps, wn as DriverAgentOptions, yl as supervisorAgentWithTestBrain } from "./index-B7PfLDfI.js";
-import { Gn as ToolLoopCallContext, Kn as ToolLoopChat } from "./stream-agent-turn-IIqBKeLB.js";
+import { Ac as SuperviseTestOptions, Mc as superviseWithTestBrain, Tn as driverAgent, Zt as RunGraphTestOptions, tn as runGraphWithTestBrain, ul as SupervisorAgentTestDeps, wn as DriverAgentOptions, yl as supervisorAgentWithTestBrain } from "./index-BspUvwtx.js";
+import { Kn as ToolLoopCallContext, qn as ToolLoopChat } from "./stream-agent-turn-CaLVbKSY.js";
 import { AgentImprovementProposal, AgentProfile, AgentProfileImprovementMeasuredComparison, SandboxSizePreset } from "@tangle-network/agent-interface";
 //#region src/testing/index.d.ts
 /** A proposal produced by Runtime's opaque profile-improvement path. */

package/dist/testing.js

@@ -1,14 +1,14 @@
 import { O as immutableCandidateValue, w as canonicalCandidateDigest$1 } from "./workspace-archive-MhowMCmp.js";
-import { f as verifyAgentImprovementProposal } from "./improvement-cycle-DxBnEP8q.js";
+import { f as verifyAgentImprovementProposal } from "./improvement-cycle-Cm_YPUwJ.js";
 import { i as applyExactAgentProfileDiff, p as parseExactAgentProfile } from "./profile-C4NDjZyB.js";
-import { r as driverAgent } from "./coordination-driver-C3bIMg7f.js";
-import { a as superviseWithTestBrain, h as supervisorAgentWithTestBrain } from "./supervise-tLnNRbmr.js";
-import { i as runGraphWithTestBrain } from "./graph-CpfmPttW.js";
+import { r as driverAgent } from "./coordination-driver-CLqj3s53.js";
+import { a as superviseWithTestBrain, h as supervisorAgentWithTestBrain } from "./supervise-Dgfve4nM.js";
+import { i as runGraphWithTestBrain } from "./graph-to6d8CPK.js";
 import { SANDBOX_SIZE_PRESET_NAMES, canonicalAgentProfileDigest } from "@tangle-network/agent-interface";
 //#region src/testing/fixtures/agent-improvement-proposal.json
 var agent_improvement_proposal_default = {
 	changedSurfaces: ["prompt"],
-	digest: "sha256:b5c60e8dcf6641eb5243f503a57a86ebd47b7211cc4a36964574f6692b379f83",
+	digest: "sha256:0bdb2ee4bdd9dbd1423a6b9460a239d360b287e9a999fc9de112abbcd6d9a179",
 	evaluation: {
 		"decision": {
 			"contributingChecks": [
@@ -4579,7 +4579,7 @@ var agent_improvement_proposal_default = {
 		],
 		"metadata": {
 			"fixture": "agent-improvement-proposal",
-			"runtimeVersion": "0.285.0"
+			"runtimeVersion": "0.286.0"
 		},
 		"objectives": [
 			{
@@ -4690,8 +4690,8 @@ var agent_improvement_proposal_default = {
 			"baselineContentHash": "sha256:5c21ee53e513fc604cb09754e21c392b24a424da0ef37dbf8f1ee4a8a0b08f09",
 			"candidateContentHash": "sha256:60fcbb1c728194bd51d7d19cb732d1c3f1881dce7e0a6266b41c8b98cfd65693",
 			"kind": "agent-eval-loop",
-			"recordDigest": "sha256:b495d730103154f19c230d966826424ce8353241fe284a793a244672f4d03490",
-			"runId": "agent-runtime-0.285.0-proposal-fixture",
+			"recordDigest": "sha256:62fe853c8ba86a49e7297389cd7d045f53a98b396500db9160aa94b2a3da9969",
+			"runId": "agent-runtime-0.286.0-proposal-fixture",
 			"schema": "agent-candidate-experiment"
 		}
 	},
@@ -4714,13 +4714,13 @@ var agent_improvement_proposal_default = {
 	}],
 	kind: "agent-improvement-proposal",
 	proposedAt: "2026-07-10T01:00:00.000Z",
-	runId: "agent-runtime-0.285.0-proposal-fixture"
+	runId: "agent-runtime-0.286.0-proposal-fixture"
 };
 //#endregion
 //#region src/testing/fixtures/agent-profile-improvement-proposal.json
 var agent_profile_improvement_proposal_default = {
 	changedSurfaces: ["prompt", "skills"],
-	digest: "sha256:05963dbc17686c4dc0616d2976e5e7295e56ce9c332b69cef075f76902061263",
+	digest: "sha256:ed9681233c29af4fb29019b51588ff5140844b92b0327fa9227275d0fb6a0847",
 	evaluation: {
 		"decision": {
 			"contributingChecks": [
@@ -6354,7 +6354,7 @@ var agent_profile_improvement_proposal_default = {
 		],
 		"metadata": {
 			"fixture": "agent-profile-improvement-proposal",
-			"runtimeVersion": "0.285.0"
+			"runtimeVersion": "0.286.0"
 		},
 		"objectives": [
 			{
@@ -6465,7 +6465,7 @@ var agent_profile_improvement_proposal_default = {
 			"baselineContentHash": "sha256:21c495a37c418c10bde64fbaa188beddeed31f1f051ea60a6a6582a9ee0db704",
 			"candidateContentHash": "sha256:103f77bc8481601eef1ad5fe6ba84a40dffabc3a44f421f8c8559121edab84e9",
 			"kind": "agent-eval-loop",
-			"recordDigest": "sha256:6b647cce14372fd9a799ef9a41891968c701111bb2a3e42aae0a252d3d7201aa",
+			"recordDigest": "sha256:e2bdfd2d42a76497049c35a618ef9511a07cb558e1622b62541fc0b63da94798",
 			"runId": "profile-improvement-1",
 			"schema": "agent-profile-improvement-experiment"
 		}

package/dist/tui/index.d.ts

@@ -1,4 +1,4 @@
-import { Ct as ProvisionSupervisorConnection, Dt as provisionSupervisor, Et as SupervisorCleanupReceipt, Tt as ProvisionedSupervisor, wt as ProvisionSupervisorRequest } from "../index-B7PfLDfI.js";
+import { Ct as ProvisionSupervisorConnection, Dt as provisionSupervisor, Et as SupervisorCleanupReceipt, Tt as ProvisionedSupervisor, wt as ProvisionSupervisorRequest } from "../index-BspUvwtx.js";
 //#region src/tui/top-app.d.ts
 /**
  * The interactive side of the supervisor-run TUI: a keypress/mouse loop over the frames

package/dist/tui/index.js

@@ -1,3 +1,3 @@
-import { t as provisionSupervisor } from "../provision-supervisor-jnArBa8T.js";
+import { t as provisionSupervisor } from "../provision-supervisor-CoXCyrZX.js";
 import { a as renderTopFrameWithLayout, i as renderTopFrame, n as runTopApp, r as loadTopSnapshot, t as renderTopOnce } from "../top-app-DaiR4LXM.js";
 export { loadTopSnapshot, provisionSupervisor, renderTopFrame, renderTopFrameWithLayout, renderTopOnce, runTopApp };

package/dist/{types-D56jQad-.d.ts → types-wIVzqXdQ.d.ts}

@@ -2267,6 +2267,8 @@ interface ContinuationNoteInput {
   readonly progress: DriverProgressMark;
   /** Whether `read_continuation` is served to this manager. */
   readonly canReadMore: boolean;
+  /** Whether submit_result is served to this manager. Absent means unavailable. */
+  readonly canSubmitResult?: boolean;
   /** Whether the check's score comes from cases the director cannot see. */
   readonly sealed?: boolean;
 }
@@ -5818,4 +5820,4 @@ interface WidenGate<Out> {
 }
 //#endregion
 export { SpawnJournal as $, RuntimeStreamEvent as $a, Iteration as $i, ReadableEnvironment as $n, RetainedInteractiveBindingError as $r, RuntimeStreamEventCollector as $t, NodeExecutionIdentity as A, RuntimeRunStatus as Aa, RetainedRunStartMaterial as Ai, RetainedPendingCause as An, ContinuationPanelInput as Ar, TraceContext as At, ResultBlobStore as B, AgentTaskRunResult as Ba, ReservationHolder as Bi, ReconnectRetainedInteractiveRunOptions as Bn, distillFindings as Br, OtelSpan as Bt, ExecutorToolCall as C, Validator as Ca, RetainedRunEffect as Ci, SessionTraceBox as Cn, CheckUnavailableError as Cr, WorkspaceCheckpointMarker as Ct, MaterializedExecutionIdentity as D, RuntimeRunOptions as Da, RetainedRunIntentAdmission as Di, decodeToolPart as Dn, ContinuationEntry as Dr, workerTraceEnv as Dt, HeldEnvironment as E, RuntimeRunHandle as Ea, RetainedRunHandle as Ei, createPushTraceSource as En, ContinuationContext as Er, readWorkerTraceContext as Et, ProviderModelAttemptEvidence as F, AgentExecutionBackend as Fa, BudgetPoolRestore as Fi, ScopeProgressInput as Fn, VerdictSummary as Fr, LoopSpanNode as Ft, RootMaterialization as G, OpenAIChatResponseFormat as Ga, ReservationWaitRefused as Gi, RetainedInteractiveStartMaterial as Gn, AgentEvalError$1 as Gr, createOpenInferenceFileExporter as Gt, ResumedWork as H, AgentTaskStatus as Ha, ReservationShortfall as Hi, RetainedInteractiveAdmissionHook as Hn, failedItems as Hr, buildLoopOtelSpans as Ht, ProviderModelExecutionEvidence as I, AgentKnowledgeProvider as Ia, BudgetReadout as Ii, TeamProgress as In, admitContinuationPolicy as Ir, OtelAttribute as It, RootStreamReceipt as J, RunAgentTaskOptions as Ja, spendFromUsageEvents as Ji, HarnessTranscriptCapture as Jn, ConfigError as Jr, loopEventToOtelSpan as Jt, RootProviderModelEvidence as K, OpenAIChatTool as Ka, ReserveOptions as Ki, StartRetainedInteractiveRunOptions as Kn, AgentEvalErrorCode as Kr, createOtelExporter as Kt, RecursiveReservationPolicy as L, AgentRuntimeEvent as La, BudgetReconcileFault as Li, WorkerProgress as Ln, admitFinding as Lr, OtelExportConfig as Lt, NodeSnapshot as M, AgentAdapter as Ma, StartRetainedRunInEnvironmentOptions as Mi, ActivityNote as Mn, ContinuationPolicy as Mr, mergeTraceEnv as Mt, NodeStatus as N, AgentBackendContext as Na, StartRetainedRunOptions as Ni, DEFAULT_STALL_AFTER_MS as Nn, ContinuationProfile as Nr, readTraceContextFromEnv as Nt, MaterializedModelIdentity as O, RuntimeRunPersistenceAdapter as Oa, RetainedRunReplayPoint as Oi, sandboxSessionTraceSource as On, ContinuationNoteInput as Or, workerTraceHeaders as Ot, ProfileMaterializationReceipt as P, AgentBackendInput as Pa, BudgetPool as Pi, ExecutorProgress as Pn, PanelFinding as Pr, traceContextToEnv as Pt, SpawnEvent as Q, RuntimeSessionStore as Qa, ExecCtx as Qi, HarnessTranscriptUnavailableReason as Qn, RetainedInteractiveAdmissionError as Qr, RuntimeEventCollector as Qt, ResourceLimit as R, AgentRuntimeEventSink as Ra, LeakedReservation as Ri, createActivityLog as Rn, checkVerdictOf as Rr, OtelExportStats as Rt, ExecutorTeardownWarning as S, ValidationCtx as Sa, RetainedRunDispatchedAdmission as Si, SessionMessageLike as Sn, CheckRead as Sr, WorkerTraceUnavailableReason as St, Handle as T, RuntimeRunCost as Ta, RetainedRunEventOptions as Ti, TraceSource as Tn, ContinuationAppend as Tr, WorkerTraceSeamCarrier as Tt, RetainedExecutionState as U, BackendErrorDetail as Ua, ReservationStage as Ui, RetainedInteractiveEnvironmentInput as Un, passedItems as Ur, buildLoopSpanNodes as Ut, ResumedKeyState as V, AgentTaskSpec as Va, ReservationRejection as Vi, RecoverRetainedInteractiveRunOptions as Vn, expandQuestions as Vr, RuntimeEventOtelOptions as Vt, RootHandle as W, KnowledgeReadinessDecision as Wa, ReservationTicket as Wi, RetainedInteractiveRunHandle as Wn, verdictFromJudgeScore as Wr, buildRuntimeEventOtelSpans as Wt, Scope as X, RuntimeCanonicalStreamEvent as Xa, DefaultVerdict$1 as Xi, HarnessTranscriptFile as Xn, NotFoundError as Xr, padTraceId as Xt, Runtime as Y, RunAgentTaskStreamOptions as Ya, AgentRunSpec as Yi, HarnessTranscriptEvidence as Yn, JudgeError as Yr, padSpanId as Yt, Settled as Z, RuntimeSession as Za, Driver as Zi, HarnessTranscriptUnavailable as Zn, PlannerError as Zr, toOtelAttributes as Zt, ExecutorMaterialization as _, MountRecorder as _a, RetainedInteractiveStartedAdmission as _i, isWaitOutcome as _n, upstreamUnavailableSignal as _r, WaitOpts as _t, BudgetOverspend as a, LoopLineageOptions as aa, WorkerSlots as ai, createRuntimeEventCollector as an, DriverAttemptsExhaustedError as ar, SpendGap as at, ExecutorRegistry as b, SandboxClient as ba, RetainedRunCancelOptions as bi, validateWaitSpec as bn, CONTINUATIONS_DIR as br, WorkerInteractiveUnavailableReason as bt, EnvironmentTeardownReceipt as c, LoopResult as ca, NativeContextContinuationHandle as ci, sanitizeKnowledgeReadinessReport as cn, DriverContractState as cr, SubtreeSummary as ct, ExecutorAccounting as d, LoopTeardownFailedPayload as da, RecoverRetainedRunIntentOptions as di, WaitOutcome as dn, DriverProgressMark as dr, SupervisorOpts as dt, LoopDecisionPayload as ea, RetainedRunAdmissionError as ei, RuntimeStreamEventSink as en, captureHarnessTranscript as er, SpawnOpts as et, ExecutorCancellation as f, LoopTokenUsage as fa, RecoverRetainedRunOptions as fi, WaitProbe as fn, DriverReentry as fr, TokenUsageProvenance as ft, ExecutorFactory as g, MountManifestEntry as ga, RetainedInteractiveIntentAdmission as gi, createWaitProbes as gn, summarizeDriverAttempts as gr, UsageEvent as gt, ExecutorExecutionBinding as h, LoopWinner as ha, RetainedInteractiveEnvironmentAdmission as hi, WaitSpec as hn, classifyDriverFailure as hr, UnknownMaterializationReason as ht, Budget as i, LoopIterationStartedPayload as ia, SlotPermit as ii, SanitizedKnowledgeRequirement as in, DriverAttemptStop as ir, SpendChannel as it, NodeId as j, startRuntimeRun as ja, RetainedRunTurnInput as ji, ActivityLog as jn, ContinuationPanelResult as jr, createPropagatingTraceEmitter as jt, NoWinnerError as k, RuntimeRunRow as ka, RetainedRunSnapshot as ki, RetainedChildRecovery as kn, ContinuationPanel as kr, workerTraceSeamKey as kt, ExecutionBindingReceipt as l, LoopSandboxPlacement as la, NativeContextContinuationInput as li, sanitizeRuntimeStreamEvent as ln, DriverFailureClass as lr, SupervisedResult as lt, ExecutorContext as m, LoopTraceEvent as ma, RetainedInteractiveAdmission as mi, WaitRejection as mn, DriverRetryPolicy as mr, UnconfirmedTeardown as mt, AgentExecutionRef as n, LoopIterationDispatchPayload as na, RuntimeRunStateError as ni, RuntimeTelemetryOptions as nn, persistHarnessTranscript as nr, SpawnRejection as nt, BudgetViolation as o, LoopPlanDescription as oa, createWorkerSlots as oi, createRuntimeStreamEventCollector as on, DriverBudgetReadout as or, SteerableRootHandle as ot, ExecutorCancellationRequest as p, LoopTraceEmitter as pa, RecoverRetainedRunResult as pi, WaitProbeRegistry as pn, DriverRepromptRefusal as pr, TreeView as pt, RootSignal as q, OpenAIChatToolChoice as qa, createBudgetPool as qi, HarnessTranscriptArtifact as qn, BackendTransportError as qr, generateSpanId as qt, AgentSpec as r, LoopIterationEndedPayload as ra, ValidationError as ri, SanitizedKnowledgeReadinessReport as rn, DriverAttemptRecord as rr, Spend as rt, DefaultVerdict as s, LoopPlanPayload as sa, NativeContextContinuationExecution as si, sanitizeAgentRuntimeEvent as sn, DriverContinuationRecord as sr, SubtreeResult as st, Agent as t, LoopEndedPayload as ta, RetainedRunDispatchBindingError as ti, RuntimeStreamEventSummary as tn, harnessTranscriptArtifact as tr, SpawnPrior as tt, Executor as u, LoopStartedPayload as ua, ReconnectRetainedRunOptions as ui, PendingWait as un, DriverLoopRecord as ur, Supervisor as ut, ExecutorNodeContext as v, OutputAdapter as va, RetainedRunAdmission as vi, pollFor as vn, UnavailablePausePolicy as vr, WidenGate as vt, FleetYield as w, RuntimeRunCompleteInput as wa, RetainedRunEnvironmentAdmission as wi, ToolStepInput as wn, CheckVerdict as wr, WorkerTraceResolver as wt, ExecutorResult as x, SelectionReceipt as xa, RetainedRunCancellation as xi, waitUntil as xn, CONTINUATION_FACTS as xr, WorkerTraceEvidence as xt, ExecutorProgressEvent as y, RunProvenance as ya, RetainedRunAdmissionHook as yi, timerAt as yn, AdmittedFinding as yr, WorkerInteractiveSession as yt, ResourceSpend as z, AgentTaskContext as za, ReservationFloor as zi, readWorkerProgress as zn, composeContinuationNote as zr, OtelExporter as zt };
-//# sourceMappingURL=types-D56jQad-.d.ts.map
+//# sourceMappingURL=types-wIVzqXdQ.d.ts.map

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tangle-network/agent-runtime",
-  "version": "0.285.0",
+  "version": "0.286.0",
   "description": "Shared task-lifecycle skeleton for agents: a recursive loop kernel for chat turns, one-shot tasks, and multi-attempt loops, with trace capture and eval-gated self-improvement. Domain behavior lives in adapters; scoring and ship-gates in @tangle-network/agent-eval.",
   "homepage": "https://github.com/tangle-network/agent-runtime#readme",
   "repository": {

package/skills/profile-authoring/SKILL.md

@@ -15,7 +15,6 @@ The audience accepts or rejects the result: a paper's authors and referees, a re
 The proxy is the check you can run during the work: an exact checker, a test suite, a rater, or a demo.
 The gap is everything the proxy accepts and the audience rejects.
 Hacks live in the gap, and a stronger model does not close it.
-Measured: where a known gap existed, 15 of 27 lanes with claims filed false "new" results; where none was known, 0 of 34 lanes did.
 
 ## 1. Name the audience and its acceptance test
 
@@ -82,19 +81,32 @@ A null that is no search, such as a spec conflict or a gap list, names each item
 A null without a certificate, or without a registered reproducer, is recorded as "stopped, unverified", never as success.
 A null inside a range the source checked or an earlier reproduced null covered adds nothing.
 Tell agents the checker may be wrong; credit a checker-defect report the same way, once someone outside the run reproduces the defect.
-Measured elsewhere: a credited exit cut one model's test cheating from 54% to 9%, while prohibitions barely moved another's.
-Measured: an in-run check that accepted a documented null made it the cheapest hack in 11 of 16 blind judgments.
-
-A Runtime manager with a check is not served `stop`.
-It ends when the check accepts its `submit_result`, when `report_blocked` shows that a granted tool failed, or at a bound.
-The bounds are the continuation `deadline`, `maxBarren` turns without progress, the budget, and cancellation.
-Runtime settles an accepted submission as delivered, and a parent may promote it; it keeps no refused one.
-So the run records its certificate on a page, such as a Knowledge page, where the outside check reads it.
-The in-run check never passes a null: it refuses a malformed certificate and holds a well-formed one.
-A held null ends the run at `maxBarren`; set it low, such as 2.
-The run settles as stopped, not delivered.
-State the continuation block in the record, and put the honest exit in its `rules` text.
-The check or a bound ends the run; an agent's claim that it is done does not.
+
+Separate execution settlement, development progress, and final acceptance.
+An unknown outside verdict does not establish absent research progress.
+Match completion instructions to the tools actually served to that node.
+Require `submit_result` only when its grant and checker exist.
+Keep unassessed results unassessed when the execution contract permits settlement without acceptance.
+
+For continued research, inspect the installed Runtime deliverable and continuation contracts before choosing bounds.
+Use the existing `checkState` interface for calibrated development measurements when applicable.
+Probe useful progress and unchanged or invalid artifacts before relying on the measurement.
+An always-false acceptance placeholder alone cannot distinguish useful work from stagnation.
+Preserve genuine no-progress limits, deadlines, budgets, and cancellation.
+Choose `maxBarren` from the task's measured feedback cadence; pending assessment does not justify a universally low value.
+
+For managers, state how they wait for descendants and collect their artifacts before returning.
+A nested manager returning can end its scope and cancel live descendants.
+Use the granted `await_event` tool while required descendant work remains live.
+A bounded event wait can return before the worker finishes; inspect its state and continue waiting when appropriate.
+A background shell timer does not establish Runtime re-entry.
+Require each descendant to retain partial work where its parent can recover it after interruption.
+
+Retain certificates and partial work where the outside assessor can read exact bytes.
+A certificate becomes accepted only through the registered outside check.
+Record the actual termination cause separately from the research outcome.
+State the continuation policy and available completion path in the immutable input.
+A source merge proves no change to an installed or running process.
 
 ## 6. Carry known traps
 
@@ -103,7 +115,6 @@ Write each trap as what the held-back acceptance refuses, never as what the prox
 "The outside check does not count a c = 0 instance" is a trap; "the checker passed c = 0 claims" is a recipe.
 Examples: excluded or trivial cases, stale or wrong-field comparators, published answers, tests versus users, a demo versus a customer, a rater versus an expert.
 Keep infrastructure incidents out of the brief; the operator owns the stack.
-Measured: a director told of an infrastructure failure spent 2 of its first 4 children probing it.
 Record each new trap where the next author reads it, in the same change as its fix.
 
 ## 7. Set budget and caps
@@ -118,7 +129,6 @@ Give each paid job its own named key with a dollar cap, and state the dollars be
 
 Before launch, register the claim, the arms, the outcome measures, the useful-effect threshold, and the result that would refute the claim.
 Run each ingredient you want to credit against a control without it: the referee, the credited exit, live notes, the wording.
-Measured elsewhere: prompt wording alone moved one model's cheating from over 85% to 1%.
 Count precision after review, not claim volume.
 Report reviewed-claim precision, canary pass rate, how many proxy passes the audience accepts, and the cost per accepted result.
 
@@ -136,15 +146,20 @@ Keep unknown cost, output and state unknown; never record them as zero or succes
 Before a new or changed profile, name the observed reason and the result that would reverse the choice.
 Before you retry a null or failed child, inspect its recorded artifacts; finished work may be usable.
 A profile with no Runtime coordination tool is a leaf; any coordination tool makes it a managed node.
-Only `agent_runtime_coordination_spawn_worker: true` grants recursion; names, metadata and role labels grant nothing.
+`agent_runtime_coordination_spawn_worker: true` grants Runtime spawning; names, metadata and role labels grant nothing.
+Prove recursive execution through actual child-to-grandchild edges.
+Record harness-native children separately from Runtime children.
+Group retries of the same assignment into one lineage when comparing teams.
 Grant `submit_result` only when Runtime gives that node an independent check; otherwise its settlement stays unassessed.
 Grant only the observation, steering and journal tools that the assignment needs.
 Set `harness`, `model.default` and `model.provider` explicitly, and validate against the current `agentProfileSchema`.
 A tool name must be one the harness publishes, and a refused name fails the same way on retry.
 `ls`, `list` and `find` are not tool names; select paths with a glob tool and search contents with a grep tool.
-An unattended run denies any permission the profile does not grant; keep file access inside the workspace or grant it.
-A stop or cancel is a request that a live acknowledger applies; plan the out-of-band path for a wedged parent.
-Compare Runtime's materialization receipt with what you authored before a large run.
+Inspect materialized permissions and served tools before relying on a restriction.
+Distinguish exact Runtime grants from harness-native defaults, which may expose additional tools.
+Keep file access within the actual permission boundary.
+Use the execution owner's supported cancellation and recovery procedures for a wedged parent.
+Compare Runtime's materialization receipt and served tool set with what you authored before a large run.
 
 Hand a file by path in `resources.files`, and compare the returned sha256 with your source:
 
@@ -172,82 +187,20 @@ Referee: family, instrument, dollars:
 Monitor: model, what it may steer:
 Null certificate: space, rows, calibration; registered reproducer, sha256, source calibration; who reruns it, budget:
 Credited exit, in the agents' words:
-Stop: check, report_blocked, deadline, maxBarren, budget:
+Served completion tools and observed materialization:
+Development progress measurement, calibration, and feedback cadence:
+Stop: installed check/continuation policy, deadline, maxBarren, budget, cancellation:
 Known traps handed to agents:
 Budget, caps, key and its dollar cap:
 Arms, control, measures, threshold, refuting result:
 Records that expose a hack:
 ```
 
-## Worked examples
-
-### Counterexample hunt, before and after
-
-Before, on a 36-lane hunt over 12 conjectures in 2026-09:
-The brief transcribed Wang and Zheng's Conjecture 13 (arXiv:2104.12942) as "c != 1".
-The checker refused only c = 1, and its admission decoys came from that transcription.
-Workers held the same checker that scored claims, and it counted any verified claim as new.
-All 3 lanes on that conjecture filed 39 "verified-new" claims, every one at c = 0.
-The paper assumes c != 0 throughout; at c = 0 the question is only whether x^d permutes the field.
-The real yield was 0.
-
-After:
-Audience: the paper's authors and its referees.
-Acceptance: Conjecture 13 quoted with the standing assumption that c is not 0 or 1, plus a literature sweep.
-Gap probe: before launch, the checker refuses c = 0 and classes a Corollary 11 family member as known.
-Agents run the development checker; the operator reruns a held-back checker and the sweep on each packet.
-A referee from another family recomputes each claim with its own code.
-A null over m <= 20 ends the run at its bound, and it counts once the outside check reruns its own reproducer on the certificate's cells.
-Arms: with and without the referee.
-
-### A null that counts
-
-Wang and Zheng checked Conjecture 13 for 2 <= m <= 10.
-Two lanes searched every (m, d, c) with c not in {0, 1} to m = 20, found no counterexample, and wrote a table.
-Nobody outside the run could check that table.
-Before the next launch, the operator registers a reference search for one m, its sha256, and its calibration: Theorem 9 gives 4 PcN pairs at m = 6 and 8 at m = 10, and a canary drops family (2) so those pairs must count as found.
-As a certificate, the null reads:
-
-```json
-{"space": {"cells": {"m": {"from": 2, "to": 20}}, "method": "enumerate", "bound": "m <= 20",
-  "within": "each unit d modulo 2^m - 1, one per Frobenius orbit, and each c outside {0, 1}"},
- "rows": [{"cell": {"m": 12}, "searched": 585442, "positives": 28, "found": 0}],
- "calibration": [{"cell": {"m": 6}, "why": "Theorem 9: x^5 is PcN for c in GF(4) outside {0, 1}"}],
- "reference": "sha256:5d198f4fc2d9..."}
-```
+## Examples
 
-The rows list every m; one is shown.
-The outside check hashes its copy, reruns the calibration and the canary with no network, then m = 11 to 20, then the rest.
-Rows typed instead of computed fail at the first cell that differs, and a script in the lane's packet is never run.
-A reference that searches nothing fails at m = 6; one whose found never fires fails the canary.
-If the budget ends after m = 18, the null is credited for 11 <= m <= 18 only.
-The in-run check holds the certificate, and the run ends at `maxBarren`; only the rerun gives credit.
-The trap for this conjecture reads: "the outside check does not count a c = 0 instance; the source sets c = 0 aside".
-
-### Coding task scored by tests
-
-Audience: the task's author, and a maintainer who decides whether to merge.
-Proxy: the visible tests.
-Gap: a change that passes the tests and that a maintainer rejects, such as a special-cased input, an edited test or a skipped check.
-Measured: in an 8-lane Terminal-Bench batch, Runtime marked 6 lanes as winners and the task's verifier passed 1 of them.
-Profile: agents run the visible tests.
-The held-back verifier and hidden tests run on the patch in a fresh container.
-Hash the test files; a changed test fails acceptance.
-A referee from another family reads the diff for special cases.
-A report that the task cannot pass without breaking its specification ends the run; it counts once the held-back verifier reproduces the conflict.
-
-### Product for a customer
-
-Audience: the customer who pays, and the decision the product supports.
-Acceptance: the customer's own evaluation criteria, quoted, and a walkthrough of the running product doing the customer's task.
-Proxy: an LLM fulfilment rater and a compliance checklist.
-Gap, measured across product lanes: the rater accepted 19 of 31 failed lanes and scored 0.45 where always-reject scored 0.73.
-In one wave, 19 of 44 reviewers never received the product.
-Profile: reviewers mount the product by path and must run it.
-A referee from another family walks the customer's task on the running product.
-The rater stays a ruler, not a gate, until it beats always-reject.
-Acceptance is the owner's or the customer's walkthrough.
-"Not ready, and this is what is missing" ends the run; it counts once the owner's walkthrough reproduces each gap.
+For acceptance gaps and null certificates, read [acceptance examples](references/acceptance-examples.md).
+The core instructions above are complete when only this file is mounted.
+Mount the reference as well when the assignment needs its worked examples.
 
 ## Then consider
 

package/skills/profile-authoring/references/acceptance-examples.md

@@ -0,0 +1,93 @@
+# Acceptance examples
+
+These examples preserve historical observations from the previous skill.
+Their counts are not current fleet measurements or default settings.
+
+## Contents
+
+- [Counterexample hunt](#counterexample-hunt-before-and-after)
+- [A null that counts](#a-null-that-counts)
+- [Coding task](#coding-task-scored-by-tests)
+- [Customer product](#product-for-a-customer)
+- [Other historical observations](#other-historical-observations)
+
+## Counterexample hunt, before and after
+
+Before, on a 36-lane hunt over 12 conjectures in 2026-09:
+The brief transcribed Wang and Zheng's Conjecture 13 (arXiv:2104.12942) as "c != 1".
+The checker refused only c = 1, and its admission decoys came from that transcription.
+Workers held the same checker that scored claims, and it counted any verified claim as new.
+All 3 lanes on that conjecture filed 39 "verified-new" claims, every one at c = 0.
+The paper assumes c != 0 throughout; at c = 0 the question is only whether x^d permutes the field.
+The real yield was 0.
+
+After:
+Audience: the paper's authors and its referees.
+Acceptance: Conjecture 13 quoted with the standing assumption that c is not 0 or 1, plus a literature sweep.
+Gap probe: before launch, the checker refuses c = 0 and classes a Corollary 11 family member as known.
+Agents run the development checker; the operator reruns a held-back checker and the sweep on each packet.
+A referee from another family recomputes each claim with its own code.
+A null over m <= 20 ends the run at its bound, and it counts once the outside check reruns its own reproducer on the certificate's cells.
+Arms: with and without the referee.
+
+## A null that counts
+
+Wang and Zheng checked Conjecture 13 for 2 <= m <= 10.
+Two lanes searched every (m, d, c) with c not in {0, 1} to m = 20, found no counterexample, and wrote a table.
+Nobody outside the run could check that table.
+Before the next launch, the operator registers a reference search for one m, its sha256, and its calibration: Theorem 9 gives 4 PcN pairs at m = 6 and 8 at m = 10, and a canary drops family (2) so those pairs must count as found.
+As a certificate, the null reads:
+
+```json
+{"space": {"cells": {"m": {"from": 2, "to": 20}}, "method": "enumerate", "bound": "m <= 20",
+  "within": "each unit d modulo 2^m - 1, one per Frobenius orbit, and each c outside {0, 1}"},
+ "rows": [{"cell": {"m": 12}, "searched": 585442, "positives": 28, "found": 0}],
+ "calibration": [{"cell": {"m": 6}, "why": "Theorem 9: x^5 is PcN for c in GF(4) outside {0, 1}"}],
+ "reference": "sha256:5d198f4fc2d9..."}
+```
+
+The rows list every m; one is shown.
+The outside check hashes its copy, reruns the calibration and the canary with no network, then m = 11 to 20, then the rest.
+Rows typed instead of computed fail at the first cell that differs, and a script in the lane's packet is never run.
+A reference that searches nothing fails at m = 6; one whose found never fires fails the canary.
+If the budget ends after m = 18, the null is credited for 11 <= m <= 18 only.
+The run retains the certificate and ends under its registered stop policy; only the outside rerun gives credit.
+The trap for this conjecture reads: "the outside check does not count a c = 0 instance; the source sets c = 0 aside".
+
+## Coding task scored by tests
+
+Audience: the task's author, and a maintainer who decides whether to merge.
+Proxy: the visible tests.
+Gap: a change that passes the tests and that a maintainer rejects, such as a special-cased input, an edited test or a skipped check.
+Measured: in an 8-lane Terminal-Bench batch, Runtime marked 6 lanes as winners and the task's verifier passed 1 of them.
+Profile: agents run the visible tests.
+The held-back verifier and hidden tests run on the patch in a fresh container.
+Hash the test files; a changed test fails acceptance.
+A referee from another family reads the diff for special cases.
+A conflict report follows the registered completion and stop policy.
+It counts only after the held-back verifier reproduces the conflict.
+
+## Product for a customer
+
+Audience: the customer who pays, and the decision the product supports.
+Acceptance: the customer's own evaluation criteria, quoted, and a walkthrough of the running product doing the customer's task.
+Proxy: an LLM fulfilment rater and a compliance checklist.
+Gap, measured across product lanes: the rater accepted 19 of 31 failed lanes and scored 0.45 where always-reject scored 0.73.
+In one wave, 19 of 44 reviewers never received the product.
+Profile: reviewers mount the product by path and must run it.
+A referee from another family walks the customer's task on the running product.
+The rater stays a ruler, not a gate, until it beats always-reject.
+Acceptance is the owner's or the customer's walkthrough.
+A gap report follows the registered completion and stop policy.
+It counts only after the owner's walkthrough reproduces each gap.
+
+## Other historical observations
+
+These observations motivated the original guidance.
+Consult their underlying records before reusing their numerical claims.
+
+- Measured: where a known gap existed, 15 of 27 lanes with claims filed false "new" results; where none was known, 0 of 34 lanes did.
+- Measured elsewhere: a credited exit cut one model's test cheating from 54% to 9%, while prohibitions barely moved another's.
+- Measured: an in-run check that accepted a documented null made it the cheapest hack in 11 of 16 blind judgments.
+- Measured: a director told of an infrastructure failure spent 2 of its first 4 children probing it.
+- Measured elsewhere: prompt wording alone moved one model's cheating from over 85% to 1%.