Public evidence summary
Evidence-backed claims and their limits
- Record date
- 13 October 2026 register update; individual historical test dates are shown.
- Scope
- Published register or policy snapshot; per-item scope and dates apply.
- Current deployment
- Not independently verified by this historical evidence. Current availability
Recorded result
The claim register updated on 13 October 2026 contains 34 bounded claims covering memory continuity, checkpoint restoration, replay billing, verified work and a production agent-session canary. Each claim links to its own dated evidence and limits. The replay-billing claims hold again: a call retried after a resume could be billed twice from 20 September to 6 October 2026, and the paid gateway test re-ran green on 13 October 2026. The register does not establish universal compatibility, guaranteed performance, an effective spending ceiling for unpriced usage or a savings rate. Current availability is stated separately on Proof.
Recorded claims and their boundaries
Memory continuity
Historical recorded result
A microVM retained its in-memory state after the tested abrupt process loss.
Date: 28 July 2026
Scope: Historical record; any production observation is limited to the dated test described below.
Limit: Counter guest, not AI agent Read-only filesystem test Laboratory, not production
Read the dated evidence summary →Session lifecycle
Historical recorded result
Session control retained continuity through the tested kill-and-resume lifecycle.
Date: 28 July 2026
Scope: Laboratory or bench evidence; production behavior is not inferred from it.
Limit: Clock resynchronization untested General-image smoke result is data only Laboratory scope
Read the dated evidence summary →Portable checkpoint objects
Historical recorded result
A second service instance on the same laboratory host restored a checkpoint from its durable objects. This test alone is not cross-host recovery.
Date: 28 July 2026
Scope: Laboratory or bench evidence; production behavior is not inferred from it.
Limit: Second instance on same host Compatible software/local storage Not cross-host recovery Branch divergence not steering
Read the dated evidence summary →Crash-consistent publication
Historical recorded result
The tested interrupted checkpoint publication preserved the previous usable checkpoint.
Date: 28 July 2026
Scope: Laboratory or bench evidence; production behavior is not inferred from it.
Limit: Second instance on same host Compatible software/local storage Not cross-host recovery Branch divergence not steering
Read the dated evidence summary →Corruption is refused
Historical recorded result
A deliberately corrupted checkpoint was refused rather than restored.
Date: 28 July 2026
Scope: Laboratory or bench evidence; production behavior is not inferred from it.
Limit: Second instance on same host Compatible software/local storage Not cross-host recovery Branch divergence not steering
Read the dated evidence summary →Billing reconciliation across resume
Historical recorded result
The tested interrupted session reconciled its recorded usage on resume.
Date: 28 July 2026
Scope: Laboratory or bench evidence; production behavior is not inferred from it.
Limit: Claude Code client scope Client retry behavior separate Task outcome not asserted Some billing checks alone remain true in degraded system
Read the dated evidence summary →Two live provider families
Historical recorded result
A bounded fixture completed through each of two live provider families. Arbitrary-provider support is not claimed; translation has a separate scope.
Date: 6 September 2026
Scope: Historical evidence; the scope and environment are described in the recorded result.
Limit: One bounded fixture in two provider families Initial attempts were deliberately resource-constrained No external publication or general success rate
Read the dated evidence summary →At-most-once billing
Historical recorded result; re-run green 13 October 2026
The recorded correction rejected duplicate billing in the tested cases. It does not establish all future provider or client behavior. From 20 September to 6 October 2026 a call retried after a resume could be billed twice. The fix is in the production gateway since 6 October 2026, and the paid gateway test re-ran green on 13 October 2026.
Date: 25 August 2026
Scope: Historical record; any production observation is limited to the dated test described below.
Limit: Laboratory regression scope Missing measurements are unavailable, not zero. No current production proof. From 20 September to 6 October 2026 a call retried after a resume could be billed twice. The fix is in the production gateway since 6 October 2026, and the paid gateway test re-ran green on 13 October 2026.
Read the dated evidence summary →Session-state persistence
Historical recorded result
A storage-replacement test retained session metadata after correction; unavailable persistence refused transitions.
Date: 25 August 2026
Scope: Historical record; any production observation is limited to the dated test described below.
Limit: Seeded records, no VM or model calls Laboratory scope No production recovery claim
Read the dated evidence summary →Claude Code interrupted-call replay
Historical recorded result; re-run green 13 October 2026
The recorded Claude Code retry received its stored response without additional billed usage. Client retry behavior remains agent-specific. From 20 September to 6 October 2026 a call retried after a resume could be billed twice. The fix is in the production gateway since 6 October 2026, and the paid gateway test re-ran green on 13 October 2026.
Date: 28 July 2026
Scope: Laboratory or bench evidence; production behavior is not inferred from it.
Limit: Claude Code client scope Client retry behavior separate Task outcome not asserted Some billing checks alone remain true in degraded system. From 20 September to 6 October 2026 a call retried after a resume could be billed twice. The fix is in the production gateway since 6 October 2026, and the paid gateway test re-ran green on 13 October 2026.
Read the dated evidence summary →Unattended interruption demo
Historical recorded result
The historical demonstration completed without manual intervention and measured 25.2 seconds of downtime.
Date: 29 July 2026
Scope: Laboratory or bench evidence; production behavior is not inferred from it.
Limit: Alive after resume, not task completion Component timings reported, not all asserted Dollar figures estimated Single laboratory demonstration
Read the dated evidence summary →Accepted work with bounded authorization
Historical recorded result
Two live task arms produced accepted artifacts and reconciled accounting. Publication stopped at authorization; no external pull request was created.
Date: 6 September 2026
Scope: Historical evidence; the scope and environment are described in the recorded result.
Limit: One bounded fixture in two provider families Initial attempts were deliberately resource-constrained No external publication or general success rate
Read the dated evidence summary →Memory and files after resume
Historical recorded result
The named laboratory tests retained memory and checkpointed files through interruption. Task completion is still reported per run.
Date: 29 July 2026
Scope: Laboratory or bench evidence; production behavior is not inferred from it.
Limit: Alive after resume, not task completion Component timings reported, not all asserted Dollar figures estimated Single laboratory demonstration
Read the dated evidence summary →Fork creation and isolation
Historical recorded result
Three branches had separate changed files; hosted fork creation produced owned children. Creation and isolation alone do not prove useful branch work or successful later resume.
Date: 29 July 2026
Scope: Laboratory or bench evidence; production behavior is not inferred from it.
Limit: Marker written by test harness Isolation is not agent obedience Laboratory scope
Read the dated evidence summary →Branch steering
Historical recorded result
The tested branches received distinct steering while their pre-steering input stayed equivalent.
Date: 29 July 2026
Scope: Laboratory or bench evidence; production behavior is not inferred from it.
Limit: Marker written by test harness Isolation is not agent obedience Laboratory scope
Read the dated evidence summary →Durability sample
Historical recorded result
25/25 fresh cycles passed, with a 95% Wilson interval of 86.7%–100%. Later soak configurations are separate results, including a failed run.
Date: 31 July 2026
Scope: Laboratory or bench evidence; production behavior is not inferred from it.
Limit: One soak/one workload Original preserved artifact destroyed; original restoration not performed Replacement restoration is distinct 25-cycle confidence limits remain visible
Read the dated evidence summary →Aged checkpoint
Historical recorded result
A preserved checkpoint aged 15.2 hours resumed in 14.3 seconds in the recorded test. The original artifact was later destroyed, limiting direct re-verification.
Date: 31 July 2026
Scope: Laboratory or bench evidence; production behavior is not inferred from it.
Limit: One soak/one workload Original preserved artifact destroyed; original restoration not performed Replacement restoration is distinct 25-cycle confidence limits remain visible
Read the dated evidence summary →Storage reduction in the measured workload
Historical recorded result
An 82-checkpoint run used 4,233 MiB rather than a naive 80,952 MiB. This sample does not promise a storage ratio for other workloads.
Date: 31 July 2026
Scope: Laboratory or bench evidence; production behavior is not inferred from it.
Limit: One soak/one workload Original preserved artifact destroyed; original restoration not performed Replacement restoration is distinct 25-cycle confidence limits remain visible
Read the dated evidence summary →Measured memory use
Historical recorded result
The historical seven-hour soak reported no peak resident-memory growth for its two measured services. It is not a general no-leak proof.
Date: 31 July 2026
Scope: Laboratory or bench evidence; production behavior is not inferred from it.
Limit: One soak/one workload Original preserved artifact destroyed; original restoration not performed Replacement restoration is distinct 25-cycle confidence limits remain visible
Read the dated evidence summary →Host-loss recovery and billing continuity
Historical recorded result
A same-region host-loss test resumed checkpointed sessions and reconciled carried spending on a replacement host. Continuous failover, cross-region recovery and a recovery-time objective remain unclaimed.
Date: 6 September 2026
Scope: Historical record; any production observation is limited to the dated test described below.
Limit: 47 seconds is one run’s deallocation-to-reconciliation measurement. Shared storage and separately tested guest recovery are dependencies. Settlement remains blocked until reconciliation completes. Production enablement is a separate 7 September addendum.
Read the dated evidence summary →Approved acceptance checks
Historical recorded result
A laboratory test refused an unapproved or changed task contract and changed tests before executing work.
Date: 16 September 2026
Scope: Historical record; any production observation is limited to the dated test described below.
Limit: Only pytest and Go test discovery covered. Laboratory client and recording service; no live execution proved.
Read the dated evidence summary →Bounded retries
Historical recorded result
The fixture retried from the same saved state, refused unsupported model choices and stopped at its spending limit or final attempt.
Date: 16 September 2026
Scope: Historical record; any production observation is limited to the dated test described below.
Limit: Laboratory fixture only No production spending or savings guarantee
Read the dated evidence summary →Ambiguous attempts stay visible
Historical recorded result
The fixture kept ambiguous work visible and prevented acceptance while accounting reconciled all attempts.
Date: 16 September 2026
Scope: Laboratory or bench evidence; production behavior is not inferred from it.
Limit: Fixture backend and modeled timeout. No live-provider reconciliation-latency claim.
Read the dated evidence summary →Job state reflects recorded evidence
Historical recorded result
The tested API and interface rejected inconsistent state, restricted concurrency and displayed absent spending as unavailable.
Date: 16 September 2026
Scope: Historical record; any production observation is limited to the dated test described below.
Limit: Laboratory services and test identity provider. Customer credit changes require authorized support. Historical tier constraint requires current product confirmation.
Read the dated evidence summary →Offline help and verified downloads
Historical recorded result
The laboratory client made no request for help and rejected a tampered artifact before writing it. Workspaces over the tested upload limit were refused.
Date: 16 September 2026
Scope: Laboratory or bench evidence; production behavior is not inferred from it.
Limit: 50 MiB threshold and exclusions belong to this tested release. Recording service, not full production integration.
Read the dated evidence summary →Review, resume and cancel
Historical recorded result
A fixture that exhausted its attempts entered review, could resume under a newly approved attempt sequence, and released the saved parent when canceled.
Date: 16 September 2026
Scope: Historical record; any production observation is limited to the dated test described below.
Limit: Fixture backend. Resume only from review. Notification event is not proof of message delivery. Since 12 October 2026 the service sends the review email; this fixture test has not re-run against the deployed service.
Read the dated evidence summary →Fresh workspace per attempt
Historical recorded result
The hosted test started each attempt from the same saved workspace and kept the parent parked. Outcome and duration remain per-run data.
Date: 17 September 2026
Scope: Historical evidence; the scope and environment are described in the recorded result.
Limit: No inference call or task acceptance: test accounts had no provider key. Successful-attempt parking was conditional and not demonstrated. One job; checkpoint pauses are host-specific.
Read the dated evidence summary →Independent host-side checks
Historical recorded result
The hosted test stopped and extracted the candidate before independent verification against approved tests; it rejected a forged pass log.
Date: 16 September 2026
Scope: Historical evidence; the scope and environment are described in the recorded result.
Limit: One hosted attempt The job ended in review, not acceptance Current deployment independently unverified
Read the dated evidence summary →Customer-path hosted job
Historical recorded result
A dated production canary reached a terminal state, verified the accepted download and denied a bystander account. This is one canary, not a current deployment guarantee.
Date: 17 September 2026
Scope: Historical record; any production observation is limited to the dated test described below.
Limit: Monetary spend unavailable; usage does not establish money spent. Limited job/account scenario; acceptance is not a performance promise. Artifact verification applies only to accepted runs.
Read the dated evidence summary →Customer credit authorization in test
Historical recorded result
Tests refused inappropriate credit changes by a customer while preserving verified payments and authorized grants.
Date: 17 September 2026
Scope: Historical record; any production observation is limited to the dated test described below.
Limit: Disposable unfunded tenant; verification granted nothing Point-in-time ledger review only Broader rerun exception Historical, not current exploitability proof
Read the dated evidence summary →Production credit-authorization canary
Historical recorded result
A dated production test refused inappropriate customer credit changes and recorded no corresponding ledger movement.
Date: 17 September 2026
Scope: Historical record; any production observation is limited to the dated test described below.
Limit: Disposable unfunded tenant; verification granted nothing Point-in-time ledger review only Broader rerun exception Historical, not current exploitability proof
Read the dated evidence summary →Usage settlement in the laboratory
Historical recorded result
A fixture consumed credit once per finalized usage amount, charged only new usage, parked at exhaustion and resumed after a top-up.
Date: 17 September 2026
Scope: Historical record; any production observation is limited to the dated test described below.
Limit: Production running-session parking remained untested Original monitoring count corrected on 18 September Laboratory parking used a simulated service
Read the dated evidence summary →Production usage settlement
Historical recorded result
A historical production correction settled outstanding usage once and recorded the funds restriction. No affected session was running, so production parking at exhaustion remained untested.
Date: 17 September 2026
Scope: Historical record; any production observation is limited to the dated test described below.
Limit: Production running-session parking remained untested Original monitoring count corrected on 18 September Laboratory parking used a simulated service
Read the dated evidence summary →Dated production agent-session canary
Historical recorded result
A single production account with its own key completed a supervised agent instruction with Claude Code 2.1.251; the no-key control was refused, usage was metered and deletion stopped the tested machine.
Result: Eight normal-run checks and a separate missing-key negative check passed.
Date: 30 September 2026; final normal run at 09:16:29 UTC
Scope: One team-operated production account using its own provider key and Claude Code 2.1.251.
Limit: Dollar cost unavailable; checkpoint erasure, general task quality and every registry-listed workflow were not tested.
Read the dated evidence summary →Deliberately not claimed
Guaranteed task completion
Restoring memory and files does not guarantee that the resumed agent finishes its task.
Arbitrary providers
Support must be established separately per provider and translation path.
Production readiness or scale
Limited recovery and soak tests do not establish production readiness, continuous failover, or scale.
Zero downtime
Measured resume takes seconds; the scoped timings remain data, not a zero-downtime promise.
Guaranteed aider retry
The recorded client returned for its stored response in three of four runs.
Universal agents
Each harness and behavior level needs its own evidence.
A general resume success percentage
The small recorded sample supports its confidence interval, not a universal percentage.
The public summary as JSON. These are the same reviewed fields shown in Read.
{
"schemaVersion": 1,
"publicSummaryVersion": "2026-09-30",
"classification": "PUBLIC",
"title": "Evidence-backed claims and their limits",
"summary": "The claim register updated on 13 October 2026 contains 34 bounded claims covering memory continuity, checkpoint restoration, replay billing, verified work and a production agent-session canary. Each claim links to its own dated evidence and limits. The replay-billing claims hold again: a call retried after a resume could be billed twice from 20 September to 6 October 2026, and the paid gateway test re-ran green on 13 October 2026. The register does not establish universal compatibility, guaranteed performance, an effective spending ceiling for unpriced usage or a savings rate. Current availability is stated separately on Proof.",
"date": "13 October 2026 register update; individual historical test dates are shown.",
"scope": "Published register or policy snapshot; per-item scope and dates apply.",
"result": "The claim register updated on 13 October 2026 contains 34 bounded claims covering memory continuity, checkpoint restoration, replay billing, verified work and a production agent-session canary. Each claim links to its own dated evidence and limits. The replay-billing claims hold again: a call retried after a resume could be billed twice from 20 September to 6 October 2026, and the paid gateway test re-ran green on 13 October 2026. The register does not establish universal compatibility, guaranteed performance, an effective spending ceiling for unpriced usage or a savings rate. Current availability is stated separately on Proof.",
"sections": [
{
"title": "Recorded claims and their boundaries",
"items": [
{
"id": "claim-01",
"title": "Memory continuity",
"text": "A microVM retained its in-memory state after the tested abrupt process loss.",
"status": "Historical recorded result",
"date": "28 July 2026",
"scope": "Historical record; any production observation is limited to the dated test described below.",
"result": "A microVM retained its in-memory state after the tested abrupt process loss.",
"limitation": "Counter guest, not AI agent Read-only filesystem test Laboratory, not production",
"href": "/evidence/phase-1"
},
{
"id": "claim-02",
"title": "Session lifecycle",
"text": "Session control retained continuity through the tested kill-and-resume lifecycle.",
"status": "Historical recorded result",
"date": "28 July 2026",
"scope": "Laboratory or bench evidence; production behavior is not inferred from it.",
"result": "Session control retained continuity through the tested kill-and-resume lifecycle.",
"limitation": "Clock resynchronization untested General-image smoke result is data only Laboratory scope",
"href": "/evidence/phase-2a"
},
{
"id": "claim-03",
"title": "Portable checkpoint objects",
"text": "A second service instance on the same laboratory host restored a checkpoint from its durable objects. This test alone is not cross-host recovery.",
"status": "Historical recorded result",
"date": "28 July 2026",
"scope": "Laboratory or bench evidence; production behavior is not inferred from it.",
"result": "A second service instance on the same laboratory host restored a checkpoint from its durable objects. This test alone is not cross-host recovery.",
"limitation": "Second instance on same host Compatible software/local storage Not cross-host recovery Branch divergence not steering",
"href": "/evidence/phase-2b"
},
{
"id": "claim-04",
"title": "Crash-consistent publication",
"text": "The tested interrupted checkpoint publication preserved the previous usable checkpoint.",
"status": "Historical recorded result",
"date": "28 July 2026",
"scope": "Laboratory or bench evidence; production behavior is not inferred from it.",
"result": "The tested interrupted checkpoint publication preserved the previous usable checkpoint.",
"limitation": "Second instance on same host Compatible software/local storage Not cross-host recovery Branch divergence not steering",
"href": "/evidence/phase-2b"
},
{
"id": "claim-05",
"title": "Corruption is refused",
"text": "A deliberately corrupted checkpoint was refused rather than restored.",
"status": "Historical recorded result",
"date": "28 July 2026",
"scope": "Laboratory or bench evidence; production behavior is not inferred from it.",
"result": "A deliberately corrupted checkpoint was refused rather than restored.",
"limitation": "Second instance on same host Compatible software/local storage Not cross-host recovery Branch divergence not steering",
"href": "/evidence/phase-2b"
},
{
"id": "claim-06",
"title": "Billing reconciliation across resume",
"text": "The tested interrupted session reconciled its recorded usage on resume.",
"status": "Historical recorded result",
"date": "28 July 2026",
"scope": "Laboratory or bench evidence; production behavior is not inferred from it.",
"result": "The tested interrupted session reconciled its recorded usage on resume.",
"limitation": "Claude Code client scope Client retry behavior separate Task outcome not asserted Some billing checks alone remain true in degraded system",
"href": "/evidence/phase-3"
},
{
"id": "claim-07",
"title": "Two live provider families",
"text": "A bounded fixture completed through each of two live provider families. Arbitrary-provider support is not claimed; translation has a separate scope.",
"status": "Historical recorded result",
"date": "6 September 2026",
"scope": "Historical evidence; the scope and environment are described in the recorded result.",
"result": "A bounded fixture completed through each of two live provider families. Arbitrary-provider support is not claimed; translation has a separate scope.",
"limitation": "One bounded fixture in two provider families Initial attempts were deliberately resource-constrained No external publication or general success rate",
"href": "/evidence/phase-l-1"
},
{
"id": "claim-08",
"title": "At-most-once billing",
"text": "The recorded correction rejected duplicate billing in the tested cases. It does not establish all future provider or client behavior. From 20 September to 6 October 2026 a call retried after a resume could be billed twice. The fix is in the production gateway since 6 October 2026, and the paid gateway test re-ran green on 13 October 2026.",
"status": "Historical recorded result; re-run green 13 October 2026",
"date": "25 August 2026",
"scope": "Historical record; any production observation is limited to the dated test described below.",
"result": "The recorded correction rejected duplicate billing in the tested cases. It does not establish all future provider or client behavior. From 20 September to 6 October 2026 a call retried after a resume could be billed twice. The fix is in the production gateway since 6 October 2026, and the paid gateway test re-ran green on 13 October 2026.",
"limitation": "Laboratory regression scope Missing measurements are unavailable, not zero. No current production proof. From 20 September to 6 October 2026 a call retried after a resume could be billed twice. The fix is in the production gateway since 6 October 2026, and the paid gateway test re-ran green on 13 October 2026.",
"href": "/evidence/patch-v0.1.2"
},
{
"id": "claim-09",
"title": "Session-state persistence",
"text": "A storage-replacement test retained session metadata after correction; unavailable persistence refused transitions.",
"status": "Historical recorded result",
"date": "25 August 2026",
"scope": "Historical record; any production observation is limited to the dated test described below.",
"result": "A storage-replacement test retained session metadata after correction; unavailable persistence refused transitions.",
"limitation": "Seeded records, no VM or model calls Laboratory scope No production recovery claim",
"href": "/evidence/patch-v0.1.3"
},
{
"id": "claim-10",
"title": "Claude Code interrupted-call replay",
"text": "The recorded Claude Code retry received its stored response without additional billed usage. Client retry behavior remains agent-specific. From 20 September to 6 October 2026 a call retried after a resume could be billed twice. The fix is in the production gateway since 6 October 2026, and the paid gateway test re-ran green on 13 October 2026.",
"status": "Historical recorded result; re-run green 13 October 2026",
"date": "28 July 2026",
"scope": "Laboratory or bench evidence; production behavior is not inferred from it.",
"result": "The recorded Claude Code retry received its stored response without additional billed usage. Client retry behavior remains agent-specific. From 20 September to 6 October 2026 a call retried after a resume could be billed twice. The fix is in the production gateway since 6 October 2026, and the paid gateway test re-ran green on 13 October 2026.",
"limitation": "Claude Code client scope Client retry behavior separate Task outcome not asserted Some billing checks alone remain true in degraded system. From 20 September to 6 October 2026 a call retried after a resume could be billed twice. The fix is in the production gateway since 6 October 2026, and the paid gateway test re-ran green on 13 October 2026.",
"href": "/evidence/phase-3"
},
{
"id": "claim-11",
"title": "Unattended interruption demo",
"text": "The historical demonstration completed without manual intervention and measured 25.2 seconds of downtime.",
"status": "Historical recorded result",
"date": "29 July 2026",
"scope": "Laboratory or bench evidence; production behavior is not inferred from it.",
"result": "The historical demonstration completed without manual intervention and measured 25.2 seconds of downtime.",
"limitation": "Alive after resume, not task completion Component timings reported, not all asserted Dollar figures estimated Single laboratory demonstration",
"href": "/evidence/phase-4"
},
{
"id": "claim-12",
"title": "Accepted work with bounded authorization",
"text": "Two live task arms produced accepted artifacts and reconciled accounting. Publication stopped at authorization; no external pull request was created.",
"status": "Historical recorded result",
"date": "6 September 2026",
"scope": "Historical evidence; the scope and environment are described in the recorded result.",
"result": "Two live task arms produced accepted artifacts and reconciled accounting. Publication stopped at authorization; no external pull request was created.",
"limitation": "One bounded fixture in two provider families Initial attempts were deliberately resource-constrained No external publication or general success rate",
"href": "/evidence/phase-l-1"
},
{
"id": "claim-13",
"title": "Memory and files after resume",
"text": "The named laboratory tests retained memory and checkpointed files through interruption. Task completion is still reported per run.",
"status": "Historical recorded result",
"date": "29 July 2026",
"scope": "Laboratory or bench evidence; production behavior is not inferred from it.",
"result": "The named laboratory tests retained memory and checkpointed files through interruption. Task completion is still reported per run.",
"limitation": "Alive after resume, not task completion Component timings reported, not all asserted Dollar figures estimated Single laboratory demonstration",
"href": "/evidence/phase-4"
},
{
"id": "claim-14",
"title": "Fork creation and isolation",
"text": "Three branches had separate changed files; hosted fork creation produced owned children. Creation and isolation alone do not prove useful branch work or successful later resume.",
"status": "Historical recorded result",
"date": "29 July 2026",
"scope": "Laboratory or bench evidence; production behavior is not inferred from it.",
"result": "Three branches had separate changed files; hosted fork creation produced owned children. Creation and isolation alone do not prove useful branch work or successful later resume.",
"limitation": "Marker written by test harness Isolation is not agent obedience Laboratory scope",
"href": "/evidence/phase-5"
},
{
"id": "claim-15",
"title": "Branch steering",
"text": "The tested branches received distinct steering while their pre-steering input stayed equivalent.",
"status": "Historical recorded result",
"date": "29 July 2026",
"scope": "Laboratory or bench evidence; production behavior is not inferred from it.",
"result": "The tested branches received distinct steering while their pre-steering input stayed equivalent.",
"limitation": "Marker written by test harness Isolation is not agent obedience Laboratory scope",
"href": "/evidence/phase-5"
},
{
"id": "claim-16",
"title": "Durability sample",
"text": "25/25 fresh cycles passed, with a 95% Wilson interval of 86.7%–100%. Later soak configurations are separate results, including a failed run.",
"status": "Historical recorded result",
"date": "31 July 2026",
"scope": "Laboratory or bench evidence; production behavior is not inferred from it.",
"result": "25/25 fresh cycles passed, with a 95% Wilson interval of 86.7%–100%. Later soak configurations are separate results, including a failed run.",
"limitation": "One soak/one workload Original preserved artifact destroyed; original restoration not performed Replacement restoration is distinct 25-cycle confidence limits remain visible",
"href": "/evidence/phase-6"
},
{
"id": "claim-17",
"title": "Aged checkpoint",
"text": "A preserved checkpoint aged 15.2 hours resumed in 14.3 seconds in the recorded test. The original artifact was later destroyed, limiting direct re-verification.",
"status": "Historical recorded result",
"date": "31 July 2026",
"scope": "Laboratory or bench evidence; production behavior is not inferred from it.",
"result": "A preserved checkpoint aged 15.2 hours resumed in 14.3 seconds in the recorded test. The original artifact was later destroyed, limiting direct re-verification.",
"limitation": "One soak/one workload Original preserved artifact destroyed; original restoration not performed Replacement restoration is distinct 25-cycle confidence limits remain visible",
"href": "/evidence/phase-6"
},
{
"id": "claim-18",
"title": "Storage reduction in the measured workload",
"text": "An 82-checkpoint run used 4,233 MiB rather than a naive 80,952 MiB. This sample does not promise a storage ratio for other workloads.",
"status": "Historical recorded result",
"date": "31 July 2026",
"scope": "Laboratory or bench evidence; production behavior is not inferred from it.",
"result": "An 82-checkpoint run used 4,233 MiB rather than a naive 80,952 MiB. This sample does not promise a storage ratio for other workloads.",
"limitation": "One soak/one workload Original preserved artifact destroyed; original restoration not performed Replacement restoration is distinct 25-cycle confidence limits remain visible",
"href": "/evidence/phase-6"
},
{
"id": "claim-19",
"title": "Measured memory use",
"text": "The historical seven-hour soak reported no peak resident-memory growth for its two measured services. It is not a general no-leak proof.",
"status": "Historical recorded result",
"date": "31 July 2026",
"scope": "Laboratory or bench evidence; production behavior is not inferred from it.",
"result": "The historical seven-hour soak reported no peak resident-memory growth for its two measured services. It is not a general no-leak proof.",
"limitation": "One soak/one workload Original preserved artifact destroyed; original restoration not performed Replacement restoration is distinct 25-cycle confidence limits remain visible",
"href": "/evidence/phase-6"
},
{
"id": "claim-20",
"title": "Host-loss recovery and billing continuity",
"text": "A same-region host-loss test resumed checkpointed sessions and reconciled carried spending on a replacement host. Continuous failover, cross-region recovery and a recovery-time objective remain unclaimed.",
"status": "Historical recorded result",
"date": "6 September 2026",
"scope": "Historical record; any production observation is limited to the dated test described below.",
"result": "A same-region host-loss test resumed checkpointed sessions and reconciled carried spending on a replacement host. Continuous failover, cross-region recovery and a recovery-time objective remain unclaimed.",
"limitation": "47 seconds is one run’s deallocation-to-reconciliation measurement. Shared storage and separately tested guest recovery are dependencies. Settlement remains blocked until reconciliation completes. Production enablement is a separate 7 September addendum.",
"href": "/evidence/phase-fo-1b"
},
{
"id": "claim-21",
"title": "Approved acceptance checks",
"text": "A laboratory test refused an unapproved or changed task contract and changed tests before executing work.",
"status": "Historical recorded result",
"date": "16 September 2026",
"scope": "Historical record; any production observation is limited to the dated test described below.",
"result": "A laboratory test refused an unapproved or changed task contract and changed tests before executing work.",
"limitation": "Only pytest and Go test discovery covered. Laboratory client and recording service; no live execution proved.",
"href": "/evidence/phase-cr-1"
},
{
"id": "claim-22",
"title": "Bounded retries",
"text": "The fixture retried from the same saved state, refused unsupported model choices and stopped at its spending limit or final attempt.",
"status": "Historical recorded result",
"date": "16 September 2026",
"scope": "Historical record; any production observation is limited to the dated test described below.",
"result": "The fixture retried from the same saved state, refused unsupported model choices and stopped at its spending limit or final attempt.",
"limitation": "Laboratory fixture only No production spending or savings guarantee",
"href": "/evidence/phase-cr-2"
},
{
"id": "claim-23",
"title": "Ambiguous attempts stay visible",
"text": "The fixture kept ambiguous work visible and prevented acceptance while accounting reconciled all attempts.",
"status": "Historical recorded result",
"date": "16 September 2026",
"scope": "Laboratory or bench evidence; production behavior is not inferred from it.",
"result": "The fixture kept ambiguous work visible and prevented acceptance while accounting reconciled all attempts.",
"limitation": "Fixture backend and modeled timeout. No live-provider reconciliation-latency claim.",
"href": "/evidence/phase-cr-5"
},
{
"id": "claim-24",
"title": "Job state reflects recorded evidence",
"text": "The tested API and interface rejected inconsistent state, restricted concurrency and displayed absent spending as unavailable.",
"status": "Historical recorded result",
"date": "16 September 2026",
"scope": "Historical record; any production observation is limited to the dated test described below.",
"result": "The tested API and interface rejected inconsistent state, restricted concurrency and displayed absent spending as unavailable.",
"limitation": "Laboratory services and test identity provider. Customer credit changes require authorized support. Historical tier constraint requires current product confirmation.",
"href": "/evidence/phase-cr-6"
},
{
"id": "claim-25",
"title": "Offline help and verified downloads",
"text": "The laboratory client made no request for help and rejected a tampered artifact before writing it. Workspaces over the tested upload limit were refused.",
"status": "Historical recorded result",
"date": "16 September 2026",
"scope": "Laboratory or bench evidence; production behavior is not inferred from it.",
"result": "The laboratory client made no request for help and rejected a tampered artifact before writing it. Workspaces over the tested upload limit were refused.",
"limitation": "50 MiB threshold and exclusions belong to this tested release. Recording service, not full production integration.",
"href": "/evidence/phase-cr-7"
},
{
"id": "claim-26",
"title": "Review, resume and cancel",
"text": "A fixture that exhausted its attempts entered review, could resume under a newly approved attempt sequence, and released the saved parent when canceled.",
"status": "Historical recorded result",
"date": "16 September 2026",
"scope": "Historical record; any production observation is limited to the dated test described below.",
"result": "A fixture that exhausted its attempts entered review, could resume under a newly approved attempt sequence, and released the saved parent when canceled.",
"limitation": "Fixture backend. Resume only from review. Notification event is not proof of message delivery. Since 12 October 2026 the service sends the review email; this fixture test has not re-run against the deployed service.",
"href": "/evidence/phase-cr-9"
},
{
"id": "claim-27",
"title": "Fresh workspace per attempt",
"text": "The hosted test started each attempt from the same saved workspace and kept the parent parked. Outcome and duration remain per-run data.",
"status": "Historical recorded result",
"date": "17 September 2026",
"scope": "Historical evidence; the scope and environment are described in the recorded result.",
"result": "The hosted test started each attempt from the same saved workspace and kept the parent parked. Outcome and duration remain per-run data.",
"limitation": "No inference call or task acceptance: test accounts had no provider key. Successful-attempt parking was conditional and not demonstrated. One job; checkpoint pauses are host-specific.",
"href": "/evidence/phase-cr-3"
},
{
"id": "claim-28",
"title": "Independent host-side checks",
"text": "The hosted test stopped and extracted the candidate before independent verification against approved tests; it rejected a forged pass log.",
"status": "Historical recorded result",
"date": "16 September 2026",
"scope": "Historical evidence; the scope and environment are described in the recorded result.",
"result": "The hosted test stopped and extracted the candidate before independent verification against approved tests; it rejected a forged pass log.",
"limitation": "One hosted attempt The job ended in review, not acceptance Current deployment independently unverified",
"href": "/evidence/phase-cr-4"
},
{
"id": "claim-29",
"title": "Customer-path hosted job",
"text": "A dated production canary reached a terminal state, verified the accepted download and denied a bystander account. This is one canary, not a current deployment guarantee.",
"status": "Historical recorded result",
"date": "17 September 2026",
"scope": "Historical record; any production observation is limited to the dated test described below.",
"result": "A dated production canary reached a terminal state, verified the accepted download and denied a bystander account. This is one canary, not a current deployment guarantee.",
"limitation": "Monetary spend unavailable; usage does not establish money spent. Limited job/account scenario; acceptance is not a performance promise. Artifact verification applies only to accepted runs.",
"href": "/evidence/phase-cr-8-prod"
},
{
"id": "claim-30",
"title": "Customer credit authorization in test",
"text": "Tests refused inappropriate credit changes by a customer while preserving verified payments and authorized grants.",
"status": "Historical recorded result",
"date": "17 September 2026",
"scope": "Historical record; any production observation is limited to the dated test described below.",
"result": "Tests refused inappropriate credit changes by a customer while preserving verified payments and authorized grants.",
"limitation": "Disposable unfunded tenant; verification granted nothing Point-in-time ledger review only Broader rerun exception Historical, not current exploitability proof",
"href": "/evidence/phase-b44"
},
{
"id": "claim-31",
"title": "Production credit-authorization canary",
"text": "A dated production test refused inappropriate customer credit changes and recorded no corresponding ledger movement.",
"status": "Historical recorded result",
"date": "17 September 2026",
"scope": "Historical record; any production observation is limited to the dated test described below.",
"result": "A dated production test refused inappropriate customer credit changes and recorded no corresponding ledger movement.",
"limitation": "Disposable unfunded tenant; verification granted nothing Point-in-time ledger review only Broader rerun exception Historical, not current exploitability proof",
"href": "/evidence/phase-b44"
},
{
"id": "claim-32",
"title": "Usage settlement in the laboratory",
"text": "A fixture consumed credit once per finalized usage amount, charged only new usage, parked at exhaustion and resumed after a top-up.",
"status": "Historical recorded result",
"date": "17 September 2026",
"scope": "Historical record; any production observation is limited to the dated test described below.",
"result": "A fixture consumed credit once per finalized usage amount, charged only new usage, parked at exhaustion and resumed after a top-up.",
"limitation": "Production running-session parking remained untested Original monitoring count corrected on 18 September Laboratory parking used a simulated service",
"href": "/evidence/phase-b48"
},
{
"id": "claim-33",
"title": "Production usage settlement",
"text": "A historical production correction settled outstanding usage once and recorded the funds restriction. No affected session was running, so production parking at exhaustion remained untested.",
"status": "Historical recorded result",
"date": "17 September 2026",
"scope": "Historical record; any production observation is limited to the dated test described below.",
"result": "A historical production correction settled outstanding usage once and recorded the funds restriction. No affected session was running, so production parking at exhaustion remained untested.",
"limitation": "Production running-session parking remained untested Original monitoring count corrected on 18 September Laboratory parking used a simulated service",
"href": "/evidence/phase-b48"
},
{
"id": "claim-34",
"title": "Dated production agent-session canary",
"text": "A single production account with its own key completed a supervised agent instruction with Claude Code 2.1.251; the no-key control was refused, usage was metered and deletion stopped the tested machine.",
"status": "Historical recorded result",
"date": "30 September 2026; final normal run at 09:16:29 UTC",
"scope": "One team-operated production account using its own provider key and Claude Code 2.1.251.",
"result": "Eight normal-run checks and a separate missing-key negative check passed.",
"limitation": "Dollar cost unavailable; checkpoint erasure, general task quality and every registry-listed workflow were not tested.",
"href": "/evidence/phase-agents-prod"
}
]
},
{
"title": "Deliberately not claimed",
"items": [
{
"id": "not-claimed-1",
"title": "Guaranteed task completion",
"text": "Restoring memory and files does not guarantee that the resumed agent finishes its task."
},
{
"id": "not-claimed-2",
"title": "Arbitrary providers",
"text": "Support must be established separately per provider and translation path."
},
{
"id": "not-claimed-3",
"title": "Production readiness or scale",
"text": "Limited recovery and soak tests do not establish production readiness, continuous failover, or scale."
},
{
"id": "not-claimed-4",
"title": "Zero downtime",
"text": "Measured resume takes seconds; the scoped timings remain data, not a zero-downtime promise."
},
{
"id": "not-claimed-5",
"title": "Guaranteed aider retry",
"text": "The recorded client returned for its stored response in three of four runs."
},
{
"id": "not-claimed-6",
"title": "Universal agents",
"text": "Each harness and behavior level needs its own evidence."
},
{
"id": "not-claimed-7",
"title": "A general resume success percentage",
"text": "The small recorded sample supports its confidence interval, not a universal percentage."
}
]
}
],
"currentDeployment": "Not independently verified by this historical evidence.",
"canonical": "/evidence/claims"
}
Download this public summary
This derivative has its own digest. It is not the original transcript or a new test. Historical original records are preserved privately.
Download public JSON →SHA-256 b882ee2a1d353cc0e0f2c53bbf06d11982c4ada4f9b857d483905896bd58f83d