Every claim. Its test. Its limits.

Full result summary: On 17 September 2026, seven assertions passed in a production-hosted check with a bring-your-own-key account. A job was accepted, an unrelated test tenant could not see it, its event history verified, and the downloaded artifact matched its recorded digest. A successor run that day repeated the checks after client corrections. Monetary spend remained unavailable: usage records did not include money, so zero was not a valid cost claim. Acceptance was an observed outcome, not a required success rate. Artifact verification was conditional on acceptance, and the record does not establish other workloads or later deployments.

Sources: Historical accepted-work test across two provider families (recorded green), Approve the task and its checks before running (recorded green), Retry from the saved state within the approved spending limit (recorded green), Each attempt starts from the same saved workspace (recorded green), Independent verification rejects a forged pass (recorded green), Unresolved attempts remain visible in the job record (recorded green), Job status and recorded results (recorded green), Offline help and verified artifact downloads (recorded green), Review, resume and cancellation after failed attempts (recorded green), A hosted job through the customer-facing service (recorded green), A hosted job through the customer-facing service

Full result summary: On 28 July 2026, a laboratory Claude Code session survived a forced stop and resume. Its records reconciled, and a completed interrupted response was replayed without being billed again. The test also detected deliberately degraded replay and record behavior. The result depends on the client returning for the held response; it is not a promise about every agent’s retry behavior. The specimen’s task outcome was reported separately and did not determine whether the infrastructure test passed.

Sources: Historical interrupted-call replay test (recorded green), Historical second-agent compatibility test (recorded green), Historical interrupted-call replay test

Full result summary: On 28 July 2026, a laboratory Claude Code session survived a forced stop and resume. Its records reconciled, and a completed interrupted response was replayed without being billed again. The test also detected deliberately degraded replay and record behavior. The result depends on the client returning for the held response; it is not a promise about every agent’s retry behavior. The specimen’s task outcome was reported separately and did not determine whether the infrastructure test passed.

Sources: Historical interrupted-call replay test (recorded green), Historical interrupted-call replay test

Full result summary: On 30 September 2026, a production canary created an agent session, bound the account’s own provider key, started its supervised machine and completed an instruction with Claude Code 2.1.251. Eight normal-run checks passed. A separate negative check refused a model call without a bound key. Usage showed a metered call; dollar cost remained unavailable because no approved model rate existed. Deletion stopped the tested machine, while checkpoint content remained subject to retention. An earlier run printed an empty runner label; a corrected final run followed. This one-account demonstration does not validate every agent, adviser, budget, approval or recovery workflow.

Sources: Agent-session production canary on 30 September 2026 (recorded green), Agent-session production canary on 30 September 2026

Full result summary: On 28 July 2026, laboratory checks ran a session through checkpoint, forced stop and resume while preserving its in-memory witness and continuing counter. Repeating the test confirmed the lifecycle could run again, and substituting a fresh start failed the continuity checks. This test did not verify clock resynchronization; a separate general-image demonstration was reported as data, not an asserted result. The evidence concerns the recorded laboratory lifecycle, not current hosted availability.

Sources: Historical session lifecycle test (recorded green), Historical checkpoint portability and integrity test (recorded green), Historical production checkpoint-and-resume canary (recorded green), Historical session lifecycle test

Full result summary: On 8 September 2026, three production-hosted assertions passed for deletion of a synthetic tenant. The customer-facing deletion operation destroyed checkpoint access, removed stored provider credentials and removed the account identity without affecting another tenant. A copied checkpoint was unrecoverable under the stated key-management conditions. The fix was deployed during this work; the result applies from that deployment, not earlier builds. The copy was made for the test because no scheduled checkpoint-store backup existed.

Sources: Deletion verified on the running hosted service (recorded green), Deletion verified on the running hosted service

Full result summary: On 3 September 2026, laboratory checks created owned fork branches, exercised terminal attachment and refused another account’s access. File-divergence checking was skipped because most branch networks were unreachable, and attachment used the stable parent. A separate production narrative reports successful owned forks and keyboard interaction through the hosted client after deployment. Final robustness changes could not be rerun on the damaged laboratory environment. The laboratory assertions and production narrative cover different scopes; neither resolves the recorded branch-network limitation.

Sources: Historical hosted attach-and-fork test (recorded green), Historical branch-isolation and steering test (recorded green), Customer-facing hosted service checks (recorded green), Historical hosted attach-and-fork test

Full result summary: On 28 July 2026, a laboratory guest resumed after its virtual-machine process was forcibly stopped. Its in-memory witness remained the same and its counter continued advancing; a fresh boot failed the continuity checks. The guest used a read-only filesystem, so the witness could not be reconstructed from a saved disk file. This was a small counter guest, not an AI agent or a production workload, and establishes only the recorded memory-continuity result.

Sources: Historical memory-continuity test (recorded green), Historical checkpoint portability and integrity test (recorded green), Historical unattended recovery demonstration (recorded green), Historical production checkpoint-and-resume canary (recorded green), Historical memory-continuity test

Full result summary: On 5 September 2026, a hosted measurement ran 20 warm and 20 cold resumes of a 2 GiB guest in interleaved order. All 40 completed. Warm p95 was 25.09 seconds and cold p95 was 47.46 seconds; medians were 18.08 and 33.36 seconds. The cold result met the 60-second target, while the warm result missed the 5-second target. These measurements describe that guest and environment, not a general latency guarantee or current service level.

Sources: Measured session resume times (recorded green), Measured session resume times

Full result summary: On 17 September 2026, laboratory tests reproduced a defect where session usage did not consume prepaid credit, then recorded correct one-time settlement, charges only for additional usage, parking at exhaustion and resume after a top-up. Production evidence later that day recorded outstanding usage debited once and the funds restriction triggered. No affected session was running, so production parking at exhaustion remained untested. An amendment on 18 September corrected the monitoring count without changing the settlement verdict. The record does not disclose account-linked balances or prove current deployment behavior.

Sources: Historical usage-meter reconciliation test (recorded green), Historical credit-ledger and funds-interlock test (recorded green), Historical usage-statement integrity test (recorded green), Customer-facing hosted service checks (recorded green), Historical production credit-authorization correction (recorded green), Historical production credit-authorization correction (recorded green), Historical prepaid-usage settlement correction (recorded green), Historical prepaid-usage settlement correction (recorded green), Historical prepaid-usage settlement correction

Full result summary: On 8 September 2026, three production-hosted assertions passed for deletion of a synthetic tenant. The customer-facing deletion operation destroyed checkpoint access, removed stored provider credentials and removed the account identity without affecting another tenant. A copied checkpoint was unrecoverable under the stated key-management conditions. The fix was deployed during this work; the result applies from that deployment, not earlier builds. The copy was made for the test because no scheduled checkpoint-store backup existed.

Sources: Deletion verified on the running hosted service (recorded green), Deletion verified on the running hosted service

Full result summary: On 1 September 2026, four recorded tests on hosted hardware reported role restrictions, auditing of privileged writes, consent-bound content access, independently reconciled financial displays and detection of altered records. Missing source data was displayed as unavailable rather than zero. These are historical test assertions about the administrative console. They do not establish the current confidential-document access model or resolve conflicting public statements about whether staff-access controls are implemented. This historical record does not independently verify the deployed staff-access behavior.

Sources: Historical staff-console control tests (recorded green), Historical onboarding and key-vault tests (recorded green), Customer-facing hosted service checks (recorded green), Historical staff-console control tests

Full result summary: On 6 September 2026, four hosted assertions passed across two live provider families. A first-stage candidate could pass independent checks; when a first attempt failed, the workspace was restored and a later stage completed the fixture. Calls were metered, and disabling escalation did not turn failure into acceptance. The tasks were tiny supplied fixtures, and recovery here was at workspace-segment level. Memory restoration was established separately; no acceptance-rate, cost-saving or general workload claim follows.

Sources: Live task completion after a failed attempt (recorded green), Retry from the saved state within the approved spending limit (recorded green), A hosted job through the customer-facing service (recorded green), Live task completion after a failed attempt

Full result summary: The production record dated 8 September 2026 reports enablement on 7 September. Its single assertion passed: all four existing credential entries were migrated to per-account protection and verified to require their account keys rather than the master key alone; a live model request still succeeded. This supports the reported migration at that time. Credential erasure depends on keeping account keys outside the database backup set. The record does not certify later entries, later deployments or independent copies of key material.

Sources: Credential erasure protection enabled in production (recorded green), Credential erasure protection enabled in production

Full result summary: A regression run on 7 September 2026, sealed in the 8 September record, verified that a session with destroyed account key material could not resume while another account’s session resumed immediately afterward without restarting the service. The earlier reported service failure was traced to an invalid test request, not damage to the service. This was an isolated engine test on hosted hardware. It did not exercise the production account-deletion route, which requires its separate canary evidence.

Sources: Historical erasure-failure isolation regression (recorded green), Historical erasure-failure isolation regression

Full result summary: On 28 July 2026, the laboratory environment passed readiness checks for running hardware-virtualized guests and verifying its pinned software. A deliberately altered integrity value was rejected. The result supports that environment’s readiness and integrity checks at that point in time. It does not demonstrate an agent workload, hosted availability or ongoing production health.

Sources: Historical environment-readiness test (recorded green), Historical environment-readiness test

Full result summary: On 3 September 2026, laboratory checks created owned fork branches, exercised terminal attachment and refused another account’s access. File-divergence checking was skipped because most branch networks were unreachable, and attachment used the stable parent. A separate production narrative reports successful owned forks and keyboard interaction through the hosted client after deployment. Final robustness changes could not be rerun on the damaged laboratory environment. The laboratory assertions and production narrative cover different scopes; neither resolves the recorded branch-network limitation.

Sources: Historical hosted attach-and-fork test (recorded green), Customer-facing hosted service checks (recorded green), Historical hosted attach-and-fork test

Full result summary: On 5 September 2026, four hosted assertions passed for moving saved sessions to another host after an isolated service failure. The tested sessions preserved in-memory state, each move produced one event and a customer notice, and duplicate resume attempts were prevented. The test used shared durable storage within one region. Work since the latest checkpoint could be lost. Billing reconciliation on the replacement host was still pending in this record; these results do not prove recovery from every host or regional failure.

Sources: Resume on another host after an interruption (recorded green), Billing state across a real host loss (recorded green), Historical checkpoint portability and integrity test (recorded green), Resume on another host after an interruption

Full result summary: On 31 August 2026, a laboratory sandbox test verified that an authenticated payment notification added credit once, repeated delivery or another notification for the same order added nothing further, and a forged notification was refused. The simulated payment, order and credit records reconciled. A mock payment service and test signing secret were used; this was not a live card purchase or a test of the hosted checkout page. KeepState’s usage statement is separate from the payment provider’s tax receipt. Later live-payment assurance requires its own evidence.

Sources: Historical payment-notification verification test (recorded green), Historical payment-notification verification test

Full result summary: Checks recorded on 1 September 2026 verified installation of the released client, successful sign-in and session start on a fresh Linux virtual machine, and rejection of altered installer/update downloads. macOS checks used a cleaned environment rather than a fresh operating-system installation. A 2 September rerun after the repository transfer verified the new distribution paths and compatibility redirects. Windows installation was not covered. These results apply to those historical client releases and environments, not all current installers or supported platforms.

Sources: Historical public-client installation checks (recorded green), Historical public-client installation checks

Full result summary: On 7 September 2026, three hosted assertions passed with the actual tested coding-agent application and real tool use. A fixture task continued after a checkpoint-boundary provider change and produced an independently verified artifact. Lineage and metering remained continuous; a disabled continuation path failed closed and fallback completed after restoring the workspace. The agent ran in an isolated copy of its environment, with transport tested separately. Quality remains model-dependent, and streaming was buffered rather than incremental.

Sources: Native coding-agent continuation across providers (recorded green), Production check of cross-provider translation (recorded green), Native coding-agent continuation across providers

Full result summary: A confirmation run on 6 September 2026 completed one bounded coding fixture through each of two live provider families. The initial attempts failed; later attempts produced accepted artifacts after independent checks. Seven checks passed, including complete-job accounting. Publication was limited to a scoped authorization: no external pull request was created. A negative control showed that verifier independence mattered. The sample is two task arms with deliberately constrained initial attempts, not a general task-success rate, comparative model ranking or cost-savings claim.

Sources: Historical accepted-work test across two provider families (recorded green), Metering and replay on an additional provider path (recorded green), Change models only while a session is saved (recorded green), Live task completion after a failed attempt (recorded green), Historical accepted-work test across two provider families

Full result summary: On 14 September 2026, the public production status reported self-service signup enabled and operational. Separately, the QA team attested that real-provider signups had completed, automatically activated accounts and received the signup credit during the preceding week. These are different evidence types: the status observation proves the setting was enabled; the successful-signup claim is a team attestation, not an independently driven browser test in this record. Current signup behavior requires a fresh end-to-end check.

Sources: Historical signup-credit and free-session test (recorded green), Historical hosted-client and provider-key trial (recorded green), Historical payment-notification verification test (recorded green), Self-service signup: status observation and team attestation (recorded green), Self-service signup: status observation and team attestation

Full result summary: On 9 September 2026, a production transactional-email check recorded a green result. The provider reported the sending domain authenticated, a DMARC policy was present in public DNS, and the deployed service’s message was accepted by the mail provider. The check did not independently derive every DNS authentication record or measure inbox placement. It did not establish delivery to a subscriber list; subscriber fan-out and an unsubscribe route were absent in this record.

Sources: Email is recorded before delivery is claimed (recorded green), Production transactional email sending check (recorded green), Production transactional email sending check

Full result summary: The record resolved on 30 August 2026 reports that the existing readiness, memory-continuity, lifecycle and checkpoint tests passed on the hosted x86 platform. Their test definitions were checked for consistency with the laboratory versions. The first run failed because its corruption test did not actually change the selected data; after that test was corrected, the negative test passed. This demonstrates the recorded platform transition, not continuous availability, portability across arbitrary architectures or a claim that the first failed result never occurred.

Sources: Historical deployment-platform portability test (recorded green), Historical deployment-platform portability test

Full result summary: The original record of 31 July 2026 and rerun of 30 August tested a second agent, aider, and a fresh laboratory installation. The gateway retained the interrupted response in all four reported runs; aider retried successfully in three. That difference is a material client limitation. The second installation used another virtual machine on the same physical computer. A later attempt was not counted as a completed trial because its environment was not ready. These results support the named agent and setup, not compatibility with every agent or guaranteed resumption.

Sources: Historical second-agent compatibility test (recorded green), Historical second-agent compatibility test

Full result summary: On 28 July 2026, a laboratory Claude Code session survived a forced stop and resume. Its records reconciled, and a completed interrupted response was replayed without being billed again. The test also detected deliberately degraded replay and record behavior. The result depends on the client returning for the held response; it is not a promise about every agent’s retry behavior. The specimen’s task outcome was reported separately and did not determine whether the infrastructure test passed.

Sources: Historical interrupted-call replay test (recorded green), Historical interrupted-call replay test

Full result summary: On 30 August 2026, a hosted-service test used a fixture credential and a simulated provider. The session completed its request using the account’s supplied key, the inspected logs and stored records contained no plaintext copy, and a session without a usable key was refused. A negative control confirmed refusal when the key could not be decrypted. Assertions were appended on 6 September 2026. This record did not make a live provider call and does not establish that arbitrary credentials can never leak.

Sources: Provider-key isolation and safe refusal without a key (recorded green), Historical hosted-client and provider-key trial (recorded green), Historical onboarding and key-vault tests (recorded green), Customer-facing hosted service checks (recorded green), Provider-key isolation and safe refusal without a key