Public methods, pricing assumptions, retention commitments and missing capabilities.
Dated 9 September 2026, this pricing-methods version corrected the earlier waitlist-only description and used customer-supplied billed storage instead of assuming a benchmark’s reduction ratio. Estimates were not quotes or savings results; provider costs were separate and an unknown provider amount left the combined total unavailable. Recorded billing terms included per-minute session billing, statement-level rounding, decimal-GB storage and customer-favorable resolution of unreconciled results after 24 hours. Taxes were handled separately. This version is superseded and is not a current price or purchase offer.
9 September 2026
SHA-256 9a52c0e2503cb72759e8863b4e761d2210516deea03e7a207390718cefb4b770
Dated 9 September 2026, this amendment retained the previous rates and billing rules while adding per-tier session-budget defaults. It also restored the earlier version’s published bytes after an unversioned change had broken their reference. Estimates remained separate from actual statements; provider costs were separate, unknown inputs were not treated as zero, and managed inference was not purchasable. The record is superseded. Its historical defaults and amounts should not be used as current product limits or purchase terms.
9 September 2026
SHA-256 dbc65b8ce2e47fc243817089ab25294b6a2180b95fed1fb798354784cff33fcf
This methods record is dated 9 September 2026; its price-book amendment is dated 14 September. It describes estimates rather than quotes or savings results. Provider charges remain separate, and an unknown provider estimate makes the combined total unavailable. Session time is measured by second, billed by minute and rounded at statement level; storage uses decimal GB averaged across the billing month. Unreconciled results are not billed and resolve in the customer’s favor after 24 hours. Statements are pre-tax; Polar handles taxes. Managed inference was not purchasable. These are recorded terms, not a current-price verification.
9 September 2026
SHA-256 32e5814674f73176cf49a567cf54671a06fb3dc2031eb4ae62cf4a64407fa477
This undated, superseded record described planned prices and a waitlist-only offering. Its outputs were estimates, not purchasable terms or observed customer savings. It separated provider charges from KeepState revenue and warned that a storage assumption came from one measured workload, so customer results would differ. Later records replaced those assumptions and the waitlist posture. This page describes the earlier estimates and their limits; the planned figures are not current prices or commitments.
unavailable: this record declares no tag to date it by
SHA-256 49eace097a92a8e8cc7e28756156951a52e715bae381abbbf44c6de4814f52b0
This undated public derivative describes six pinned build components and records that five download-address fields were removed. It includes a digest tying the derivative’s provenance to the earlier source record. One component entry mentions an image obtained on 28 July 2026; that is not a date for the whole record or a new verification run. The record supports historical reproducibility only. It does not certify current component versions, vulnerability status, deployment state or the completeness of a software bill of materials.
28 July 2026
SHA-256 4b859c7a4c0f01a7af3f85a0cf7664fb1a085bab2f344fc8174209e27b090ba3
The 13 October 2026 record retains provider, architecture, checkpoint-version and recovery limits. The recorded agent-session runner is Claude Code 2.1.251. Current restrictions include customer-supplied provider keys, different token limits for different call paths, a short approval window after which a request is closed, an ineffective Cruise spending ceiling while usage is unpriced, and Cruise acceptance that rests on tests the checked code can influence. From 20 September to 6 October 2026 a call retried after a resume could be billed twice; the fix is in production and its paid gateway test re-ran green on 13 October 2026. Agent budget pausing and production credit exhaustion remain unverified. Historical retention gaps, 30 open findings and the small reliability sample remain visible below.
13 October 2026 source update; older measurement dates remain distinct.
SHA-256 8c25407f8553960c0fbb4ced3a7ec406444a244e1e5c329ac4941dff9f13e6a3
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.
3 September 2026
SHA-256 58f778bbb7b374c2d95744cb02da09754eea0b8bf1c7c6cf668d1e7202f0abb7
On 16 September 2026, four laboratory assertions passed for Cruise preparation. The client drafted a task manifest, required explicit approval, and refused execution if the approved task or tests changed. A repository without discoverable tests was rejected. Preparation and approval made no service requests in the test. Discovery covered pytest and Go tests only; other test systems were not established by this record. This is a dated client-contract result, not current production verification.
16 September 2026
SHA-256 1454b727e3458527efdf48514e58249646da9795cf48633651a5a64541c63f5f
On 17 September 2026, six infrastructure assertions passed for isolated attempts from one saved workspace, restored content, periodic save points and cleanup on hosted infrastructure. The addendum records that every attempt failed: the test accounts lacked the required provider key, so no inference call or task acceptance occurred. Successful-attempt parking was conditional and was not demonstrated. This one-job run supports the tested infrastructure behavior, not task success or acceptance rates. Checkpoint pauses remain host-specific measurements.
17 September 2026
SHA-256 232922814bcaa8ccc750199e94823390bbeef282e368271a6c3d029801382d6f
On 16 September 2026, a hosted-service test recorded four passing checks: the candidate was extracted and the child stopped before verification; the verifier ran on the host using approved tests; verdicts carried signed test references; and a forged pass log was rejected. A negative control showed that compromised verifier independence could invalidate the result. The tested job ended in review after its single attempt failed. This establishes the tested verification boundary, not universal task success or current deployment state.
16 September 2026
SHA-256 06847a6d6cd62b195eecaf4d0242658d1029af5b0882e0832921f58d97526d1d
On 16 September 2026, a laboratory fixture completed a three-step job after two failed attempts. Each retry started from the same saved state. The test refused an unsupported model choice, stopped when the remaining budget could not cover another attempt, and made no retry after the final attempt. A negative control detected a disabled budget check. These are fixture results, not measured production costs or a savings guarantee.
16 September 2026
SHA-256 2414e8210d3cda40ad5368e754b6b11819673550635b2d3d760687432f6b1ae0
On 16 September 2026, four laboratory assertions passed for job accounting. Attempts progressed through their recorded lifecycle, reported spend matched reconciled attempts, and the event history passed its integrity check. When usage remained unresolved, the attempt stayed visible and the job could not be accepted. A negative control caught an attempted acceptance despite that ambiguity. These checks used a fixture service with a simulated reconciliation timeout; they did not measure a live provider’s reporting delay.
16 September 2026
SHA-256 985057b9be5adf13e922b3fa7a83aa955efd78ceaac489c955812878c139b382
The original laboratory record is dated 16 September 2026. A 17 September amendment changed test funding after customer credit grants were disallowed; the revised normal and negative-control runs each passed eight checks. They rejected a changed approval digest and a second concurrent free-tier job, verified event history, prevented acceptance of unresolved attempts and resume outside review, and displayed missing spend as unavailable. Customers could not grant themselves credit. These local tests used a test identity service and do not establish current production limits or availability.
16 September 2026
SHA-256 020bc61bf22b83c839f5f2627c5abdb843c4b5b01b6dc3b7825ed3a9226b8df5
On 16 September 2026, five laboratory assertions passed for the Cruise client. Twenty help invocations made no service requests, status and logs preserved the recorded states and event order, and a modified artifact was refused before writing an output file. Workspaces larger than 50 MiB were rejected before upload; the measured size excluded Git and KeepState metadata directories. This record covers the tested client contract and recording service, not every current client release.
16 September 2026
SHA-256 2eaa77fe705e1e7c5937d7dd854f4374caea3cbc47db8e798bd6d7202481fc4c
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.
17 September 2026
SHA-256 b0fb11b0d92b6d48cb970989c9ed8d20a376425b44b17decdbe09b92516946d7
On 16 September 2026, four fixture-based laboratory assertions passed. Exhausting the approved attempt plan moved a job into review with a notification event. Extending the plan from review allowed the tested job to complete; cancellation released the retained workspace. The final permitted attempt was not silently retried. Resume was allowed only from review, and the retained parent remained the customer’s recovery option until release. This record does not prove live notification delivery or production availability.
16 September 2026
SHA-256 2c6a20cd20025c649c26f4e362de87e46762e75b46ac1e62046ba060090758ed
On 3 September 2026, a team demonstration on the production fleet created three branches from one checkpoint. A file written in one branch was absent from the other branches, and their records retained the fork relationship. The demonstration used a base image without model calls; it was performed through team tooling, not the then-available public hosted client. Test sessions were cleaned up and no customer was billed. This demonstrates that specific fork example, not general branch performance or present client availability.
3 September 2026
SHA-256 9a90576a95e3da2aab9c2f607a27828ad1661dc3c5335d56d573722f7cfecf55
On 30 August 2026, a two-account test on hosted hardware verified that each owner could operate its own session and that forged or cross-account attempts to read, attach, checkpoint or resume were refused without changing the target session. The test exercised the engine’s account-ownership boundary using test-generated account assertions. It was not a full end-to-end test of external sign-in or today’s document portal. Legacy laboratory sessions had a separate local-access scope, explicitly outside the owned-session claim.
30 August 2026
SHA-256 be2d29315571159976fe6d5365917f3db0d29dd6ad395cd8b8796c2ae4cd09e5
On 30 August 2026, a test on hosted hardware used an account token to create, read and checkpoint an owned session through the control plane. Another account’s token and a revoked token were refused. The tested components were colocated; public-domain transport, real Google/GitHub sign-in and real provider-key use were not established by this test and awaited separate setup. The result demonstrates that recorded ownership and token-revocation path, not a complete customer onboarding or payment journey.
30 August 2026
SHA-256 ce4b8500cf2d5d6a2c7db30c7fee4e1c0042f73a085bf57c8f4a6e84fc90d122
The record sealed on 31 August 2026 reports checks run on 30 August showing that the authorized control plane could create and read an owned session across the private network. Public and unauthenticated transport attempts were refused, including plaintext and untrusted client certificates. Network restrictions and certificate authentication were tested independently. This was a same-region control-path test. It did not test moving a checkpoint to another host, recovery after host loss or today’s network configuration.
31 August 2026
SHA-256 897f8b1406c700fdbd6174f0b4fc79c5f1c7755e56fc3699347f8f90cd46ab75
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.
30 August 2026; 6 September 2026
SHA-256 b1a0f63a324ed0a3a74acc6a4b74281078f38742c354244fc696dbfa3792a364
On 31 August 2026, a hosted-hardware test verified that the services restarted after a forced process stop and returned automatically after the host rebooted. A previously checkpointed and stopped counter session resumed with its in-memory witness preserved and its counter advancing. The negative test detected the absence of automatic restart. This test used a small counter specimen and a saved checkpoint; it does not promise preservation of unsaved work, arbitrary workload recovery or failover to another host.
31 August 2026
SHA-256 e4040a90b7708c38726a03565a683814368f16aaae13d0cd7a93fb0768dcfec9
On 31 August 2026, three hosted assertions passed for the termination guard. An ordinary termination request with unsaved work was refused and the session stayed alive, with safer choices explained. Explicit forced termination proceeded; a session with a saved checkpoint could be terminated normally. The protection was enforced by the service as well as the client. This guard protects unsaved work, not all destructive actions, and an explicit force request can still discard that work.
31 August 2026
SHA-256 e1ff85ed86fd97e830f9743ec86cc8ff1cca7378dad6613c14eae2edcb56aeef
On 30 August 2026, laboratory checks verified that an account and its session survived a service restart, deletion removed identifying account data while retaining a tombstone, and the deleted account’s session was refused. Altered sign-in callback state created no account. Session-cookie protections were also checked. The login round trip used a test identity provider; Google and GitHub redirect construction was checked without contacting those providers. This is not a live-provider authentication test or proof of document-specific authorization.
30 August 2026
SHA-256 d46069f772e886b2f75278b0d2e54f9b2e76f107f6772dea0ea0c39d5b38e037
On 30 August 2026, laboratory tests completed the client’s device-login flow, bound the resulting token to the approving account and verified that revocation took effect on the next request. Expired and mismatched approval codes were rejected. The inspected logs and database did not contain the fixture token; it appeared only in the tested client credential store. This exercised the Linux storage path. The macOS keychain path was not tested here, and the result is not a general claim that no token can ever leak.
30 August 2026
SHA-256 4530775c77edfd4b4fa1053b4e2445eca4a901335c2bd6c59c915aa8112e4591
On 31 August 2026, laboratory tests verified that low-credit alerts were calculated from server-side account records, appeared once in the console and were not falsely marked emailed when email was unavailable. With a mock delivery endpoint, an alert reached that endpoint before being marked sent. Live authenticated email was outside this test. A 17 September amendment changed test funding; its attempted rerun was not held because the laboratory environment was unreachable. The original result remains historical, and the aborted attempt is not a passing rerun.
31 August 2026
SHA-256 3df439a8dee204ee8f39dce19216cd7acdc01e8d6dadd66086513ed366e423e2
On 31 August 2026, a laboratory test recorded three passing assertions: accounts could not activate themselves, authorized activation was audited, and queued accounts received no start instructions while active accounts received usable welcome instructions. A negative control confirmed that the activation restriction mattered. The welcome test checked instruction structure, not the production destination. This historical result does not establish that the same activation policy or interface is deployed today.
31 August 2026
SHA-256 f2b0cd36ddc6995566bb360f28f8130fbf386ce9346f8967eb7f2db514067abc
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.
1 September 2026
SHA-256 a38739bae41227f07902ddc560df26edffbc6f31e824f648dfd2fc41b783d931
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.
30 August 2026
SHA-256 83a0ee51f94eece1ac50ee63e483fdde54bad0e2695ba21f75c25c6343caf164
The record sealed on 31 August 2026 reports hosted-client checks run on 30 August: a token created a session on the live fleet, read its usage information and stopped it. A revoked token was refused before creating a session. A separate team-operated trial on 31 August completed a real model call using the team’s own provider key. That real-key trial is narrative evidence, not one of the automated gate assertions. These two scopes are distinct and do not establish behavior for every workload or current release.
31 August 2026
SHA-256 cb379a2f2f4864d1de4e8c15da0c5a7ea63c1930f5d38267f99f0dc2f2de250a
On 30 August 2026, laboratory checks verified that onboarding state survived a restart, stored a fixture provider key in encrypted form and returned only a masked key reference. Cross-account attempts to read or change the vault were refused. The service also refused an unsafe key-storage configuration. The fixture key was deliberately not a real provider credential, and no live provider call was made. The evidence applies to those inspected responses and stores; it does not certify every current production secret-handling path.
30 August 2026
SHA-256 5b4bad3840840f073027fc67f28bc0eb0d20594319b4abe023d4fbd7b0f79faa
On 31 August 2026, a test on hosted hardware reconciled a usage statement against its token, session-time and storage records. The total matched its priced lines, an altered meter record failed reconciliation, and an unreadable source was shown as unavailable rather than zero. A real billed model call supplied the token measurement. This test did not exercise card handling or checkout. It verifies the recorded statement and reconciliation checks, not today’s prices or every future bill.
31 August 2026
SHA-256 56ff6b06be45b38a3b73beb2816ebaac255b2bf49eada0216d6a18bccf27097a
On 31 August 2026, a hosted-hardware test with a real virtual-machine session verified that repeated credit operations did not duplicate a debit and that exhaustion checkpointed and parked the session, which later resumed after a top-up. A 17 September amendment changed the test because customers were no longer allowed to invoke credit-changing operations. The amended ledger checks run in isolation, with settlement supported by separate evidence. The original assertions describe the earlier interface; they are not proof that those customer operations remain available or that the amended test reran in production.
31 August 2026
SHA-256 9990196fe693c08a2324dac00162105a71c993f17d25e1d8c5d445325da30361
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.
31 August 2026
SHA-256 8e43b99b24707f5bccaea226e7008d93d44ac16377cdc2e388e5c29a9b740154
On 31 August 2026, a test on hosted hardware verified that a downloadable usage statement matched its source statement and priced lines, retained a verifiable relationship to the prior statement and detected alteration. It was explicitly a usage statement, not a tax receipt. A 17 September amendment changed how the test account was funded after customer credit-changing operations were closed. This version says the amended rerun was still pending an authorized test window. The original passing result does not establish a completed amended production rerun.
31 August 2026
SHA-256 6894c7cff44a269405d171482ebf7342ca2c3fe00f03d4b36a3b44065e7b8289
On 31 August 2026, a hosted rehearsal passed all ten steps in the recorded customer journey and two summary assertions. It covered activation, a supplied provider key, crediting, model use, pause/resume, statement reconciliation and a refund returning the balance to zero. Revoked access and a modified statement were rejected. The identity provider and payment service were test systems: this was a rehearsal, not proof that a real new customer completed live checkout or received a real refund.
31 August 2026
SHA-256 af6f562e09a21d0a0b8f334c8281da5e1512faf2685245d34960bcb38b325130
The original record dated 2 September 2026 reports three passing assertions across laboratory and hosted email checks. Simulated interruptions left mail queued, and sent status required provider acceptance. The provider verified the domain’s email authentication configuration and accepted a welcome message. The interruption tests used a mock provider; provider acceptance does not prove inbox delivery or receiver authentication headers. A 17 September amendment changed test funding, but its rerun was not held because the laboratory host was unreachable. The amendment therefore adds no new passing run.
2 September 2026
SHA-256 fd9fbe4515d8da15381028030521b702d59f1e013248f845ea8f63717f36d069
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.
1 September 2026
SHA-256 a32be7100c2a323ddb0f016ab48f6e5fcf1dd7f1e903a38a366da8288feb033f
On 1 September 2026, laboratory tests with real virtual-machine sessions verified automatic account activation, a one-time signup credit, concurrency limits and checkpoint-and-park behavior. Repeat signup by the same authenticated identity did not receive another grant. The record’s parking behavior was based on elapsed runtime, not inactivity, and some abuse controls were approximations. Production self-service was disabled in this historical version. A 17 September amendment changed the credit-exhaustion check to an isolated internal test; that amendment is not itself a new full production run. This record does not establish current signup grants or availability.
1 September 2026
SHA-256 c0ef207bcedf9d05a325c3c0cb86051c382e48273a108a7865c52d61c3a6bb84
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.
5 September 2026
SHA-256 6631f8f6109f55df3790dee38126e6a3eca82f630988e87b8ace9d5a1461be82
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.
5 September 2026
SHA-256 9a6423ffce297074b0f03f3c4e2f496c99f8221735b3d74c5301659db153210a
The body records website checks on 6 September 2026, although its viewer metadata omits a date and environment. Built pages used source-derived evidence titles and rejected a claimed result with an empty evidence reference. A separate served-page review found references on all 19 checked proven rows and confirmed corrected catalogue and security wording. These are publication-consistency checks, not fresh tests of the underlying engine or security controls. The record contains no machine-readable overall verdict or assertion set.
6 September 2026
SHA-256 5bc776872d32ff3abf5082f3b89375984b01f8eaae3a0701e73e2e6f88ea7564
On 7 September 2026, four assertions passed using production binaries with disposable data. The customer deletion path removed credentials and made checkpoint data unrecoverable through the tested backup-recovery path. It did not report deletion if key destruction could not be confirmed. A second tenant remained restorable. Production services were not changed by this test; it exercised their deletion code in isolation. Recovery protection depends on keys being excluded from the restored backup.
7 September 2026
SHA-256 09d74e2b1a816de75bcca80dac66623a99b60bfd0d841d520089397bac1560b1
On 7 September 2026, an engine test on hosted hardware verified that an opted-in session produced an encrypted transcript readable with the customer’s key, while an opted-out session produced no transcript in the record subsystem. Alteration of a recorded entry was detected. This record explicitly says the feature was not enabled in production. The separate, bounded response-replay cache is outside the transcript subsystem, so the opt-out result is not a claim that no response bytes can exist anywhere. Production enablement requires separately reviewed privacy disclosures and evidence.
7 September 2026
SHA-256 9cd0d2e8854355f5d364e6535d0bd1d9e2b497d74dd171f74114efd43f772452
On 7 September 2026, four hosted assertions passed for continuing a tested task across provider protocols at a checkpoint boundary. The live continuation included tool use and an independently verified artifact, while lineage and metering remained continuous. A failed continuation was refused and the prior workspace restored for fallback. Deterministic translation checks were separate from the live run. The result is fixture-specific and model-dependent; streaming translation was unsupported and explicitly refused in this version.
7 September 2026
SHA-256 2547131d5db0b3bd8e73c397df53cc52418b9f142a5e7bf0f83cbd762130429a
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.
7 September 2026
SHA-256 d348669d507b6c35a41557a69431d6bddb89d6ee2cc9b5623e98d054a0306469
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.
8 September 2026
SHA-256 41fcee28dc6d20ea359799d82914ce5e40390caf1653aea131e5e8e5513047a8
The credential-erasure record dated 8 September 2026 reports three passing hosted test assertions. A database backup recovered the test credential before account deletion but could not afterward, even with the master key retained. A negative control confirmed that destroying the account key was necessary. The base test required older entries to be rewritten; an appended production section dated 7 September reports migrating and verifying all four existing entries. Erasure protection depends on keeping account keys outside the database backup set. The record does not establish later deployments or independent copies of those keys.
8 September 2026
SHA-256 57eb51e7fa1125f56189d60e13cf78f12b14f54f7dae6e582af25ce529baf88b
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.
8 September 2026
SHA-256 5500bb986e45d5804edb8062749a20f73b60158ba4ca55a008c03524d6310a85
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.
7 September 2026
SHA-256 47b3b86414279376a3aaf1c568da6239523c3cb911b51d21bd4ac573fd17cd26
On 7 September 2026, deterministic checks converted a real Claude Code transcript into inputs for other agent tools and produced a separate task handoff. The record reports the conversation information lost in each conversion. No live target agent was run in this test, and no successful task continuation is claimed. Each supported source/target pair needs its own live completion test before it can be advertised as harness switching. The task handoff remained usable independently of transcript conversion.
7 September 2026
SHA-256 2f0a1d44b9aa33478874550ef3acf6df5d8ed8e42b2d53eebd27a12db1e7417b
On 8 September 2026, regression and production-canary evidence reported that deleting a synthetic account removed its active-session resources before destroying its key. The prior behavior left unusable session residue and caused health checks to fail. The corrected deletion refused to report success if teardown failed. Billing-related tombstone records were retained under the stated policy. The production check used a synthetic tenant; unrelated resources from a concurrent test were cleaned separately. This is dated evidence for that cleanup path and does not remove later open account-deletion findings.
8 September 2026
SHA-256 26704373c42bb529c2c8034fdabadbee388e195c1ac2610992338f5c7c6ae413
The run completed in September 2026 and was sealed as failed on 8 September. On the encrypted hosted configuration, a six-hour segment restored its hours-old state and stayed within the other measured durability limits, but only 71 of 72 checkpoints completed because one snapshot exceeded its deadline. The record therefore remains failed. A later amendment points to a successful rerun completed on 19 September; that result belongs to the separate rerun and does not erase this defect or its original outcome.
2026-09-08
SHA-256 48b945792f6afc7ef2d862c55ba75626a8f8c8ae612ee93916bd4dded01a65f4
The 8 September 2026 record reports a correction for checkpoints that exceeded the previous fixed deadline. Tests verified an adjusted time allowance, bounded recovery after a timeout and explicit recording of failed checkpoints. The page reports deployment verification before the later soak resumed. Measurements came from one host and one storage class; the calibration did not track storage that degraded while the service was running. A prematurely started soak on the previous build was stopped, so it is not evidence for the corrected build.
8 September 2026
SHA-256 897850d31ccec495eb30fd47f4dabd0d6a901c7c17e0af15210743bd0bf9c873
On 17 September 2026, laboratory tests reproduced and then rejected unauthorized customer credit changes after the correction. A separate production check with a disposable unfunded account confirmed that customer credentials could not change credit, balances stayed unchanged and legitimate ledger reading still worked. A point-in-time ledger review found no unexpected grant-reference categories in the surviving records. That review is not a guarantee that no misuse ever occurred. Some broader laboratory reruns were deferred under a documented exception; the scope of this production check remains credit authorization.
17 September 2026
SHA-256 fca0972967cc45b032dff1f9c96c425edc6ff3e3f80ec9cf9e307e1d7bd8f695
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.
17 September 2026
SHA-256 ed9a2bc43183f70250b7a8c1c02ccedf9b0a3bbeaf67e48d6c8aee028221e870
From 18 to 19 September 2026, one hosted six-hour segment completed all 72 checkpoints after the earlier deadline defect was corrected. Six recorded assertions passed, covering the segment duration, checkpoint count, cadence, bounded memory and absence of checkpoint failures, plus a control against pooling unrelated runs. This rerun used one guest size on one host. It does not turn the earlier failed run into a pass or establish reliability for larger guests, other workloads or the current deployment.
19 September 2026
SHA-256 a747d5e220ecf35d71cf2ba131a3d0f92aec1f4b4d967e02f33eae131c24a754
On 18 September 2026, isolated tests reproduced a wrong-unit configuration that delivered less signup credit than advertised, then passed after the correction. The new checks rejected the mistaken unit and used the approved price-book value when encountering the stored error. The production section reports deployment and remediation of the undercredited account; repeating the remediation did not add credit again. The automated assertions used an isolated test identity provider with no live fleet or model calls. Production remediation is a separate recorded observation, not an independently repeated audit test.
18 September 2026
SHA-256 d2d9e4c7ce6cf291fcf7660e4a016cbc777e59ea4db472ea9ae1f0886a7811bf
On 18 September 2026, isolated tests verified that an old credit-setting override stopped governing after the approved price book changed, with the reason made visible. An explicitly reconfirmed setting could then apply under the new version. The record reports production deployment, but states that no override was active, so the deployment itself did not exercise removal of a stale override. The functional proof came from an isolated test identity provider without live fleet or model calls. This record does not define today’s signup-credit amount.
18 September 2026
SHA-256 6583cd016cb25ee517dd3639040d9bdc91f90657aac2f5c6076a8bdfaae5bad5
The 18 September 2026 record discloses that some Google signups did not receive the credit advertised on the website. Isolated tests reproduced the failure and then verified that eligible signups received the approved credit even when their identity provider supplied no account-age information. The production section reports deployment, correction of affected Google accounts and clearing of stale low-credit alerts. The automated tests did not use live identity providers or the fleet. This record preserves the customer impact and reported remediation without publishing account identities or balances.
18 September 2026
SHA-256 54bdfab73e735fb16d69c4d6dd8584ae74b434b60d4529183adb0db5260c21e2
On 18 September 2026, the activation test was corrected to use the approved credit amount and to verify a real park before claiming resume success. Repeated laboratory runs then passed. A new review of billing-test freshness found additional unresolved items: the payment test’s measurement had become stale, the end-to-end paid rehearsal had not been rerun and an earlier soak remained failed. The freshness review therefore remained red. Successful activation retests did not resolve those separate assurance gaps, and an older green payment result was explicitly retained as historical evidence.
18 September 2026
SHA-256 140928397e2c591dc240034a5821f3ac62c49234d8e1ca7976a09094cfc52eb3
On 19 September 2026, a synthetic tenant used the production customer routes to create, checkpoint, stop and resume one session. Its memory-backed witness persisted and its counter continued; the completed run cleaned up its test resources. This used one image, one guest size and one resume, with synthetic paid status. It tested the checkpoint route’s orderly stop, not a hard machine kill. Two earlier aborted attempts leaked sessions and were recorded as an open account-deletion issue; the successful cleanup does not erase that limitation.
19 September 2026
SHA-256 ec50e9d1246c836a3b8b3083705fd8bd194837d3b06822de5de5ee88d7e69241
On 8 September 2026, one production-hosted assertion passed for a real translated request on the team’s account. The request reached a live compatible model, was recorded and metered, and the service stayed healthy. This was a one-call production check, not a broad task-success study. Streaming responses were buffered rather than incremental, and replayed thinking blocks were omitted during translation. Those limitations remain part of the result; current behavior requires separate deployment verification.
8 September 2026
SHA-256 cca23e3786f1c5bb9bae39dc886d15652c8ab1f0c8a15acc3ea18c0f1e2778ce
On 8 September 2026, six production-hosted assertions passed for synthetic tenants. The check covered authenticated account data, onboarding redirection, write-only credential submission, an authorized session fork, attachment negotiation and billing displays. An unrelated real tenant remained unchanged. Limits matter: the credential was synthetic, the paid state was simulated, attachment covered the handshake rather than an interactive session, and the billing account had no spend. The rendered console and larger fork fan-outs were not verified.
8 September 2026
SHA-256 2ae6457ebb7ac25567c8fb2ec1085099ed306f2ded7102c1d7eb2f5ba0b6f4c3
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.
14 September 2026
SHA-256 e00b21da676fcc4d1de9f931ee263c9978a731510e30db0ff53a94db4053439b
Generated on 10 September 2026 for release v0.1.7, this historical snapshot lists 78 test records: 76 marked green, one not requiring a rerun and one failed. The failed encrypted checkpoint run completed 71 of its required 72 checkpoints. The open-durability list was empty in this snapshot; that does not establish absence of defects or override later findings. Each record has its own test date, environment and limitation. This archive is not a current release status or a universal product pass.
10 September 2026
SHA-256 ca697430b34bd0936c5cc746253ba79b2cf92219acba05e32381c8547db64915
This historical snapshot preserves the test-status register as published on 10 September 2026. It includes a failed encrypted soak that completed fewer checkpoints than required, and a prior soak rerun marked unnecessary for a narrowly scoped change. Those statuses describe this snapshot, not today’s deployment. Green entries apply only to their recorded assertions and environments. The snapshot does not carry one overall execution date; use each underlying test record for its date and limitations, and the current register for later results.
10 September 2026
SHA-256 bd784a8e471c74a50f1bf5047f08183f76f147f4d725b1891b7800a45d9e7192
This superseded record is labelled with the 31 August 2026 release and includes runs and an amendment dated 7 September. The first hosted six-hour segment restored an hours-old checkpoint and kept memory bounded, but produced 68 of the required 72 checkpoints and remained failed. After correcting the test driver, a separate six-hour segment achieved 72 checkpoints and restored its earlier state. The original failed segment and later passing segment are separate runs; they do not demonstrate continuous success across both segments.
31 August 2026
SHA-256 ae217a8e932d1c3c67831f71f63631e8e5dc671369d2609ed3aca21e1a471990
The reissued record spans the 31 August 2026 release and 7 September test evidence. It preserves the original hosted segment’s failure to reach 72 checkpoints: only 68 were recorded, although recovery of old state and bounded memory passed. A later run using the corrected driver reached 72 checkpoints across its own six hours and restored the earlier checkpoint. Reissue did not turn the first result into a pass. This is evidence for two identified test segments, not a guarantee for every workload or current release.
31 August 2026
SHA-256 26957af2be3437d5e8c491ed8b7c58a327493cd58223865ca820a505b1db6530
On 7 September 2026, three assertions passed for the accepted-work workflow on a plain Linux customer-environment stand-in. A failed attempt restarted from a saved segment boundary, a later attempt completed, and the artifact was independently checked after the remote round trip. The environment’s limitations appeared in the result and missing limitation declarations were rejected. Plain-host recovery did not rewind memory. This was a tested integration configuration, not proof for every customer host or workload.
7 September 2026
SHA-256 050b9a39f9156a460ca60f61a18a3b4ce7a8f399e6572376bbaa2d86ea10e894
On 7 September 2026, a hosted-hardware test forked one task workspace into three branches. Two branches reached authorized model providers, performed work in their own workspaces and produced accepted candidates under independent verification. The unreachable branch was counted as failed rather than omitted. Exactly one verified result received scoped publication authorization. This was authorization for a review branch or pull request, not an actual external publication. Larger branch counts were not tested.
7 September 2026
SHA-256 fba361dcf92c7ecd98850bc5f7763f8fc0f476bbd04db012f3899f36ce12e503
On 7 September 2026, six recorded assertions passed for account-level erasure on hosted infrastructure. Deleted-account checkpoints became unreadable after key destruction, including from the tested backup copy, while another account’s matching content remained restorable. Identity was removed from retained billing records. The rollout used team/test data; a cleanup error also removed the test canary’s chunks, with no reported customer impact. A fresh post-cleanup checkpoint subsequently resumed. This record does not establish that every backup or future deployment follows the same key-management conditions.
7 September 2026
SHA-256 af71c71b8bce506a116b80e3c2e9bb7f0684c2eb916028560b1d4e6f3d05a24b
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.
6 September 2026
SHA-256 e40191cd5273dafc122272a0ab3fdc5270cb620ec2b34e61f24a8a0c0af9b60d
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.
6 September 2026
SHA-256 090e39655857d0cf8a768f32b8e1ddd7a8437e99dbce08e372450d701a88aa97
The record combines a hosted failover rehearsal dated 6 September 2026 and production enablement dated 7 September. Six assertions report billing continuity after real host deallocation: recorded spend survived, an identical retry was not billed again, and budget enforcement used the carried balance. Settlement stayed blocked until reconciliation passed. One run measured 47 seconds from deallocation to reconciliation; this is not an availability promise. The enablement addendum records a live production check. Session recovery relied on a separately tested mechanism, and regional disaster recovery was not established.
6 September 2026
SHA-256 475d5224a72c13083a0cf78861dca169389264720ee75a880fff60579275dbcd
On 6 September 2026, six deterministic assertions passed for the acceptance layer. Approved task conditions could not change silently; verification was separated from agent-controlled material; publication was limited to an authorized review branch or pull request; and unresolved attempts blocked settlement. Complete-job accounting matched an independent calculation, while unreadable inputs were shown as unavailable. These were isolated deterministic checks. A live multi-model task and real external publication were separate matters; this record does not establish either.
6 September 2026
SHA-256 e36d06f9ae95129c6ce3322c6c2199f45ffd2a8d73fc6836d9118dd6341e4cef
On 6 September 2026, two deterministic publication assertions passed. Evidence titles and summaries were derived from each record’s declared subject, conflicting manually supplied titles were rejected, and records missing required metadata stopped generation. This checks the consistency of the evidence catalogue, not the truth of every underlying product claim. It covers the record-generation side; whether the public website actually rendered the correct titles required a separate website check.
6 September 2026
SHA-256 ba94cabef6ed80263340d52aa057a60cd41cdacd6036c7a88709d6460defbd76
On 6 September 2026, three deterministic publication assertions passed. Claimed results had to reference specific assertions present in the cited records, rather than just naming a test or release. The check reconciled 152 assertions across 42 test records and rejected empty or invalid references. This establishes traceability within the inspected records, not whether each product result is correct or currently deployed. The public website’s rendering was a separate scope.
6 September 2026
SHA-256 fded36106e582db6b5ff91419a86d9a45a3d4fb11c65728a53fd886ca39a97c7
Retention policy v1.0 was approved on 9 September 2026 and published as final on 14 September. Checkpoints do not expire automatically while an account exists; destroying the account key makes them unreadable. Deleting a session erases the checkpoint content only it held: built and tested, served in production since 6 October 2026, and extended to sessions saved before then by a re-seal on 10 October 2026. A production check through the customer's own deletion route is still owed. No automatic expiry job is implemented, and several retention periods rely on procedure. Secrets written into a workspace are included in subsequent checkpoints.
9 September 2026 policy; 13 October 2026 implementation update.
SHA-256 83e7d769c10c3843c945985648b26e040a7c81402798a2792fd82c885391f439
On 5 September 2026, eight hosted assertions passed across two supported provider paths. Test sessions preserved in-memory state through save, interruption and resume; real provider calls were metered; identical retries avoided a second charge; missing credential access failed closed. The tests used a base guest exercising the integration configuration, rather than proving every behavior of a full agent application. Whether a request is retried remains the client’s decision. These results apply to the tested combinations.
5 September 2026
SHA-256 7c52d671b9253971703dc522d96c6d5d17f068b3bda023b827cdda11394ca7e3
This undated support matrix lists two tested integration families: Codex and a compatible-provider path with aider as a reference client. For each it records session recovery, metered provider use and replay of an identical interrupted request without a second charge. These are separate support properties, not a blanket promise that every agent workflow works. The underlying recovery tests used a base guest exercising the integration configuration. Clients decide whether to issue a retry, and the replay claim applies only to an identical request. Current supported versions require separate confirmation.
unavailable: this record declares no tag to date it by
SHA-256 e2d0d5f6ebb5d2a58f41b7f53f82ad5adf522127a87bb956abc36aba256458a5
On 4 September 2026, six hosted assertions passed for an additional compatible provider path. Deterministic tests checked credential routing, usage accounting, ordered records and replay of an identical request without another charge. A separate live provider call also succeeded and its identical retry was replayed. The live gateway ran in isolation and did not prove the full agent-plus-session-interruption combination. The replay guarantee concerns identical retries; it does not mean all similar requests are free.
4 September 2026
SHA-256 94adbd33db6726cf79f85056384b4a48c60d12c4b512ebc466fd1e9e8bc86337
On 4 September 2026, three hosted test assertions passed for changing a session’s model. A change was refused while the session was running, accepted while it was checkpointed, and used by the next request without resetting the request lineage. These tests used a simulated compatible provider. They covered compatible clients through the tested routing path; native clients for the other protocol were explicitly outside this result. Live continuation quality was not measured here.
4 September 2026
SHA-256 31ae8cf53629859bf68998042a39e8344432ee7148b4aebb434ae28660168e16
The record was written on 31 July 2026 after a release test could not be rerun under its existing baseline and compatibility rules. A resolution added on 30 August 2026 records amended rules and a passing normal and negative-control rerun in v0.1.4. The earlier blockage remains part of the test history. This does not establish universal agent support: aider retried interrupted work in three of four recorded runs.
31 July 2026; 30 August 2026
SHA-256 94fb10f2a0be4d06be464f8ded01ac35faf73a47a84079515e81d040743535d0
The collected ledger contains 184 redaction entries dated from 17 to 25 September 2026. It records that public evidence bytes were replaced, with separate original and derivative hashes and intended replacement or withdrawal notices. Redaction is a publication change, not a new successful product test. This summary records the publication change without republishing withdrawn source content. Ledger entries alone do not establish that every historical address, search copy or cache now serves safe content; those surfaces require separate verification.
25 September 2026
SHA-256 69a40778c2e7a0230176469138ff973decef69ce26dd778c8884c996e96c3c66
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.
30 September 2026; final normal run at 09:16:29 UTC
SHA-256 ca4f75297374c6cfa5b394141004297596d32981170ba3af28a3f5904f338724
The per-session deletion record dated 4 October 2026 reports fourteen passing test assertions on the laboratory platform, with none left unproven. After one session was deleted, the content only that session held could not be read from any of its checkpoints or backup copies, while the account's other sessions stayed readable and could still resume. Two deliberately broken controls, one that skipped the deletion and one that deleted too much, each failed as they should. Sessions saved in the older format must be re-sealed before they can be deleted one at a time. Production deployment followed on 6 October 2026 and is not part of this record.
4 October 2026
SHA-256 d7d074174bef18483c93a1c266b86fae71fb6f1af0db60456e670f42f7e4e5a9
The encrypted-checkpoint soak record dated 5 October 2026 reports ten passing test assertions over 180 minutes on the laboratory platform. Two encrypted sessions were checkpointed and resumed 306 times, every file check matched, and all 28,211 stored chunks were sealed. A deliberately broken control with encryption turned off left its chunks readable and was caught, as it should be. The soak covers one test scenario, not every workload. Production deployment followed on 6 October 2026 and is not part of this record.
5 October 2026
SHA-256 e7bb3e8bf884804f1821c7317a2f30eea182123f46b483d833ff29237eb2b153
The re-seal record dated 10 October 2026 reports seven passing test assertions on the laboratory platform, with none left unproven. After an account was re-sealed under a new key, a session saved in the older format and one saved in the current format both resumed byte for byte, content sealed under the old key could no longer be read from the live store or a backup copy, and the formerly older-format session could then be deleted on its own. A re-seal interrupted partway left the session readable and completed when run again, a session under a preservation hold was refused untouched, and two deliberately broken controls each failed as they should.
10 October 2026
SHA-256 0a013c609e89891d6bd06e3ebf1854908c619dc44199d2471299d523506a8b34