Review topic:
Review a workflowCurrent public evidence · Reviewed October 16, 2026
Payment control for AI agents. Backed by scoped evidence.
Inspect the evidence behind one shared payment approval, from Agent handoff and unresolved payments to checks at execution time.
Verified
One controlled test connects signed source IDs, lost payment response, A-to-B handoff and actual process restart. The original UNKNOWN reservation remains. The native Increase Sandbox record has a separate scope.
Inspect the connected controlled testStill to qualify
Your source, deployed credentials, release paths and recovery ownership require customer acceptance. Real-money production remains NO_GO. No availability percentage is established.
Find your remaining conditionsAvailable now
Free Workspace Sandbox for one supported workflow and its evidence. The synthetic SDK demo and installed runtime checks each retain their own scope. Production requires a qualified scope and operating approval.
See pricing and qualificationThe current Agent records below are separate controlled profiles. The earlier local Payment API reference and RC15's provider candidates concern own-account USD cash movement, not supplier payment. Supplier-payment and check-to-ACH illustrations describe a broader principle; those workflows are outside the earlier local reference qualification.
Current controlled profiles, native Sandbox, SDK and earlier engineering records retain separate sources, amounts and counts. A successful verification is historical evidence, not a payment permit.
Agent payment evidence · Reviewed October 16, 2026
One connected controlled test follows the original through lost response, Agent handoff and process restart.
Retained original
USD 1.00
The test amount stays with the original after the lost reply. Assigning B does not create another payment approval.
Replacement: HOLDOriginal outcome: UNKNOWN
Inspect this test's recordTwo explicitly signed source IDs resolve to one original USD 1.00 approval.
One controlled payment request reaches the provider model. Its response is intentionally lost.
The original USD 1.00 remains UNKNOWN and reserved. B receives no new payment capacity.
A's stale request is refused. B's alternative payment route stays on HOLD.
An actual forced termination and a new process retain the original, reservation and history.
Controlled qualification. Test signing keys, fictional provider replies and a test-only browser adapter. This USD 1.00 test, the Workspace learning example and the USD 10 Increase reference retain separate scopes. Customer acceptance and native execution remain unqualified.
Inspect the claim and its limits
Four separate controlled qualification profiles. Integration in one source tree does not establish joint native qualification.
Agent handoff Controlled
B takes over. The original stays reserved.
Handing over the work does not grant another right to pay.
02ee8208324e8f85ca8fc953e35b9c610ff1c397joint-execution-02ee8208324e8f85ca8fc953e35b9c610ff1c397-38812917019-1ad74367807df7676814f79a53d6e101dde5aac0a5cd33bc468619b66af45c031Source-bound release records report separate retained-evidence replay. This is not an external audit, independent host attestation or customer acceptance.
Read the public evidence summaryOriginal records and reproduction materials are shared within an agreed technical review.
Request supporting recordsCurrent installation Controlled
Changed or expired evidence holds new effects.
A previously passing installation check does not become permanent permission.
411967472bfbbc0b7ad510b8f4df5d5c5260b118qa-pipeline-38834881004-workspace-1bbc3a2001643df68fed516e4d6ceeed0c6b41856addfd811be9cb4bf59724a69Source-bound release records report separate retained-evidence replay. This is not an external audit, independent host attestation or customer acceptance.
Read the public evidence summaryOriginal records and reproduction materials are shared within an agreed technical review.
Request supporting recordsVersion continuity Controlled
Compaction and adapter changes retain the original accounting.
Maintenance must preserve unresolved reservations and completed-payment consumption.
fbd5df6c9e64efd725bc0f4f4fa2b314b4da7ff3authority-evolution-fbd5df6c9e64efd725bc0f4f4fa2b314b4da7ff3-38827543234-12357c5deafdb22c7e5fe87513cc61e64b7ae68ba1ea1bc4e5bcfdebe35b94c8eSource-bound release records report separate retained-evidence replay. This is not an external audit, independent host attestation or customer acceptance.
Read the public evidence summaryOriginal records and reproduction materials are shared within an agreed technical review.
Request supporting recordsCompany ownership Controlled
One registered obligation stays bound to its first original.
A changed actor, route or restarted installation cannot register a second original for that obligation within this profile.
8b9192056bd9493d3c6929c1335779e2e3fea551original-owner-registry-8b9192056bd9493d3c6929c1335779e2e3fea551-38842313357-runtime-1ec0b42ce2241548c0e02249eb3867aa305ecce5bdd26a45af33db3be4385f99aSource-bound release records report separate retained-evidence replay. This is not an external audit, independent host attestation or customer acceptance.
Read the public evidence summaryOriginal records and reproduction materials are shared within an agreed technical review.
Request supporting recordsOpen at this review. A normal request expected to return ACCEPT instead returned HOLD/JOURNAL_REPAIR_REQUIRED after one controlled provider CREATE.
The same-source rerun and retained replay passed. The original failure is retained and its cause is not established.
A passing rerun does not resolve the first failure or establish an availability percentage.
Scope: controlled loopback HTTP. Source 2af16ae9338a5ea5f859180397c5147265ff63bf.
Your company's next proof
The free Workspace helps you start with one workflow. Your source, installed paths and operating conditions require their own acceptance.
Confirm the exact organization, obligation, version, amount and separate financial approval. Source access and Agent receiving proof do not approve a payment.
Review the installed code, funding account, credentials, final gates and known bypass paths. Customer custody and native enforcement must match the agreed scope.
Verify authoritative original outcomes, unresolved reservations, restart behavior and fresh replacement approval. Reconcile the retained records with the provider.
Agree recovery ownership, independent trust, customer acceptance and operational approval. Availability and production disaster recovery need measured evidence.
Current operating boundary: customer acceptance NOT_PERFORMED; native execution NOT_QUALIFIED for these four profiles; real-money Production NO_GO. A historical record or successful verification grants no current execution authority.
We are building toward company-owned payment authority that can persist across Agents and systems. Cross-platform network enforcement remains a direction requiring further qualification.
Explore the prepared synthetic demo · No real payment or installation qualification.
Earlier evidence · Original dates and scopes preserved
The October 9 Increase Sandbox reference, October 11 control profiles, SDK, RC15 and assessment records retain their own dates, amounts and counts. Counts are not combined.
Installed control and authority evidence · October 11, 2026
Later installed releases preserve the original approval and complete authority-event payloads, with source-bound incident controls and independent Python and Node verification.
The tested Increase Sandbox reference retains UNKNOWN reservations and exact original identities, commits incident controls to its witnessed history, and checks interruption repair and same-version reinstall. Independent implementations recalculate the retained economic prefixes against the source, provider-original inventory and witness evidence required by the profile.
Scope: one fictional own-account USD 10 workflow, one Linux execution host and two paths using Increase. Sandbox settlement and return are simulations; no customer funds are activated.
A separately qualified control profile checks local database and whole-execution snapshot restore, retired credentials, signed responses and exact-original recovery while paused. Its provider-effect campaign uses a controlled synthetic model with real loopback mTLS.
This separate campaign does not qualify a combined native and remote-control deployment, payment-key theft or a consistent rollback of the independent control host.
These later records remain separate from the October 9 installed reference, SDK, RC15 and assessment populations. Actual customer source and path coverage, Owner-to-host-to-public-Workspace acceptance, deployed credential custody and failure-domain separation, measured availability and operational approval remain open. Real-money production is not approved.
Controlled installed profile · October 16, 2026
A separate, finite installed BILL record preserves one financial owner while a reviewed executor succeeds another.
The controlled profile keeps the original approval, unresolved UNKNOWN reservations and PAID consumption while its reviewed executor changes from v1 to v2. The same retained payment history and separate retained witness stay with the original owner.
Both adapters are fixed and reviewed. Provider effects and signing custody are controlled test fixtures. Arbitrary upgrades, legacy-profile migration, financial key rotation and host-loss recovery require their own qualification.
Native BILL execution remains NOT_QUALIFIED; external customer acceptance remains NOT_PERFORMED; real-money Production remains NO_GO. The separate normal-progress journal-repair finding remains open. These records retain their own sources and populations; their counts are not combined.
Installed reference workflow · October 9, 2026
Scheduled own-account funding and an operator replacement path share one obligation in the installed runtime.
27 passed checks against actual Increase sandbox APIs. HALTSEAL controlled two test-company configurations. Customer installations: zero.
One Linux host. USD 10 between accounts belonging to one fictional Increase entity. Both software paths use Increase. HALTSEAL controlled the test companies and issued the payment approvals.
The internal qualification record applies to the tested release and environment. Customer workflows, independent review, cross-provider enforcement, failover between hosts and production disaster recovery require separate qualification.
Customer installations remain zero. Real-money production is not approved. No availability SLO or reliability percentage is established.
A manual bank portal or another credential outside the connected paths remains outside this control. These results are separate from the SDK, RC15 and offline assessment populations below; their historical counts remain unchanged.
SDK v2 · 0.2.0-rc.2 · October 7, 2026
Python and JavaScript reject mismatched operation or obligation IDs, incomplete attempt records and inconsistent nested histories. Both apply a total caller deadline across the request and response.
Each language passes 29 synthetic HTTP cases, 101 identity/record cases and seven zero-send input checks. CI covers Python 3.12/3.13 and Node 22/24. Clean builds reproduce the published wheel and tarball bytes.
The client checks use fixed HTTP fixtures. The sandbox v2 record separately reports retained-engine checks with RC4 clients and an Increase sandbox contract probe. The October 9 installed Increase reference has its own bounded sandbox results. Customer workflows, cross-provider recovery and production availability remain unqualified.
Preserved Payment API record · 0.1.0rc4
Public clients, an exact contract and recorded original-recovery behavior. The local evaluator is available by request.
RC4 records 144 Python tests and 28 separate JavaScript transport tests, six real SIGKILL boundaries and two local installations using identical SDK archives.
These are synthetic USD 10 own-account fixtures on one physical host. Native provider calls, customer installations and independent operators remain 0. Production remains NO_GO; no availability SLO is established.
A lost create response retains the operation ID. A lost resume, cancel or recover response retains the existing attempt ID. UNKNOWN exposure stays reserved; no financial call retries automatically.
This API release, the RC15 reference and the assessment kit are separate records. Their test counts are not combined. Customer installation and provider qualification require their own evidence.
Engineering evidence · RC15 R1
Four local scenarios. Inspect the state, the proposed action and the observed result. Each view connects to its published source.
Scenario views based on the public summary · not original event traces
The original effect remains possible.
A different path. The same payment capacity.
Replacement held while the original effect remained possible.
RC15 R1 · observed case 1 · public summary
{
"situation": "Original response lost",
"result": "Replacement held while the original effect remained possible"
}A new process owns execution.
Its execution authority is stale.
No additional release granted to the stale executor.
RC15 R1 · observed case 2 · public summary
{
"situation": "Previous executor returns after ownership handoff",
"result": "No additional release granted to the stale executor"
}Creation or funding outcome is unknown.
This is not a third provider.
Next protected path held while the replacement effect remained possible.
RC15 R1 · observed case 3 · public summary
{
"situation": "Replacement creation or funding response lost",
"result": "Next protected path held while the replacement effect remained possible"
}The prior attempt can no longer pay.
Fresh current checks are required.
One fresh successor release allowed in the declared local case.
RC15 R1 · observed case 4 · public summary
{
"situation": "Authoritative no-effect closure and fresh current checks",
"result": "One fresh successor release allowed in the declared local case"
}Authoritative no-effect closure restored capacity in the declared local case. A fresh request with current authority permitted one successor release. Closure or an old ACCEPT alone did not authorize it.
The supplied RC15 release and October 5 local recheck agree. These populations are separate; they do not measure a reliability percentage.
In this RC15 R1 record, native acceptance remains blocked; the integrated native workflow has not run. Provider finality, every in-scope customer release path and an accepted recovery plan must be demonstrated before customer approval. Real-money production is not approved.
Completed-payment response: in the local full-amount PAID fixture, available capacity remained zero and a fresh replacement returned HOLD, with no replacement transfer. The public control model and demos do not define a released execution API contract.
Known exclusion: joint-history rollback is not protected. Additional exposure was reproduced outside the protected scope. Inspect the recovery limits.
Selected restarts and recovery cases were checked locally. Host migration is not implemented; separate-host failover and production disaster recovery are not validated.
A known exclusion was reproduced: Joint-history rollback is not protected. Additional exposure was reproduced outside the protected scope. There is no general crash-recovery or durability guarantee.
Independent bank settlement is unproven. The candidates are Stripe and a Wise sandbox path; the next protected path is not a third provider. Byte hashes and semantic replay are not independent attestation.
Current assessment release · v1.1.0
The offline workbench separates a workflow finding from qualification to enforce. A diagnosis, a reference match and submitted evidence do not grant payment authority.
Ten qualification gates assess execution readiness. The offline readout does not authorize:
Execution readiness remains NOT QUALIFIED · DO NOT ENFORCE.
October 2 local recheck: 158 passed; one optional Korean PDF font check skipped. The supplied release used the font and reports 159/159 passed, no skips. Both report replays and the package inventory matched. No PDF was generated in this recheck.
Native browser integration was blocked in the supplied v1.1.0 runtime. The earlier 25 bridged UI checks belong to v1.0.0. Website browser checks do not validate the kit or payment engine.
Offline Shadow example
Compare what was known at decision time with what happened later.
The earlier attempt remains unresolved.
Same supplier. Same obligation.
The fallback proceeded.
The earlier attempt may still pay.
Added after the decision. Not used to produce the HOLD.
Matched information and decision time are declared in this fixture. This is not a customer incident or proof of prevented loss.
The fixture declares matching context. No payment was blocked. A later outcome does not rewrite the earlier decision or establish prevented loss.
How to evaluate a Shadow recordA retained failure
Q04b · unsafe local fixture · synthetic
One observed payment does not establish a valid admission. Required closure evidence was missing when the fallback was allowed.
The original did not pay.
It had not reached the workflow.
Required evidence was absent.
The matched local case that correctly holds the fallback passes. Incomplete observation or attribution remains INCONCLUSIVE.
Q05b retains a bounded competing-replacement comparison; it does not cover every distributed race. The earlier unsafe campaign already failed overall. Later findings clarify admission separately from observed effects.
Evidence boundary
Recorded behavior within the declared Increase sandbox reference and separate local synthetic scenarios. Failed and inconclusive findings remain in their original records.
Customer coverage, production provider finality, exhaustive concurrency, production security and reliability, independent reproduction or economic value.
No production certification or guaranteed safety is implied. Partial payments, FX, refunds, batches and production disaster recovery require separate evaluation. Not reached, not evaluated, and inconclusive are not PASS.
The cause of a historical local storage failure remains unconfirmed. Later local checks do not guarantee future durability. Production storage and recovery require separate validation.
Production readiness remains unestablished. Buyer-specific coverage, outcome evidence, recovery and accountable owners must be validated before customer approval.
Review the public control modelTechnical review
Source records and reproduction materials are available by request.
Request the evidence recordsPlease keep your first message non-confidential.
Include the workflow, case or claim you want to examine. Sharing scope and terms are agreed before delivery.
Release history
Historical counts are not combined with RC15 R1 or v1.1.0.
Historical engineering release: 48 local unit tests, 12 process campaigns and 10 evidence mutations, with a separately reproduced joint-history rollback exclusion. The October 2 recheck agreed. These counts are not added to RC15 results.
RC14 R1 public summary · revision 2Historical release: 108 scenario combinations matched, 123 unit tests passed in the supplied log, and 25 UI checks used bridged local transport. The earlier local recheck recorded 122 passed and one optional PDF font check skipped. Native browser HTTP was not validated. These UI results are not v1.1.0 results.
HALTSEAL Qualification Kit v0.4.2: 286 unit and regression tests passed per complete run; 36 scenarios × 4 modes = 144 reference executions; 120 legacy and 72 repeated fault-injected executions, recorded separately. Reference findings: 71 PASS, 33 FAIL, 40 INCONCLUSIVE. Offline Shadow: 8 fictional proposals / 7 obligations; 2 ACCEPT, 3 HOLD, 2 REFUSE, 1 INCONCLUSIVE.
Actual buyer paths verified: 0. Native provider calls: 0. Repeat runs are not independent observations. These assessment results do not validate the execution-control engine.
| Material | Scope and record |
|---|---|
| Assessment and offline Shadow | v0.4.2 · v0.4.1 · v0.4.0 · v0.3.0. Public summaries; detailed records by request. |
| Execution-control reference | RC12 R2. Separate from assessment software; production qualification is not established. |
| Buyer review | RC12.1 R1. A review of the RC12 R2 reference, not an installable runtime or new engine validation. |
| Earlier engine record | RC8.0.1. Historical summary. Local, synthetic findings do not validate later engine releases. |
Public payment sources provide context, not evidence of HALTSEAL performance.
Research, sources and limitsSeparate diagnostic release
Run the free CLI/Python package against sixteen authored synthetic expectations. These diagnostic cases are separate from the engineering and assessment counts above.
Inspect and run the lab