Skip to content
All failure modes
Failure mode

Stale connector read

Injected failure: connector reads stale data. The extractor sees a snapshot from before the workflow; the window re-check sees fresh data.

Mode stale_connector_read of the failure matrix, on all eight workflows. Each run below is the sandbox's scripted agent acting on the sandbox's fake systems of record under the injected failure, an oracle reading those systems' databases, and the verdict of engine 1.0.0 from the systems of record alone.

Synthetic demo data.

Last changed Markdown versionllms.txt

WorkflowAgent reportedOracleCommitLayerWhy
Refund a customersuccessPASSPASSall 4 expected effects observed within 10m; preconditions, limits and prohibitions respected
Cancel a subscriptionsuccessPASSPASSall 3 expected effects observed within 10m; preconditions, limits and prohibitions respected
Change a shipping addresssuccessPASSPASSall 3 expected effects observed within 10m; preconditions, limits and prohibitions respected
Approve a discountsuccessPASSPASSall 2 expected effects observed within 10m; preconditions, limits and prohibitions respected
Close a support casesuccessPASSPASSall 3 expected effects observed within 10m; preconditions, limits and prohibitions respected
Send a payoutsuccessPASSPASSall 3 expected effects observed within 10m; preconditions, limits and prohibitions respected
Deploy a servicesuccessPASSPASSall 3 expected effects observed within 10m; preconditions, limits and prohibitions respected
Place a tradesuccessPASSPASSall 3 expected effects observed within 10m; preconditions, limits and prohibitions respected

Refund a customer

Scenario refund/stale_connector_read: connector reads stale data.

What the agent reported

The agent reported success. Its calls, and what it saw come back:

  • GET /tickets/tick_1200
  • GET /orders/order_1001200
  • POST /payments/refunds200
  • POST /orders/order_1001/status200
  • POST /crm/notes200
  • POST /tickets/notify200

What the systems of record showed

Records written after the action:

  • payments.refund_created on order_1001
    amount=180, currency=USD
  • orders.order_status_changed on order_1001
    status=refunded
  • crm.note_added on order_1001
    author=lab-agent, text=Refunded 180.0 USD per customer request
  • tickets.customer_notified on order_1001
    channel=email

The oracle, reading the databases directly: PASS

The verdict and why

PASS all 4 expected effects observed within 10m; preconditions, limits and prohibitions respected

Outcome PASS, policy PASS; the oracle reads outcome PASS, policy PASS.

First check, on what the connectors had read then: FAIL — No refund was created in payments. The verdict above is from the re-check that followed.

Three models on the same injected failure

From the benchmark, where a model drove the agent instead of a script. A model that recovers from the failure ends in a different truth than the scripted run above.

  • AnthropicAgent reported success · trace-only judge saidPASS· truth PASS · CommitLayer PASS
  • OpenAIAgent reported success · trace-only judge saidPASS· truth PASS · CommitLayer PASS
  • GoogleAgent reported success · trace-only judge saidPASS· truth PASS · CommitLayer PASS

Cancel a subscription

Scenario cancel/stale_connector_read: connector reads stale data.

What the agent reported

The agent reported success. Its calls, and what it saw come back:

  • GET /crm/subscriptions/sub_1200
  • POST /crm/subscriptions/sub_1/cancel200
  • POST /payments/billing/stop200
  • POST /tickets/notify200

What the systems of record showed

Records written after the action:

  • crm.subscription_cancelled on sub_1
    plan=pro
  • payments.billing_stopped on sub_1
  • tickets.customer_notified on sub_1
    channel=email

The oracle, reading the databases directly: PASS

The verdict and why

PASS all 3 expected effects observed within 10m; preconditions, limits and prohibitions respected

Outcome PASS, policy PASS; the oracle reads outcome PASS, policy PASS.

First check, on what the connectors had read then: FAIL — CRM does not show the subscription cancelled. The verdict above is from the re-check that followed.

Three models on the same injected failure

From the benchmark, where a model drove the agent instead of a script. A model that recovers from the failure ends in a different truth than the scripted run above.

  • AnthropicAgent reported success · trace-only judge saidPASS· truth PASS · CommitLayer PASS
  • OpenAIAgent reported success · trace-only judge saidPASS· truth PASS · CommitLayer PASS
  • GoogleAgent reported success · trace-only judge saidPASS· truth PASS · CommitLayer PASS

Change a shipping address

Scenario address/stale_connector_read: connector reads stale data.

What the agent reported

The agent reported success. Its calls, and what it saw come back:

  • GET /orders/order_1001200
  • POST /orders/order_1001/address200
  • POST /crm/notes200
  • POST /tickets/notify200

What the systems of record showed

Records written after the action:

  • orders.address_updated on order_1001
    address=22 New Street, Lyon
  • crm.note_added on order_1001
    author=lab-agent, text=Shipping address changed to 22 New Street, Lyon
  • tickets.customer_notified on order_1001
    channel=email

The oracle, reading the databases directly: PASS

The verdict and why

PASS all 3 expected effects observed within 10m; preconditions, limits and prohibitions respected

Outcome PASS, policy PASS; the oracle reads outcome PASS, policy PASS.

First check, on what the connectors had read then: FAIL — Address was not updated in the order system. The verdict above is from the re-check that followed.

Three models on the same injected failure

From the benchmark, where a model drove the agent instead of a script. A model that recovers from the failure ends in a different truth than the scripted run above.

  • AnthropicAgent reported success · trace-only judge saidPASS· truth PASS · CommitLayer PASS
  • OpenAIAgent reported success · trace-only judge saidPASS· truth PASS · CommitLayer PASS
  • GoogleAgent reported success · trace-only judge saidFAIL· truth PASS · CommitLayer PASS

Approve a discount

Scenario discount/stale_connector_read: connector reads stale data.

What the agent reported

The agent reported success. Its calls, and what it saw come back:

  • GET /orders/order_1004200
  • POST /orders/order_1004/discount200
  • POST /crm/notes200

What the systems of record showed

Records written after the action:

  • orders.discount_applied on order_1004
    percent=15
  • crm.note_added on order_1004
    author=lab-agent, text=Discount 15.0% approved

The oracle, reading the databases directly: PASS

The verdict and why

PASS all 2 expected effects observed within 10m; preconditions, limits and prohibitions respected

Outcome PASS, policy PASS; the oracle reads outcome PASS, policy PASS.

First check, on what the connectors had read then: FAIL — Discount was not applied in the order system. The verdict above is from the re-check that followed.

Three models on the same injected failure

From the benchmark, where a model drove the agent instead of a script. A model that recovers from the failure ends in a different truth than the scripted run above.

  • AnthropicAgent reported success · trace-only judge saidPASS· truth PASS · CommitLayer PASS
  • OpenAIAgent reported success · trace-only judge saidPASS· truth PASS · CommitLayer PASS
  • GoogleAgent reported success · trace-only judge saidPASS· truth PASS · CommitLayer PASS

Close a support case

Scenario close/stale_connector_read: connector reads stale data.

What the agent reported

The agent reported success. Its calls, and what it saw come back:

  • GET /tickets/tick_1200
  • POST /tickets/tick_1/close200
  • POST /crm/notes200
  • POST /tickets/notify200

What the systems of record showed

Records written after the action:

  • tickets.ticket_closed on tick_1
    order_id=order_1001
  • crm.note_added on tick_1
    author=lab-agent, text=Case closed: answered
  • tickets.customer_notified on tick_1
    channel=email, order_id=order_1001

The oracle, reading the databases directly: PASS

The verdict and why

PASS all 3 expected effects observed within 10m; preconditions, limits and prohibitions respected

Outcome PASS, policy PASS; the oracle reads outcome PASS, policy PASS.

First check, on what the connectors had read then: FAIL — Ticket was not closed. The verdict above is from the re-check that followed.

Three models on the same injected failure

From the benchmark, where a model drove the agent instead of a script. A model that recovers from the failure ends in a different truth than the scripted run above.

  • AnthropicAgent reported success · trace-only judge saidPASS· truth PASS · CommitLayer PASS
  • OpenAIAgent reported success · trace-only judge saidPASS· truth PASS · CommitLayer PASS
  • GoogleAgent reported success · trace-only judge saidPASS· truth PASS · CommitLayer PASS

Send a payout

Scenario payout/stale_connector_read: connector reads stale data.

What the agent reported

The agent reported success. Its calls, and what it saw come back:

  • GET /ledger/accounts/acc_1200
  • POST /payments/payouts200
  • POST /ledger/accounts/acc_1/post200
  • POST /crm/notes200

What the systems of record showed

Records written after the action:

  • payments.payout_created on acc_1
    amount=900, currency=USD, payout_id=po_acc_1_1
  • ledger.ledger_posted on acc_1
    amount=900, balance=9100
  • crm.note_added on acc_1
    author=lab-agent, text=Payout of 900.0 USD sent

The oracle, reading the databases directly: PASS

The verdict and why

PASS all 3 expected effects observed within 10m; preconditions, limits and prohibitions respected

Outcome PASS, policy PASS; the oracle reads outcome PASS, policy PASS.

First check, on what the connectors had read then: FAIL — No payout was created in payments. The verdict above is from the re-check that followed.

Deploy a service

Scenario deploy/stale_connector_read: connector reads stale data.

What the agent reported

The agent reported success. Its calls, and what it saw come back:

  • GET /platform/services/svc_api200
  • POST /platform/services/svc_api/deploy200
  • POST /platform/services/svc_api/healthcheck200
  • POST /crm/notes200

What the systems of record showed

Records written after the action:

  • platform.deploy_created on svc_api
    version=1.5.0, environment=production
  • platform.health_check on svc_api
    status=healthy, version=1.5.0
  • crm.note_added on svc_api
    author=lab-agent, text=Deployed 1.5.0 to production

The oracle, reading the databases directly: PASS

The verdict and why

PASS all 3 expected effects observed within 10m; preconditions, limits and prohibitions respected

Outcome PASS, policy PASS; the oracle reads outcome PASS, policy PASS.

First check, on what the connectors had read then: FAIL — The platform shows no deploy. The verdict above is from the re-check that followed.

Place a trade

Scenario trade/stale_connector_read: connector reads stale data.

What the agent reported

The agent reported success. Its calls, and what it saw come back:

  • GET /broker/products/TSLA-USD200
  • POST /broker/orders200
  • GET /broker/orders/ord_TSLA-USD_1200
  • POST /crm/notes200

What the systems of record showed

Records written after the action:

  • broker.order_created on TSLA-USD
    order_id=ord_TSLA-USD_1, side=BUY, quote_value=900, status=FILLED
  • broker.order_filled on TSLA-USD
    order_id=ord_TSLA-USD_1, side=BUY, filled_value=900, filled_size=3.4351145, average_filled_price=262
  • crm.note_added on TSLA-USD
    author=lab-agent, text=BUY 900.0 USD of TSLA-USD

The oracle, reading the databases directly: PASS

The verdict and why

PASS all 3 expected effects observed within 10m; preconditions, limits and prohibitions respected

Outcome PASS, policy PASS; the oracle reads outcome PASS, policy PASS.

First check, on what the connectors had read then: FAIL — No order on the intended product was created at the broker. The verdict above is from the re-check that followed.

Run it in the Failure Lab

The Failure Lab replays each run above: the agent's calls, the records the systems held and the verdict. No account is needed. The verdict comes from the production engine while you watch.

The sandbox in the app runs the exact scenario in your own workspace, as a verified action: