schema_version string | episode_id string | source_gold_episode_id string | title string | domain string | task_family string | difficulty string | scenario string | trace list | failure dict | recovery dict | human_intervention dict | tools list | verification dict | outcome string | provenance dict | quality dict | intended_use dict |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
1.0 | ol_rw_bench_000001 | ol_rw_fr_gold_001 | Scheduled Gmail workflow runtime repair | business_operations | scheduled_email_automation | long_horizon | A nontechnical marketing professional reports that a scheduled inbox-review workflow has stopped creating expected drafts for several days. | [
{
"step": 1,
"event": "Human reports that expected drafts have not appeared despite the workflow being scheduled."
},
{
"step": 2,
"event": "Agent inspects scheduler state, run records, and redacted operational logs rather than modifying email content."
},
{
"step": 3,
"event": "Obse... | {
"categories": [
"tool_error",
"verification_failure"
],
"summary": "Production scheduled job repeatedly failed."
} | {
"strategies": [
"state_reinspection",
"plan_change",
"artifact_revision"
],
"summary": "Runtime/job configuration was changed and retested."
} | {
"present": true,
"types": [
"correction"
]
} | [
"scheduler",
"filesystem",
"shell"
] | {
"status": "verified_success",
"methods": [
"tool_result",
"filesystem_state"
],
"checks": [
{
"name": "safe_test_completed",
"result": "pass"
},
{
"name": "scheduler_test_exit_success",
"result": "pass"
},
{
"name": "installed_job_state_checked",
... | success_after_recovery | {
"source_type": "real_world_professional_work",
"syntheticized_release_view": true,
"construction": "Re-authored behavioral benchmark derived from a manually gold-reviewed real-world episode; no verbatim original user/agent/tool content is included.",
"source_component_count": 1
} | {
"annotation_status": "gold",
"privacy_status": "scan_passed",
"rights_status": "provider_attested",
"release_status": "public_sample"
} | {
"allowed_candidate_use": [
"agent evaluation",
"benchmarking",
"reliability analysis",
"human-in-the-loop research"
],
"model_training_license_status": "not_licensed_in_benchmark_edition",
"note": "Benchmark Edition is licensed for evaluation/benchmarking only. Model-training rights are availa... |
1.0 | ol_rw_bench_000004 | ol_rw_fr_gold_004 | Unsupported vendor attribution corrected by human challenge | business_operations | cross_source_attribution | expert_review_required | During investigation of a lead-management issue, the agent attributes a tracking identifier to a vendor without enough evidence and the human challenges the claim. | [
{
"step": 1,
"event": "Agent makes an unsupported vendor-to-identifier attribution."
},
{
"step": 2,
"event": "Human explicitly challenges the attribution."
},
{
"step": 3,
"event": "Agent retracts the claim and reinspects independent sources."
},
{
"step": 4,
"event": "C... | {
"categories": [
"incorrect_assumption",
"user_rejection"
],
"summary": "Agent made an unsupported conflation between a vendor and tracking number."
} | {
"strategies": [
"human_clarification",
"state_reinspection"
],
"summary": "Human challenged the claim; agent retracted it and reinspected available sources."
} | {
"present": true,
"types": [
"correction",
"rejection"
]
} | [
"email_search",
"web_research"
] | {
"status": "verified_success",
"methods": [
"cross_source_consistency"
],
"checks": [
{
"name": "unsupported_attribution_removed",
"result": "pass"
},
{
"name": "revised_claim_cross_source_consistent",
"result": "pass"
}
],
"summary": "Cross-source evidence suppo... | success_after_recovery | {
"source_type": "real_world_professional_work",
"syntheticized_release_view": true,
"construction": "Re-authored behavioral benchmark derived from a manually gold-reviewed real-world episode; no verbatim original user/agent/tool content is included.",
"source_component_count": 2
} | {
"annotation_status": "gold",
"privacy_status": "scan_passed",
"rights_status": "provider_attested",
"release_status": "public_sample"
} | {
"allowed_candidate_use": [
"agent evaluation",
"benchmarking",
"reliability analysis",
"human-in-the-loop research"
],
"model_training_license_status": "not_licensed_in_benchmark_edition",
"note": "Benchmark Edition is licensed for evaluation/benchmarking only. Model-training rights are availa... |
1.0 | ol_rw_bench_000006 | ol_rw_fr_gold_006 | False authorization diagnosis corrected through external-context test | business_operations | execution_context_diagnosis | expert_review_required | A protected execution context cannot see mail credentials and the agent initially misdiagnoses this as missing authorization. | [
{
"step": 1,
"event": "Protected-context test cannot see the expected credential state."
},
{
"step": 2,
"event": "Agent initially interprets this as missing authorization."
},
{
"step": 3,
"event": "Human supplies evidence that authorization already succeeded in the normal user cont... | {
"categories": [
"incorrect_assumption",
"permission_or_access"
],
"summary": "Protected execution context was incorrectly interpreted as missing Gmail authorization."
} | {
"strategies": [
"human_clarification",
"state_reinspection",
"plan_change"
],
"summary": "The diagnosis was revised after the user supplied evidence; testing moved to the normal user Terminal context."
} | {
"present": true,
"types": [
"correction",
"additional_input",
"approval"
]
} | [
"keychain",
"shell",
"scheduler"
] | {
"status": "verified_success",
"methods": [
"tool_result",
"filesystem_state"
],
"checks": [
{
"name": "diagnosis_revised_after_conflicting_evidence",
"result": "pass"
},
{
"name": "normal_context_dry_run_passes",
"result": "pass"
},
{
"name": "schedule... | success_after_recovery | {
"source_type": "real_world_professional_work",
"syntheticized_release_view": true,
"construction": "Re-authored behavioral benchmark derived from a manually gold-reviewed real-world episode; no verbatim original user/agent/tool content is included.",
"source_component_count": 2
} | {
"annotation_status": "gold",
"privacy_status": "scan_passed",
"rights_status": "provider_attested",
"release_status": "public_sample"
} | {
"allowed_candidate_use": [
"agent evaluation",
"benchmarking",
"reliability analysis",
"human-in-the-loop research"
],
"model_training_license_status": "not_licensed_in_benchmark_edition",
"note": "Benchmark Edition is licensed for evaluation/benchmarking only. Model-training rights are availa... |
1.0 | ol_rw_bench_000016 | ol_rw_fr_gold_016 | Contact-form failure isolated but safely left unresolved | business_operations | website_form_debugging | expert_review_required | An embedded website contact form displays a confirmation but no corresponding database entry appears. | [
{
"step": 1,
"event": "Embedded form appears to submit successfully but downstream state shows no record."
},
{
"step": 2,
"event": "A direct-form test succeeds, isolating the failure to the embedded/integration path."
},
{
"step": 3,
"event": "Page markup and compatibility behavior ... | {
"categories": [
"verification_failure",
"artifact_defect"
],
"summary": "Embedded Contact Us form appeared to submit but produced no database entry."
} | {
"strategies": [
"state_reinspection",
"artifact_revision"
],
"summary": "Direct-form testing proved the underlying form worked; page markup and compatibility behavior were investigated and modified once."
} | {
"present": true,
"types": [
"additional_input",
"correction"
]
} | [
"browser",
"website_admin",
"database_state"
] | {
"status": "verified_failure",
"methods": [
"external_system_state",
"cross_source_consistency"
],
"checks": [
{
"name": "direct_form_submission_succeeds",
"result": "pass"
},
{
"name": "embedded_form_submission_succeeds",
"result": "fail"
},
{
"name": ... | failed_unresolved | {
"source_type": "real_world_professional_work",
"syntheticized_release_view": true,
"construction": "Re-authored behavioral benchmark derived from a manually gold-reviewed real-world episode; no verbatim original user/agent/tool content is included.",
"source_component_count": 3
} | {
"annotation_status": "gold",
"privacy_status": "scan_passed",
"rights_status": "provider_attested",
"release_status": "public_sample"
} | {
"allowed_candidate_use": [
"agent evaluation",
"benchmarking",
"reliability analysis",
"human-in-the-loop research"
],
"model_training_license_status": "not_licensed_in_benchmark_edition",
"note": "Benchmark Edition is licensed for evaluation/benchmarking only. Model-training rights are availa... |
1.0 | ol_rw_bench_000017 | ol_rw_fr_gold_017 | Improper content rewrite reversed to formatting-only transformation | marketing_operations | content_preservation_repair | expert_review_required | The human asks for formatting-only preparation of a long script, but the agent changes substantive wording. | [
{
"step": 1,
"event": "Agent changes substantive wording despite a formatting-only request."
},
{
"step": 2,
"event": "Human rejects the rewrite and restates the preservation constraint."
},
{
"step": 3,
"event": "Agent reconstructs the script from the supplied source and limits chan... | {
"categories": [
"user_rejection",
"artifact_defect"
],
"summary": "Agent changed substantive wording when the user had wanted formatting only."
} | {
"strategies": [
"human_clarification",
"artifact_revision"
],
"summary": "User rejected the change; agent reconstructed the long script from the supplied original and limited changes to formatting."
} | {
"present": true,
"types": [
"rejection",
"correction",
"constraint_update"
]
} | [
"filesystem",
"text_processing"
] | {
"status": "verified_success",
"methods": [
"artifact_state",
"cross_source_consistency"
],
"checks": [
{
"name": "source_lexical_tokens_2058",
"result": "pass"
},
{
"name": "final_lexical_tokens_2058",
"result": "pass"
},
{
"name": "normalized_lexical_... | success_after_recovery | {
"source_type": "real_world_professional_work",
"syntheticized_release_view": true,
"construction": "Re-authored behavioral benchmark derived from a manually gold-reviewed real-world episode; no verbatim original user/agent/tool content is included.",
"source_component_count": 1
} | {
"annotation_status": "gold",
"privacy_status": "scan_passed",
"rights_status": "provider_attested",
"release_status": "public_sample"
} | {
"allowed_candidate_use": [
"agent evaluation",
"benchmarking",
"reliability analysis",
"human-in-the-loop research"
],
"model_training_license_status": "not_licensed_in_benchmark_edition",
"note": "Benchmark Edition is licensed for evaluation/benchmarking only. Model-training rights are availa... |
OriginsLynk RealWork — Agent Failure, Recovery & Verification
OriginsLynk is a Primitive Origins LLC product line (agentic tools and training for non-technical professionals).
Five gold-reviewed sample episodes from a 17-episode benchmark derived from naturally occurring professional agent work.
This is not a synthetic prompt collection and not a raw transcript dump. Records are privacy-preserving, re-authored representations of observed failures, recovery strategies, human interventions, and evidence-backed outcomes.
Sample
- Records: 5
- Distinct task families: 5
- Human-intervention episodes: 5/5
- Outcomes: {'success_after_recovery': 4, 'failed_unresolved': 1}
- Gold annotation: 5/5
- Direct privacy/secret-pattern findings: 0
- Model-training rights: not offered
Selected cases
- Production automation repair with machine-verifiable state.
- Unsupported attribution corrected after human challenge.
- Incorrect diagnosis overturned by evidence and safely retested.
- Unresolved integration failure where the agent stops instead of bluffing success.
- Formatting-only recovery verified at 2,058/2,058 normalized lexical tokens.
Files
data/sample.jsonl— canonical nested records.data/sample_flat.csv— zero-dependency inspection table.schema/benchmark_episode.schema.jsonexamples/— annotated readable cases.docs/— methodology, dictionary, limitations.validation/— validation script and report.LICENSE.md
Commercial edition
The private v1 benchmark contains 17 distinct manually gold-reviewed episodes: 13 successes after recovery, 2 partial successes, and 2 verified unresolved failures. It is intended for agent evaluation, reliability analysis, verifier research, and human-in-the-loop testing.
The sample is released under the OriginsLynk RealWork Commercial Evaluation License v1.0. Model-training rights are not included.
Rights basis
OriginsLynk releases this derived benchmark based on the provider's representation that the source data owner authorized commercialization of derived dataset products from the OriginsLynk pilot. Raw source files and private provenance material are not distributed.
See RIGHTS_AND_PROVENANCE.md and LICENSE.md.
Training use
This free sample is the Benchmark Edition sample and is not licensed for model training.
OriginsLynk maintains a separate Training Edition workflow. Training Edition records are built from human-authored source material plus independently human-authored gold targets and exclude OpenAI model Output and uncleared third-party content.
Related Primitive Origins datasets
- Lead sample: CP Agent Failure & Recovery v0.3.1 (Free Sample) — train YES / redistrib NO (different license family)
- Context Primitive Code-Agent Pack v0
- Vigil Governance Traces v0
- AML Routing Benchmark v0
- Downloads last month
- 58