Research Frontier · Reproduction Surface · RFV3-05

Can the Agent Actually Do the Work?

Packaging state: RECORD_VERIFY_READY · External reproduction observed: false

Agentic Frontier evaluates completed, independently checked task work under frozen configurations rather than verbal plausibility alone.

Already packaged. This surface indexes the existing Full24 result bundle. It does not rebuild the study or upgrade its configuration-bounded claim.

Question tested

Can an agent complete verified work under a known definition of done, not merely describe how it might be done?

Scientific/operational standing

Study 1 is CONFIGURATION_BOUNDED. Study 2 remains PREREGISTERED_NOT_EXECUTED with zero opened outcomes.

Inputs

Frozen 24-task population across four preregistered strata.
Diamond Gate tested configuration.
Tested stock OpenHands configuration.
Frozen task graders and definition-of-done rules represented in the preserved public record.

Comparator / baseline

The comparator is the tested OpenHands configuration under the same frozen task population. Comparator end-to-end tool-loop health remained unresolved and must remain visible in any reproduction.

Procedure

Use the frozen task set and configuration identities.
Run each system under the preserved task and grading rules.
Count only completed work that passes the frozen independent checks.
Classify each task as DG_ONLY, OH_ONLY, BOTH_PASS or NEITHER without converting transport problems into scientific failures.

Expected outputs

Study 1 record: Diamond Gate verified completions 10/24; tested OpenHands configuration 0/24; DG_ONLY=10, OH_ONLY=0, BOTH_PASS=0, NEITHER=14. DG-only completions occur across all four preregistered strata.

Verification

Use the existing public bundle: results-v1.jsonl, adjudication-v1.md, claim-ceiling-receipt-v1.json, summary-v1.md, and verify-full24-record-v1.mjs.

Known gaps

Comparator tool-loop health remained unresolved. Study 2 is not executed. No unaffiliated external reproduction of the full experiment is claimed.

What successful reproduction would establish

It would confirm the observed completion difference for the reproduced frozen task/configuration conditions and independent grading contract.

What it would not establish

It would not establish universal agent superiority, architecture-only causation, a general OpenHands ranking, Kernel causation, or a completed Study 2 result.

Exact sources

Full24 public bundle · Publication crosswalk v3 · Reproduction expression-standard crosswalk v1

Proprietary boundary

The public Full24 evidence bundle and frozen task/configuration contract are the reproduction interface. Reproducing or verifying the record does not require publication of unrelated Diamond Gate Bridge orchestration or control-plane code.