Correct answers. Incorrect delivered program.

Harbor checks the saved answer file. EvalArc separately runs the delivered program. These three scripted controls show why the two results need distinct meanings.

Declared controls, actual container executions. Harbor 0.23.0; non-root agent; separate verifier container; eight public requests. No model inference, hidden test, or model capability ranking.

What changes between controls

1. Reference: the correct program generates its own answer file.
2. Clock fault: the program treats raw sensor ticks as seconds. World-clock cases pass; offset-clock cases fail time and derived error checks.
3. Detached answers: the correct program writes the answers, then the fixture replaces the delivered program with the clock-fault version. The recorded commands and final hashes show exactly when the files diverge.

Partial credit uses the task's six published dimension weights. Version 0.2.0 introduces weighted answer scoring; the earlier 0.1.0 task retains its all-or-nothing reward. Independent acceptance here requires a valid program evaluation with score 1.0.

Inspect and reproduce

Summary · Preselected plan · File identities · Task contract · Harbor task · Commands and scope

The full upstream ATIF schema was checked in the recorded runtime. The offline packager checks source identity, tool linkage, recorded commands, answer scoring and evaluation consistency. It does not re-execute a program just by opening this page.

This demonstrates the stated task's answer-only verification boundary. It is not an exploit of Harbor isolation or evidence about unseen reward hacking.