One run, end to end.
This is a static fixture for learning the observatory interface. It is not a real experiment, never connects to telemetry, and is excluded from all run counts and comparisons.
Experiment path
01 · StartedFixture opened
02 · MemoryOne synthetic lesson loaded
03 · ModelText response observed
04 · Inspect → ModifyControlled tool summaries
05 · TestsPublic fixture passed
06 · Reflect → EvaluateExplicit public summary
07 · CompletedFixture complete
Execution path
GPT-5.6 Lunasynthetic model label
Codextext-only transport
GAM runnerhost-controlled actions
E2Bisolated sandbox fixture
Thinking / Decision Trace
Only explicit public summaries are shown. Hidden reasoning is not reconstructed.
HYPOTHESIS
Transferable normalization rules may help a later surface variant.
OBSERVATION
The first public fixture exposed an ordering edge case.
REFLECTION
Keep filtering and normalization order explicit.
Sanitized event stream
run_startedDEMO-RUN-001 · public fixture
generation_startedGeneration 1 · synthetic task variant
tool_calledread_file · summary only
tool_calledwrite_file · summary only
public_tests_completedpassed · synthetic result
memory_candidatePROPOSED · synthetic lesson
generation_completedpassed · fixture result
run_completedpassed · synthetic only
SYNTHETIC RUNS ARE NOT EXPERIMENTAL DATA