LUNA · EVALUATION EVIDENCE
AI quality
What we tested. What still needs work.
Loading the latest published results…
Application checks
Real Luna application logic with fictional identities and simulated model responses. The browser journey is included.
About these results
A pass means the selected code checks succeeded. It does not measure real-model understanding or certify production reliability. Browser journeys use test identities. Unavailable and unreviewed checks remain visible.
REAL MODEL · BOUNDED SAMPLE
GPT-5.6 Luna
Four fictional replies are a small sample. Literal checks look for specific expected facts; semantic quality still needs human review. Cost is service-accounted usage for this sample, not a price estimate for future use.
INDEPENDENT TEST CASES
Held-out assessment
These independently authored checks exercise application contracts using simulated responses. They do not establish real-model quality.
Earlier assessment correction
The first assessment silently ignored unsupported inputs. Independent review corrected its six apparent passes to two contract passes and four unexecuted cases. The evaluator now rejects those inputs. A new set was sealed for the current assessment.
COMPARISON ONLY · DEFAULT UNCHANGED
GPT-5.4 Mini comparison
The same two fictional scenarios as the earlier Luna sample, with four sequential replies and no paid judge. Different run dates and service versions mean this is a bounded comparison, not evidence that one model is better.
Observable request traces
Completed comparisons publish spans for identity, consent, context, provider call and response. No hidden reasoning, credentials or conversation text is published. Retrieval and tools are not used by this chat-only run.
FICTIONAL GOLDEN CANDIDATE · HUMAN REVIEW PENDING
One story, every stage
An original fictional journey with expected inputs, outputs and safeguards. This example is not an executed end-to-end result.
The dataset is maintained as separate story, source, label, retrieval and stage files. Each stage below has its own JSON link.
Loading the fictional journey…
Coverage across the application
| Layer | Passed | Failed / errors | Pending / blocked | Not applicable |
|---|
Autonomous agent routing and vector ingestion are absent from this product. Their absence is not a passing test. Layer counts remain separate from real-model sample results.
Before production readiness
- Human review of model correctness, faithfulness and uncertainty.
- Judge calibration against genuine human labels.
- Production reliability, load, recovery and native authentication evidence.
- Approved measurement of real user outcomes.
This is a published snapshot, updated after reviewed runs. Opening this page does not run evaluations or make paid model calls. Full private captures and reviewer notes stay in the workspace. Sanitized execution spans for the comparison are shown above.