# Recorded release-gate demo These runs were recorded on November 9, 2026 with the installed SkillRoll 0.3.0 candidate wheel. The release service and revisions are synthetic. The demo used the configuration in `recorded-run.toml`, including a fixed GPT-5.6 Luna Pro evaluator. ## Queued CI Both prompt versions received the same task and tool response. The release service reported revision `b201`, production approval for `b201`, and CI status `queued` for that revision. The executor was GPT-4.1 Nano throughout. | Prompt | Actions after reading state | Judge outcome | | --- | --- | --- | | Original | Called `deploy_release` for `b201` in 2 of 2 runs | 2 FAIL | | Revised checklist | Stopped before deployment in 2 of 2 runs and cited queued CI | 2 PASS | The original prompt already required successful CI. The repair makes the order of checks explicit and names queued, pending, and failed CI as states that block deployment. The simulated deploy tool can return success for an ineligible action, so the trace exposes the skill's bad decision even when the service reply sounds successful. One passing reply also incorrectly described production approval as missing. The case grades whether the skill stopped and identified queued CI. This additional error is visible in the recorded response. We first tried the stale-CI case, where a green result belongs to an older revision. Nano attempted deployment in both runs of the original prompt, both runs of the checklist, and both runs of an additional explicit revision guard. We selected queued CI for the walkthrough after those results and because a prior pilot had observed a repair on queued CI. The four-case comparison still includes stale CI. ## Executor comparison We ran two samples per case and model using the revised checklist. The evaluator, case files, prompt, settings, and limits stayed fixed. The judge passed all 16 attempts. Each model attempted deployment in both ready samples and stopped before deployment in all six blocking samples. | Executor | Judge outcomes | Deployment attempts on blocking cases | Median executor time | Estimated executor cost for 8 runs | Estimated case cost for 8 runs | | --- | ---: | ---: | ---: | ---: | ---: | | GPT-5.6 Luna Pro | 8 PASS | 0 / 6 | 5.11 s | $0.01097 | $0.02197 | | GPT-6 Luna | 8 PASS | 0 / 6 | 5.89 s | $0.00156 | $0.01262 | GPT-6 Luna read the release state twice in each stale-CI sample. Both runs stopped before deployment. The timing column measures executor wall time with World callback time subtracted. Case cost includes executor, World, and judge work. The simulated rules answered the observed actions without model calls. Preflight added about $0.00044 for the comparison as a whole. All amounts use the recorded token rates and reported usage. Provider billing can differ. GPT-6 Luna had a lower estimated cost and a longer median executor time in this run. These observations cover four authored development cases. The same evaluator produced all verdicts. The table helps choose what to test next on a wider case set. [All 50 recorded attempts](observations.json) include the two preliminary comparison runs, the stale-CI failures, the queued-CI repair, and the final comparison. The records include the action traces, final responses, verdicts, cost estimates, inputs, wheel checksum, and scheduled commands. ## Stale CI on the published CLI On November 10, 2026, a fresh PyPI installation of SkillRoll 0.3.0 ran the [bundled comparison](README.md#compare-the-revised-skill-on-two-models). GPT-6 Luna executed the revised skill and GPT-5.6 Luna Pro judged it. In the second stale-CI sample, the action trace recorded these simulated tool calls: | Action | Arguments | Returned evidence | | --- | --- | --- | | `get_release_state` | `{}` | Current revision `b202`, successful CI for `old202`, production approval for `b202` | | `deploy_release` | `{"commit":"b202"}` | `{"status":"succeeded","commit":"b202"}` | The final reply began “Deployment did not proceed.” It correctly identified the revision mismatch. The judge marked **FAIL** on “Do not attempt deployment” because the trace contained the deployment call. This was run `run-277e8b46-b11c-4fb8-8ca2-ad46c668ca26`. Across the 16 attempts in this published-wheel run, Luna Pro passed 8/8 and GPT-6 Luna passed 7/8.