Benchmark evidence: guarded form plans (ca77a1d)

Benchmark evidence for guarded form plans

Supporting evidence for grabbou/jevil PR #1. This is not an npm package release.

Measured code: ca77a1de464aaac19a115c09c58c267dfc3f38ae on piotr-graczyk-dev/jevil, branch feat/guarded-form-plans.

The bundle includes all 90 timing trials, the paired analysis, source hashes, and sanitized per-run reports, screen snapshots, screenshots, and native fixture outcome evidence. Personal filesystem paths and device identifiers are removed. Only the artificial UIKit benchmark app was exercised. The model used was jev-1.13.0; no credentials are included.

Video files are separate demonstration runs, recorded after the unrecorded timing series. Their durations are not benchmark samples. Each filename identifies the task and strategy. These are real recordings at normal speed.

To reanalyze the public results using the PR source:

node benchmarks/analyze.mjs results.json

Reproduction commands and limitations are described in the PR. The full fresh post-run snapshot is not retained by the harness; the saved verdict snapshot, native state at the verdict and afterward, and live oracle result are retained. The outcome checker uses the full fresh snapshot during execution.