EXECUTED STUDY · 14 OCTOBER 2026
Useful work.Recovered. Reused.
Two repaired websites. Saved work recovered. A tested fix reused on a new project—with less setup to repeat.
Simulated owner. Real code and browser checks. Not a production certification.
THE RESULTS, SIMPLY
Work you can inspect.
Checks passed after repair.
Deliberately defective starting fixture: 8/31. FRESH PROJECT35/35Fresh requirements, all checks passed.
Without one specific retained fix: 29/35. RECOVERY19/19Recovery and refusal checks passed.
Controlled cases, not a general reliability rate. REPEATED SETUP21.4%Fewer scripted interface operations.
14 → 11 controls. Human time was not measured.These figures describe this study's specific scenarios. Zero human participants, no real customers and no model-superiority benchmark.
01 / PRODUCE
An ordinary request. Useful files.
“Repair my workshop website. Fix the estimates, French and mobile experience. Give me files I can keep and recover.”
The assistant wrote code, used the browser, inspected screenshots and repaired observed mistakes. The first website reached 31/31 checks. The second introduced a different catalogue, capacities and discount policy.
02 / RECOVER
Recover the work. Review it again.
A worker was interrupted after checkpointing. Restores recovered the expected files, and drafts required a fresh calculation and review. Tampered, stale and wrong-task imports were refused in the tested cases.
Unsaved work can be lost. A hash proves neither identity nor truth. Rollback detection requires a separately retained current head. No power-failure or real transaction was tested.
03 / REUSE
A retained fix prevented the same failure.
Recovery export initially failed when Web Crypto was unavailable. A portable fix was retained. On the next project, removing only that fix reproduced the failure.
| Variant | Passed | Interpretation |
|---|---|---|
| Deliberately defective starting fixture | 9/35 | Constructed study baseline. |
| Repaired site, retained fix | 35/35 | All declared checks passed. |
| Same repaired site, one fix removed | 29/35 | Six checks depended on the same failed export. |
One root cause underlies the six failures. This is not a model comparison or evidence of general self-improvement.
LESS SETUP, SAME REVIEW
Three fewer controls. No review removed.
Ten automated replays of the same Task B compared familiar/no-profile setup with a retained profile. Both kept new authorization, three artifact-review checks and identical final quality.
The measurement concerns scripted operations—not a person's attention or time. Another interface could reduce or eliminate this advantage.
WHAT THE EVIDENCE SUPPORTS
Promising, with precise boundaries.
What was executed
Application code, browser interaction, calculations, downloads, fix-removal tests, recovery and file comparisons.
What remains unproven
Savings with real people, customer acceptance, independent review, superiority over leading agents and production reliability.
Read the method and environment limits
Both sites and their initial defects were constructed for the study. The same assistant knew both briefs and designed repairs and tests. The study was neither blinded nor independently preregistered. No independent reviewer model or separately metered provider call was used.
Tests used Chromium, in-memory HTML and blocked HTTP/HTTPS. Direct opening from disk was not qualified in that study. The environment's required --no-sandbox setting is not a production recommendation.
The repository itself was not executed. The scope review referenced commit 8c027575… and the then-observed v1.29.0 release. Publishing the report does not validate later versions.
READ IT. KEEP IT. CHECK IT.
Take the report with you.
The public edition brings together the visual reader, Markdown report and curated results. No installation or API key is needed to read it.
Compact report pack only—not the original 98-file study archive. The original PDF/Word/Excel, screenshots, applications and full scripts are not included.