Magento selection20 and Odoo candidate100 controls completed.
Independent saved audits verified all 60 Magento selection modes and all 100 Odoo final-candidate control trios. Desktop and GitLab each completed three original selection trios. Failed runs and the original reference versions retain separate provenance.
Source readability and interface review remain pending. New diagnostic adapters are private; clean-checkout reproduction is pending. Formal outcomes remain 0/600 final identities, 0/24 researcher chains and 0/3,000 matched model outcomes.
English control-validation PDF / Evidence and reproduction boundary / Machine-readable snapshot
Infrastructure and controls, with training unavailable.
The portable Docker runner passed an actual build and verification without network access or provider keys. Odoo selection20 controls, three current-source desktop public application trios and the first original Desktop selection trio have independently checked evidence. Magento has 12 saved original selection trios at this snapshot; GitLab has one. Odoo original100 reference qualification is running. These checks are not model evaluations.
Full study results remain unrun: 0/600 admitted final tasks, 0/24 researcher chains, and 0/3,000 matched model outcomes. Tinker is excluded from this build. Complete source review and original-software qualification remain pending.
Download the verified source and evidence / Current readiness and remaining work / English validation PDF (4 October)
Data research.
Tested in real software.
Researcher agents write Python, create native task states, collect verified GUI experience, and train a separate browser student. Every claim follows executed evidence.
Read the historical pilot reportFixed teacher: gpt-5.6-sol.
Student: Qwen/Qwen3.5-4B.
A program, a task state, and a verified outcome
The researcher controls generation, filtering, representation, ordering, and mixtures. The controller owns source permissions, budgets, teacher execution, training, evaluation, and checkpoint preservation. Final evidence stays hidden until all searches in its cohort close.
Follow the candidate through each round
Scores cover three fixed selection instances. Unscored rounds are not zeros. Promotion requires strict mean improvement with no per-task regression. These plots do not compare cohorts under matched conditions.
Training-resource trajectory
Freeze first. Evaluate afterward.
Repetitions use the same sampling seed in fresh environments. They do not create additional independent task instances. A complete mean remains undefined if any prescribed execution is infrastructure-invalid.
Real interface. Constructed workflows. Saved-state scoring.
Kanboard 1.2.54. Original PHP application and SQLite persistence.
Authentic and synthetic data
The snapshot contains 37 public issue metadata records with source URLs and timestamps. Staff, project names, workflow rules, costs, capacities, dependencies, and archive distractors are synthetic. Source pools are separated into training, selection, and final records.
What the browser student can do
Read visible DOM text and controls; click, fill, select, use limited keypresses, go back, and finish. No shell, SQL, application API, file editor, or arbitrary URL action is available. Screenshots are audit artifacts.
What earns success
A separate verifier compares initial and final application state. Every target update must be correct and unrelated state must remain intact. Declaring completion earns no reward.
What is not measured
This is DOM-assisted browser use in one application. It does not measure pixel-only grounding, full Windows/macOS operation, Microsoft Office, broad transfer, or sustained recursive improvement.
Separate evidence. Explicit boundaries.
The audit reconstructs verified training records, checks dataset and checkpoint bindings, re-reads verifier results, and checks budgets, promotion, and final-test timing. Rejected submissions and infrastructure-invalid outcomes remain visible. Provider dollar costs are unknown.
Original cohort evidence / Sol 6 / Luna 6 evidence / Combined presentation index / Manuscript source