# Full-scope acceptance tracking

The goal is still active. The following distinguishes partial artifacts from
proof of the requested end state. Nothing here narrows the original requirements.

| Requested deliverable | Current evidence | Remaining proof/work |
|---|---|---|
| 1. Chosen architecture and reasoning | Mean, GRU, Transformer implementations; experiment reports | Final selection on realistic tasks and target hardware |
| 2. Parameter-count target | First trained mean baseline: 128,900 | Evaluate requested 50M–1.5B range and smallest reliable option |
| 3. Model implementation | Learned classifier and pointer | History, planning, argument spans, all action types |
| 4. HF training pipeline | PyTorchModelHubMixin save/load, CPU trainer | Third-party data ingestion, pinned base checkpoints, reproducible larger runs |
| 5. Dataset schema | Synthetic v1 JSONL and manifest | Normalized real trajectories, license/privacy gates |
| 6. Synthetic generator | Five layouts, three action classes | Multi-step, dynamic, modal, scrolling, ordinal and adversarial tasks |
| 7. Trajectory recorder | HMAC metadata SQLite | Reviewed redacted training trajectories and outcome linkage |
| 8. Experience memory | Not implemented | Reusable retrieval with poisoned-data and privacy protection |
| 9. SHA-256 subsystem | Canonical state and signatures | Correct full-state cache identity and validated cache reuse |
| 10. Compact DOM | Semantic rows, frame/shadow handling | Delta integration, bounds, faithful ARIA, value/history representation |
| 11. Action DSL | Strict 18-action parser | Actual tokenizer comparisons where generative policies are evaluated |
| 12. Safe executor | Authority, stale node, role and host policy gates | Network/effect isolation, redirects/downloads, security audit |
| 13. Recovery controller | Loop rejection and observe action | Learned recovery from enumerated browser failures |
| 14. Completion verifier | Host callback bound to task | General evidence-based user-goal verification |
| 15. CPU runtime | Actual local CPU inference | Linux EPYC two-vCPU end-to-end validation |
| 16. Quantized model | Linear INT8 artifact exported, reloaded and benchmarked; slower than FP32 | Broader quantization/runtime comparison and deployment selection |
| 17. Benchmark harness | Browser fixtures and CPU microbenchmarks | Real sites, baselines, full metric set and held-out audit |
| 18. Regression suite | Parser, state, Chromium, model, registry tests | Complete action/security/generalization coverage |
| 19. Model registry | Immutable copies, integrity and promotion/rollback gates | Full evaluator evidence and real champion; none promoted |
| 20. Continual improvement | Offline trainer; promotion gates | Failure mining, curated dataset deltas, replay and automated cycles |
| 21. Reproducible commands | README and module CLIs | Clean Linux install verification and dependency locking |
| 22. Measured CPU/RAM/latency | Local reports with measurement scope | Target VPS and full service peak measurements |
| 23. Known limitations | README, DATASET, STATUS, experiment reports | Keep updated as scope expands |
| 24. Roadmap | STATUS experiment order | Evidence-driven next-generation plan after full evaluation |

Five meaningful iterations are required but are not a substitute for acceptance.
The production promotion gate requires target-hardware evidence and task,
generalization, recovery and security metrics that current microbenchmarks lack.
Registry unit tests use labelled fake fixtures to test transitions; those fixtures
are never registered in the actual project model registry.