Reference
Roadmap
This is the proposed order of work. It is a sequence, not a schedule: no dates are given, because dates that cannot be supported would mislead. Each phase lists what would be published and how completion can be checked.
Status. Phase 0 is in progress. Later phases are proposals.
Phase 0: Foundations
Goal. State the agenda in public and invite criticism.
| Deliverable | Check |
|---|---|
| This documentation | Every source is listed in References |
| Public issue channel | Corrections are logged and answered |
Phase 1: Measurement
Goal. An open evaluation harness with reproducible baselines.
| Deliverable | Check |
|---|---|
| Evaluation tasks written in Inspect | Anyone can run them with one command |
| Baseline results on open-weight models | Logs published, with confidence intervals |
| A private held-out set | Custody and rotation policy published |
The method is in Evaluation harness.
Phase 2: Safe agents
Goal. Open tooling for running agents under control.
| Deliverable | Check |
|---|---|
| Sandbox and policy gate | Reference implementation and tests |
| Monitoring and audit log | Published false-negative rates |
| Control evaluations | Red-team results reported with the safety and usefulness trade-off |
The method is in Safety engineering and Control and oversight.
Phase 3: Research
Goal. Original results on the open problems listed in the Research section.
| Candidate question | Why it is tractable |
|---|---|
| Reliability of agents at 95% and 99% thresholds | Needs careful measurement more than compute |
| Cost-normalised evaluation | Uses existing models and public tasks |
| Auditing games for interpretability tools | Small models are enough to test methods |
Each result is published with code, data manifest and logs.
Phase 4: Network
Goal. A governed, disclosed mechanism for sharing.
| Deliverable | Check |
|---|---|
| Governance specification | Published before adoption |
| Token disclosures | The six items listed in Token |
| Independent audit | Report published in full |
What would change the roadmap
- Evidence that a proposed method does not work. The page describing it will say so.
- A material change in the law that applies.
- A decision on the items listed as undecided in Project status.
How to hold the project to this
Each deliverable above has a check that an outsider can perform. A phase is complete when its checks pass in public, not when it is announced.