Recursive self-improvement.

Position

This is the lab's most closely held research line. We do this work — we do not pretend it can be wished away — but we do it under protocols that we believe are stricter than those used at any centralized lab pursuing this capability. We do it openly enough to be reviewable. We accept that the right outcome is sometimes a halt, and that the value of the protocol is precisely that the halt actually happens.

The relevant numbers: in the protocol's operating history, the pre-registered stopping condition has been triggered three times. Each triggering halted the run, quarantined the candidate checkpoint, and activated the paired interpretability cell. Each halt has been followed by a public delayed-release report describing the trigger, the analysis, and the protocol revisions adopted.

Research threads

• Modification-under-review (MUR). The protocol used when a candidate model proposes a change to its own training procedure, architecture, or evaluation criteria. Decoupled into proposal, evaluation, and incorporation phases; each phase has a pre-registered stopping condition and a paired interpretability cell. Reference: 25/05.
• Capability-evaluation methodology. What we measure, how we measure it, and what the stopping conditions look like. Published in the form of an axis-wide methodology document (24/12), updated quarterly. Each update is reviewed by the cross-axis methodology review pool.
• Isolated compute enclaves. Compute pool capacity that is contractually segregated for RSI runs — separate accelerators, separate networking, separate trace storage. The federation's interconnect agreements include enclave clauses (see news/datacenter-jakarta).
• Stopping conditions. The hard part. Cells in this axis spend significant time on the question of what evaluation metrics, behaviours, or trace properties should be tripwires. The methodology is pre-registered before any run begins; updates are non-trivial governance events.
• Paired interpretability. Every RSI cell is paired with an interpretability cell. The pairing is not advisory: the interpretability cell has rolling read-access to checkpoints, training logs, and proposal commits, and can call a halt that the RSI cell cannot overrule.

On the question of whether to do this at all

What is shared and what is held

Foundational methodology — the MUR protocol, the capability-evaluation methodology, the stopping-condition specification — is released openly. Specific run reports are released with a 90-day delay and an accompanying safety analysis. Run-specific capability findings that the paired interpretability cell flags as dual-use are held indefinitely until the cell signs off.

We do not publish the candidate checkpoints from RSI runs. We do publish the protocol artifacts: training logs, trace digests, and pre/post comparison reports, redacted where the interpretability cell requires.

Where this connects

To agentic engineering. The sandboxed self-modification capability is co-developed with the agentic axis. The agentic axis owns the substrate; the RSI axis owns the experiment and the safety protocol.

To interpretability. Every RSI cell is paired with an interpretability cell. We treat this pairing as the most important single piece of structural infrastructure in the lab.

To governance. Changes to the methodology or to the stopping conditions are non-trivial governance events that require a long-tenured-contributor quorum.

Recent halts

Three halts in protocol history: 22-04 (first triggered, July 2022), 24-13 (October 2024), 25-19 (August 2025). Each halt has produced a public delayed-release report — see 23/02, 25/06, and the forthcoming 25-19 report in late 2025.