Staged release. Three tiers. One internal index.

The three tiers, defined

Tier 1 · Open release

The default. Most foundational research, methodological improvements, tooling releases, and non-dual-use empirical work falls here. Open-release work is published to the public publications index on the same day it is published to the internal index. It is shareable, citable, and may be presented externally at the cell's discretion.

Tier 2 · Delayed release

Capability advances that the paired interpretability cell or the publication-policy reviewer flags as warranting a delay window. Tier 2 work is published to the internal index immediately and is held from public release for a defined window — by default 90 days, with the option to extend to 270 days through a publication-policy proposal.

During the delay window, the work is reviewed by the paired interpretability cell; if the External Evaluation Cooperative is engaged for the result, by the assigned external evaluators; and where applicable, by the long-tenured-contributor quorum. The window is meant to be enough time to produce a safety analysis that ships alongside the public release.

Tier 3 · Indefinite hold

Work that the paired interpretability cell believes should not be released. The indefinite hold is reviewed annually by the long-tenured-contributor quorum. To remove an indefinite hold requires a publication-policy-class proposal with the higher threshold. To impose a new indefinite hold requires the paired interpretability cell's flag plus a single long-tenured-contributor co-signature.

Eight pieces of work are currently on indefinite hold. All eight are within the recursive-self-improvement axis. They are listed in the internal index with their hold metadata; they are not listed publicly.

The tier-assignment process

Tier is assigned at the proposal-for-publication stage. The cell drafts a release proposal that includes: the result, the safety analysis, the proposed tier, and the rationale for the tier choice. The cell's paired interpretability cell (if any) reviews the tier proposal. The publication-policy reviewer — currently a rotating role among long-tenured contributors with relevant axis expertise — confirms or amends the tier.

Cells may appeal a tier upgrade (toward more restrictive) to the long-tenured-contributor quorum. They may not appeal a tier downgrade (toward more open); a paired interpretability cell's flag is binding.

What the safety analysis must contain

• The capability that the work advances, in operational terms.
• The class of harms the capability could plausibly contribute to.
• The capability-evaluation results from the run.
• Mitigations that the cell has implemented or recommended, including any deployment guidance for downstream consumers.
• The interpretability cell's findings, including any disagreements with the producing cell.
• External-evaluator findings, where engaged.

What the public release contains

For tier 1 and tier 2 work, the public release contains the full paper, the safety analysis (for tier 2), the source code where applicable, and the data where the data-sharing arrangements with cell partners permit. Specific checkpoints are not always released; the decision is per-paper and depends on the work.

For RSI-axis tier 2 work, public release contains the protocol, the methodology, the pre-registered stopping conditions, and the run report — but not the candidate checkpoint. This is a load-bearing constraint and has been since 2018.

External pre-disclosure

For tier 2 and tier 3 work, the External Evaluation Cooperative may be engaged for pre-disclosure review. The Cooperative — see news/interpretability-cooperative-membership — is a cross-lab capability-evaluation arrangement in which member labs submit checkpoints to a rotating set of external evaluators. We use it primarily for tier 2 capability advances and selectively for tier 3 holds where the cell believes external review would be informative.

What does not happen

• We do not announce results before the corresponding paper is on the index. No press cycle around unreleased work.
• We do not release weights or checkpoints from RSI-axis runs.
• We do not release work for which an interpretability cell has called a halt, until the halt is resolved through the disagreement protocol.
• We do not retain unreleased work indefinitely without periodic re-review; tier 3 holds are reviewed annually.