Research Safety Company
News
← News

Responsible Scaling Policy, v1.0

A policy is worth the moment it costs something. This one names the conditions under which work stops, written before there is any work large enough to stop.

A staircase of three capability levels. Level one requires evaluation before release, level two external review, and level three stops work until the risk is mitigated. L1 evaluate before release L2 external review required L3 work stops until mitigated capability increases
Figure 1 Each tread is a level, and carries what has to be true to keep working at it. The requirement gets heavier as capability rises, and only the last one costs anything: it is the only one that can stop the work.

This policy is published at version 1.0 and carries a version number for a reason: a commitment that can be edited without trace is not a commitment. Every future change to it will change this number and say what moved.

What it says

Capability is treated as a ladder rather than a line. At each step, something must be true before the next step is taken, and the requirement gets heavier as capability rises: evaluation before release, then external review, then a stop.

The third of those is the only one that costs anything. A policy full of evaluations and reviews is a description of good practice. A policy that contains the sentence work stops, with conditions attached that somebody else can check, is the part that can be broken.

Written early on purpose

Nothing here currently comes close to the thresholds this policy describes. That is exactly why it is being published now.

A scaling policy written by an organisation that already has the capability is a document written under pressure, by people with something to lose by the answer. Written beforehand, the same sentences cost nothing today and are much harder to quietly soften later, because the version history will show it.

What it does not do

It does not define the thresholds precisely enough to be adjudicated by an outsider. Nobody’s does, and pretending otherwise would be the easy dishonesty here.

It does not bind anyone else, and it is not a standard. It is a statement of what this organisation will do, published so that departing from it is visible.

Notes

1 Version. v1.0, published 4 May 2026. Subsequent versions will carry a changelog stating what changed and why. A silent edit to this document should be treated as a failure of the policy itself.
2 The figure is schematic. The three levels are illustrative of the structure. The policy text governs; where the two disagree, the text wins.