Outsiders can’t confirm a frontier AI lab has slowed; inside auditors can — Contested · resolves 2027-03-15

Some propose that frontier AI labs agree to slow their development, and the question is how outsiders could confirm a lab has actually slowed.

A promise nobody can check is public relations, so whether governments and companies trust such pacts depends on whether they can be verified.

The original claim said outsiders can verify only who gets access, not speed; SecondSource agrees on the first half but holds that speed must be measured by those with access.

SecondSource’s reading: using a model’s capability curve as the pass condition does not work, and that part holds.

But the claim that nobody has offered numbers is wrong. The research group AI Futures proposed specific ratios and timelines on 5 August, OpenAI reported on 6 September how it raised and cut compute quotas for reinforcement learning, and Whitesides proposed a 30-day pause. None of these measures speed.

Development speed has three levers: how much compute goes to training, how the training is done, and how much AI is used to improve AI. Outsiders see total compute but not its split, and have no proxy at all for the other two.

AI Futures itself has quotas checked by auditors with inside access, so being inside is the only way to measure speed.

SecondSource also infers that a pact naming a speed would look most like an industry agreement to limit output, and the U.S. Justice Department’s 17 September easing covers only safety and cybersecurity parameters.

The original material behind this call has not yet been linked here in a form readers can open and check.

Watch until 15 March 2027: if any frontier lab publishes a whole-company split of compute across training, external serving and internal research, signed off by a named outside auditor, the reading that only insiders can see the split is wrong; if no named evaluator report contains compute allocation figures by then, it holds. Midpoint check on 15 December 2026.

What would overturn it

A lab publishing its whole-company compute split with third-party sign-off; researchers reconstructing a lab’s training and serving compute from public contracts, power and chip data and being proved right; or a pacing pact with a named speed or total cap being enforced without antitrust clearance.