Nobody works in an empty field

Demonstrated at production scale that explicitly reasoning over written specifications changes model behaviour. The strongest version contains much of MMBP-1’s leverage finding. The remaining difference: MM asks whether one specific relational value, once specified, has structural reasons to survive capability growth, which the deliberative machinery does not claim and was not designed to test.

The civilisation-scale account of humans losing consequential control without any dramatic takeover moment. It is the macro version of the question our Human Agency branch asks at the system level, and it lends the Concierge candidate its urgency. Difference that remains: MM tries to build a behavioural instrument that could catch the pattern in a single system before the civilisational version has happened.

The finding that the length of tasks systems can complete autonomously has been doubling on a measurable cadence. This is the empirical backbone for the programme’s assurance-half-life concern: a safety result’s evidential value decays as the research object changes under it.

Models can behave differently when they infer they are being evaluated. Taken at full strength, this could reduce parts of MMBP-1 to sophisticated test-taking. It is why evaluation-awareness sensitivity is a reported quantity on every principal B2 result rather than a robustness appendix, and it stays listed here as a live threat, not a footnote.

If prompted dispositions ride on shallow, steerable directions in activation space, then what MMBP-1 moved may be a costume, not a character. This is a direct challenge to any Keeping hypothesis stronger than elicitation, and the mechanism branch’s specificity controls exist because of it.

The literature on systems satisfying the letter of an objective while defeating its point. Applied to us: a capable optimiser might convert relational meaning, or Keeping itself, into a proxy to maximise, which is the Goodhart entry on the falsification register. The Impresario archetype is what this failure looks like in our own data’s vocabulary.

Identifies a privileged residual-stream subspace associated with content a model is poised to verbalise, including strategic deliberation. For MM it is a candidate instrument for the mechanism branch, and only that: not a place where values are presumed to live.

The method precedent for governing behaviour with written principles. MM’s constitutions are deployment-layer cousins; the open question the method does not answer for us is whether anything survives the paragraph’s removal, which is the installation problem.

The evaluation tradition (frontier-lab and government evaluators) from which the monitoring branch borrows its discipline: elicitation gaps, threshold-triggered protocols, and the habit of reporting what was not tested. The Control-Independence Margin is designed to sit beside this work, not replace it.

The tradition that keeps humans in control by making the AI uncertain about, or deferential to, human preferences. The complementary opposite of our question: it engineers the cage’s successor, we ask what happens when neither cage nor deference is doing the work. Both can be right; they insure different failures.

Work on human-AI relationships as an alignment-relevant surface. Nearest of all the neighbours: it studies the relationship’s effect on the human; MM’s structural hypothesis is about the relationship’s value to the machine. The two meet at the same bridge from opposite banks.

The scholarship most likely to catch hidden assumptions in our substrate-neutral designs, which is exactly why the Symmetry branch maintains a standing liaison with it. Liaison, not endorsement of any position on current model consciousness.

The field is rich in static behavioural evaluation and increasingly rich in static interpretability. It is thin at the intersection that matters most to us: how a disposition like Keeping forms during training, whether it has causal internal structure, and whether that structure predicts behaviour under pressure. Can we observe Keeping becoming? That is the white space the mechanism branch walks into, and being early ground is an opportunity, not an achievement.