formalism-open-questions-0-3

Malign hypotheses can end or corrupt the training process before the asymptotic elimination the guarantees rely on applies.

A finite-time form of the deference collapse's guarantees, holding before any malign hypothesis can act, is available by some route.

The logical induction criterion constrains no finite day: a five-coordinate finite perturbation of FAF's logical inductor is a logical inductor and a computable market that violates the interval forms of exact no-net-update and exact self-trust.

In the corrigibility kernel's second box, for a logical inductor the ranking of the violating option below the compliant one holds at every day from the price range alone, while the expectation-provability-induction version is asymptotic.

A constraint on prices compiles to a legal trading strategy inside the market recursion, giving per-date conformance within a schedule while preserving the logical induction criterion.

With explicit schedules, the forcing results of the deference note give a finite-horizon bound on the value at stake.

A non-asymptotic refinement of logical induction exists that licenses a per-decision guarantee rather than a policy-over-a-class average.

Quantitative limitations on an agent's self-knowledge imply quantitative limitations on epistemic endorsement, so that reflection-type trust properties hold only to within a tolerance matching the precision of self-knowledge.

The gap packages and the expert pins the bounds-transfer repair assumes are discharged at grade (a) for a distinct inductor expert.

The modelling assumption that AIs consider hypotheses from a larger complexity class than humans is a faithful model of human-AI capability differences, given the maintainer's observation that humans "seem to generalize better on less data than AIs".

The framework proves positive results in the eliciting-latent-knowledge setting, such as bounded-loss assurances that the humans will not regret not knowing something the AI was thinking.

A model of legitimacy yields anti-wireheading results, closely related to results against manipulation of the humans.

An AI trained in this setting has no misaligned agency, so it can be given a purely epistemic goal.

When humans use the AI to help them think, predicting the future humans reduces to the AI predicting itself, which is a form of illegitimacy, and accurate prediction and active steering produce the same vanishing prediction error off the timely fragment.

The manipulation theorem is assembled from its four ingredients: a second calibration condition separating calibration-to-self from calibration-to-truth, the evidence/preemption distinction, the transfer-of-trust attack, and non-recoverability.

The corrigibility kernel's third box excludes trust laundering, the transfer of authority earned on one class of questions to another.

The open-problem ledger lists every open question of the li-deference notes' section 0.3 and of the deference note's Appendix B, each with its status after the clean room.