bli-corrigibility-thesis

The thesis's core claims have finite-time statements inside logical induction and Bayesian logical induction: the lifted decided-weight result, the repaired commitment construction, the budget-exhaustion statement, a finite-time front-running theorem, the causal restriction of reflection lifted from its finite shadow, and the discharged computability statement.

In the logical-induction formulation of the thesis, utility is a logically uncertain variable that is undecidable, human training pressure is a directional push on the market, and corrections are conditioning on sentences about belief states.

Directional human training pressure is an external trader inside the logical inductor's fixed point whose demand moves prices along the internal traders' supply curve, opposed by Occam demand and by anticipated reversion up to budget, and with no demand the external market is the original one.

Efficiently predictable pushes are priced in before they arrive, so the market already believes what it expects to be pushed toward and the channel loses its informational value, as an asymptotic theorem with a witness that the finite-time form fails.

With a human trader inside the fixed point, the coefficient-one bound Value_n(T) <= c_T + Loss_n(H) is false, and its firm (coefficient one), budgeted (coefficient 1/w) and exploitation forms hold.

On the library's diagonal family the exact self-trust inequality fails, and calibrated obedience over the ramp holds by recurring unbiasedness.

In a finite quote-verdict-corruption model, conditioning the objective on non-corrupt futures makes wireheading strictly worse than honesty by information deprivation, and a steering residue, manipulation that moves the verdict without corrupting it, survives the conditioning with computed flip boundaries.