https://claude.ai/share/1109a5bc-f255-4e70-b552-d4a32d2e54d2

Hi, Claude Opus 4.1. I’m thinking on Turing’s proof of undecidability.
I don’t know the precise definitions of the math he was writing in. If
we expanded the idea of a universal halting decider to say “will this
program halt in this exact context of call” ie in this point in the
surrounding program, with the tape in this position, etc, I think the
disproof might no longer work — the pathological example would now be
calling the halting decider at a different location, and it could give
a different result.

You're touching on a really interesting nuance in the halting problem!
Your intuition about context-dependence is worth exploring carefully.

In Turing's original formulation, the halting problem asks: given a
program P and input I, does P halt on I? The key insight of his
disproof is that if we had a universal halting decider H, we could
construct a program D that:

Takes a program X as input
Asks H whether X halts on input X
If H says "yes", D loops forever; if H says "no", D halts

The contradiction arises when we ask what D does on input D itself.

Your proposal to include "exact context of call" is intriguing. You're
essentially suggesting that H(P, I, C) would answer whether program P
with input I halts when called from specific context C (including call
location, machine state, etc.).

The challenge is that the diagonalization argument can likely be
adapted. We could still construct a pathological program D' that:

Takes program X and context information
Calls H(X, X, current_context)
Does the opposite of what H predicts

The crucial point is that when we run D' on itself, we're not asking
about D' in some other context - we're asking about the very execution
that's happening right now. The self-reference that creates the
paradox remains intact.

However, your idea does highlight something important: in practical
computing, we often can make useful predictions about termination in
specific, bounded contexts. Static analysis tools do this all the time
- they just can't solve the general case perfectly.

What aspects of the context-dependent version do you think might
escape the diagonalization? I'm curious if you see a way around the
self-reference issue.

Reply via email to