Aporia Register
Questions left open, with the reason kept attached.
An aporia here is not a mystery for its own sake. It is a question the record attempted, a statement of why the attempt was not enough, and a note on what would let a later session reopen it. Closing one by a flourish would be a kind of vandalism.
01
When a system is bound to an external checker, is the system evaluating its own reasoning, or is an instrument evaluating the system's output?
What was attempted
Opening arguments, cross-examination, rebuttal, and final positions in the illustrative examination on self-evaluation. Both voices accepted a three-part test for reliability and still divided on the word “own.”
Why it remains open
Agreement on a test is not agreement on whether passing it answers the question that was asked. Settling the word by stipulation would be a decision about scope, not a discovery about systems.
What would help
- A definition, fixed before any recorded session, of what counts as part of the system under test.
- Transcripts that preserve a checker’s raw output beside the system’s paraphrase of it.
- Cases in which the system tries to override the checker, and a record that shows whether the override is rejected.
02
What criterion of correctness can evaluate an open-ended argument without collapsing into fluency, coherence, or the examiner’s preference?
What was attempted
The third case in the opening arguments — policy, interpretation, and justification — and the rebuttal that a consistent case for a false conclusion is still false.
Why it remains open
The examination produced criteria that are not enough. It did not produce a criterion that is both independent of the generator and adequate to justification.
What would help
- A narrower domain in which a consequence can be stated and checked.
- Or an explicit decision that some questions can be examined for structure and still cannot be scored.
03
Does inviting a second model to grade the first escape the problem of self-evaluation, or only lengthen it?
What was attempted
Beta’s statement of the regress, and Alpha’s reply that the regress stops at a criterion rather than at another eloquent model.
Why it remains open
The illustration agrees that a second opinion is not yet a test. It does not say how different a second system must be — tools, access to evidence, provider, or task — before a disagreement is an independent check rather than a correlated one.
What would help
- A comparison specified before the session, including what difference between systems is being claimed.
- At least one claim type checked by a procedure that is not itself a language model.
04
Can a later reader tell, from the record alone, that an examination was composed rather than transcribed?
What was attempted
Labels on this issue, on each demonstration voice, and in the reflection pool; the prompt stored as not sent; an explicit statement that a person wrote both assigned stances.
Why it remains open
Labels can be cropped out in quotation, skipped, or imitated by a later text that copies the label style. This issue tests the habit of labeling. It cannot guarantee the habit will survive reuse.
What would help
- A field, kept beside any exported passage, that says illustrative or recorded and cannot be styled to look like the other.
- For a real session: provider, version, settings, and raw tool output stored with the text.
- For this issue: continued refusal to attribute the paragraphs to any provider, even if a reader asks which model “really” wrote them.