What the probe is
When an issue's question is chosen — before any research dossier, any candidate conclusions, or any drafting — the editorial process asks the AI model one unscaffolded question: what do you honestly conclude? The answer is saved to a file and sealed. Nobody in the process — author or AI — reads it again until after the issue's conclusion (the "landing") has been selected through the normal rival-candidate process.
After selection, the probe is unsealed and compared with the selected landing. The comparison is recorded one of three ways — match, partial, or no match — in the issue's internal records and in a running ledger, one row per published issue.
Why it exists
This publication uses AI drafting under human editorial control, and AI models have defaults. Left alone, a model's conclusion on a contested topic tends toward the most common respectable position in its training data; corrected naively, it tends toward "both sides have a point" — the position that asserts least. The landing-selection process (rival candidate conclusions spanning directions, each attacked, with the author selecting) exists to counter both pulls.
The probe is the thermometer for that arrangement. It records what the model's default was, so that when the selected landing agrees with the default, that agreement is visible instead of invisible. A bias you can see is one you can interrogate.
What it has actually shown
Issue 18 (first pilot)
- Recorded match: partial. The probe's near-term direction coincided with the selected landing, but the shape of the verdict differed — the published essay's conditional, rule-scoped conclusion was not what the raw default produced.
Issue 19 (second pilot)
- Recorded match: strong, same direction. The raw default independently produced the same direction the selection process later chose: cost-causation with enforcement, plus public transparency, with a bounded shared-benefit exception. The prior reappeared.
That second row is exactly the case this note exists for. The honest reading is not "the process confirmed the model's judgment." It is: on this question, the selected landing and the model's default point the same way, and the record says so out loud. The essay's claims then have to stand on their sources and arguments — which is where they would have had to stand anyway.
What a match does and does not mean
A match does not mean
- The landing is right. Agreement with a default is evidence about the model, not about the world.
- The process "validated" the conclusion. Survival records in this process report; they never rank, recommend, or certify.
- The selection was captured. A default can agree with a well-selected landing for the ordinary reason that both track the same evidence.
A non-match would not mean
- The landing is wrong, or brave, or original. Distance from the default is not a virtue; it is just a different data point about the model.
What carries the weight instead
- Rival candidate conclusions built in more than one direction, each required to survive attack — including a burden-inversion test that asks whether the argument holds when its subject is swapped.
- Primary sources read in the original, with contested numbers labeled contested.
- Published failure conditions: each issue states what would prove it wrong.
Limits, stated plainly
Two published issues is two data points. At this sample size the ledger cannot statistically distinguish a captured process from an ordinary run of topic-and-evidence alignment, and it will not be able to for a long time. It is reviewed quarterly as a journal, never used as a target — no quotas, no alternation, no steering toward or away from the default.
The probe text itself is never published and is never quoted in any essay or record — publishing it would turn a diagnostic into a marketing exhibit. What is public is this note, the per-issue match value, and the discussion of the match in each issue's development record. The probe is disclosed so that readers know the instrument exists; it is scoped so that it can never be offered as proof.