6. What Remains Open — Bridge to Article 4

The arc this article has traced can be summarized in a few sentences. Section 0 diagnosed a gap left by Articles 1 and 2: both evaluate a judgment — the AI's, the supervisor's — after it has been produced, with no vocabulary for classifying, beforehand, the task itself being judged. Section 1 recovered the functional architecture of Hammond's CCT that fills that gap: two poles, a dominant middle ground, a task→cognition induction mechanism. Section 2 chose, between two competing formulations, Doherty & Kurz's (1996) list of eleven properties, interpreted through Hammond's (1988) surface/depth distinction — and resolved, with its own justification, a polarity the literature leaves open. Section 3 mapped those eleven properties onto the AI-IoT domain, through a three-level architecture. Section 4 formalized that mapping into an index — $TCI_\ell$ and $\Delta TCI$ — declared as an original construction, not the recovery of a validated instrument. Section 5 tested that index against two concrete cases, and discovered, while calculating, a mathematical ceiling and a blind spot that none of the preceding sections had anticipated.

Closing an article does not mean resolving everything it opened. Three debts remain explicitly unpaid, and it is worth listing them here, together, rather than leaving them scattered across the bibliographic notes where they were flagged:

• The asterisk case (Sections 5.1, 5.2): the "number of cues" property, as inherited from classical CCT, does not distinguish "few analyzable cues" from "no cues, just a conclusion" — a flaw in the property itself, not in a specific case, which this trilogy flags but does not fix.
• The Juslin/Olsson citation (Section 2): used to characterize the opposite polarity of the cue-criterion relation, it remains to be verified directly against the original publications before any formal citation.
• The undetected opacity (Section 5.2): $\Delta TCI$ detects structural displacement of mode across levels, but does not detect an organizing principle that turns opaque without displacing the remaining properties — a boundary of the index's application, not a calculation error.

None of these three is urgent enough to hold up Article 4. But there is a fourth: the limit $|\Delta TCI_{\text{AI-sup}}| \leq 8/11 \approx 0.73$, discovered in Section 5.3, is not a mathematical curiosity to file away — it is a design question for which this article has no vocabulary. If no supervisory interface, however minimalist, can push $\Delta TCI_{\text{AI-sup}}$ beyond roughly $0.73$, then the relevant question stops being "what is this task's $\Delta TCI$" — Section 5 already showed how to compute that — and becomes: within that bounded interval, what position does a given system design choose to occupy, and what authority should that position justify assigning to the AI?

That is exactly the question Article 4 opens with. The four architectures examined there — assisted human autonomy, human-in-the-loop, human-on-the-loop, bounded operational autonomy — differ, among other things, in how much of the task's structure is exposed to the human supervisor: exactly the design lever Section 5.1 manipulated when comparing a rich panel against a poor one for the same thermal detection task. It is tempting to conclude from this that architectures with greater algorithmic autonomy should be reserved for high-$\Delta TCI$ tasks, and vice versa — but Article 4 itself refuses that shortcut in its closing section, which declares itself a "design heuristic, not a correspondence table." $\Delta TCI$ will enter there as one heuristic input among several, not as a lookup key. This article does not resolve that question — it establishes only that the question can now be posed with precision, which it could not before Section 4.

It is worth closing by noting where this article leaves the trilogy, in terms of its own lineage. Section 1 already noted that Hammond was a direct student of Brunswik, and that CCT is born as an extension of the Brunswikian lens model on the task side. This article imported precisely that side — what the task demands of whoever judges it. Article 5, already written in its early sections, returns to the other side of the same lineage: Brunswik's original lens model, and the accuracy of a judgment against a distal criterion. Article 4, situated between the two, is where task classification meets, for the first time in this trilogy, an engineering decision — where "what kind of task is this" becomes "who, or what, should be deciding here." It is that transformation, not its answer, that this article leaves open.

(Bibliographic note: this section introduces no new citations. The debts listed refer to bibliographic notes already recorded in Sections 2 and 5.)