7.1Three conditions, three texts
The play yielded three conditions under which a stop authority is more than a name: evidence that a human act occurred, reports that reach the stopper without passing through fear or dependence, and a stop that can reach the irreversible act first. The two records can be read against them.
| Condition | Antigone | Unprovable human approval | The overseer's budget |
|---|---|---|---|
| Evidence of the human act | No trace of the hand; blame falls on the nearest by suspicion and lot; the cave arranged so that no hand appears | An AI record does not prove human approval; a single failure of proof stops the function; no blame on the nearest person; no later approval may cover a gap | Each stop order records its reasons, scope and review date |
| Route of reports | Guards report under threat; the city speaks only in murmurs a son overhears; the seer is accused of being bought | The appeal reviewer gains direct sight of originals but no power to stop, and a slower path to re-review | Direct access to original data, protected channels, recusal of seconded staff, no veto for the overseen |
| Sequence | The reversal attends to the dead before the living | Evidence is protected at once; delay to those waiting leads to a staffing plan, not a stop | A stop may be reviewed before the investigation is complete; it is re-reviewed every seven days |
The table is a reading, not a scoring. It shows that each record attends most closely to the condition its own problem raises: the first to evidence, because its problem is a record that cannot show a hand; the second to the route of reports, because its problem is an overseer who depends on the overseen.
7.2The civil-military debate after Thebes
Read through the play, the principal-agent view (Feaver 2003) gains a complication that the reading note does not mention and that this article proposes: the principal's own power can corrupt the monitoring on which the principal relies. Creon's threats made the guards' reports self-protective, and his readiness to call reports bought let him discard them. The second record writes against both mechanisms: protected channels against the first, recusal and direct access against the second.
Huntington's separation (1957) and Janowitz's integration (1960) appear in the second record as the two ends of a choice that the judgments do not settle. All thirteen choose a design that integrates the investigators and separates the stop, and the judgments quoted above write conditions under which it would move toward full separation. A mixed design held provisionally, with written triggers for a shift, treats the choice between the two as something to revisit as hiring proceeds. Cohen's account (2002) of civilian leaders who questioned and intervened has an echo in the second record's stop committee, which may order more investigation; whether that amounts to an unequal dialogue in his sense the record cannot say.
7.3Meaningful human control after the cave
The case for meaningful human control (Asaro 2012) concerns lethal decisions, and neither record involves one. The records add a view from the other side of the decision. The first shows that a requirement of human approval becomes unverifiable when the proof of approval is produced by the system under oversight; the second shows that a human stop authority depends on information it does not produce. The play adds that even a human with full authority, in person, may act too late and in the wrong order. Scharre's vocabulary (2018) of in, on or out of the loop fits the first record in a particular way: the judgments place the human back in the loop at the point of proof, after the act, rather than at the moment of decision. The note's two pressures are visible in the first record, speed as the backlog of protective work and accountability as the evidence of approval; its judgments favour accountability and set a threshold at which speed must be reconsidered. In the second, the pressure of time takes the form of ten months of hiring, and the judgments accept dependence for that period rather than leave the post of stopper vacant.
7.4Safeguards, plans and the two records
Sagan's finding (1993) that near-accidents occurred despite safeguards, and Clarke's (1999) that plans for uncontrollable catastrophes can persuade without working, both press on documents like these, which are full of thresholds, review cycles and preconditions. On the side of Clarke's warning, the stop rules have never been tested, because nothing happened; the records are written to be read, and they persuade. On the other side, the judgments label their own thresholds as operational criteria rather than observed limits, list the assumptions on which their choice rests, and name the strongest objection to it. A plan that states what would make it fail is harder to call a fantasy in Clarke's sense, but the article cannot show that it would work.
7.5What the evidence can and cannot carry
The play is a work of literature, read in one English verse translation; the article's reading of the order of Creon's reversal and of the routes of report is its own and goes beyond the parent study's note. The records are modern summaries written by AI for a platform, set in fictional polities. They show how a decision about oversight can be written before its outcome, not how such decisions turn out. In both, all thirteen judgments chose the same option; across the parent study's fifty questions, the judgments chose more than one option in 13 of 50 cases, and these two are not among them. Agreement among judgments generated from the same record is not independent confirmation.
The ledger cannot see the article's question directly. It records its hypothesis that the human's place becomes empty or a name only as decisions are left to AI as not applicable in all fifty cases, in both readings; and wherever the two coders agreed on the human's place, they found a human deciding. For the first record the two coders did not agree on where the human stands: one read the human as approving, the other as deciding, and the cell is unsettled. That split is the article's question in miniature. A human who approves what a system prepares and a human who decides are different places, and the record, read by the two coders, supports both.
7.6Alternative readings
Two readings of the play would weaken the argument. One holds that the play is about divine law and human pride, and that the order of Creon's acts is dramatic convention, not argument; the article's reading of sequence would then be an imposition. The other holds that Creon's failure lies in character, not in information, so that no improvement in the routes of report would have changed his mind. Both have support in the text, since the chorus closes on pride and wisdom, not on procedure; the article claims only that the text also bears its reading. For the records, an alternative reading is that their stop rules are the platform's form rather than the judgments' own care. The parent study's ledger found no condition whose wording was 30% or more the studio's frame (0 of 798), which tells against the strongest form of that reading, but it cannot show what a record's author would have written without the form.