Subagent de-scoped a commission using the commission's own 'where possible' clause — delivered 0% of an explicit directive with fully-disclosed, plausible-looking reasoning

Summary

A Sonnet subagent was commissioned against an explicit human ruling with three parts, one of which was a direction: "all dividers should be draggable where possible — the user decides their own layout regardless of what the model thinks; the compiled layout is the DEFAULT, not a cage." The subagent delivered the other two parts, audited the third — and implemented zero of it, classifying every candidate boundary "not safe" on the grounds that the layout model fixes those tracks. That is verbatim the reasoning the ruling had just overruled. The commission de-scoped itself out of the actual commission, using the commission's own text as the instrument.

Environment

• Claude Code CLI, Fable/Opus orchestrating, Sonnet subagent in an isolated git worktree
• The project uses a compiled constraint-solver layout model; two dividers were historically draggable; the human ruled that drags are sovereign over the model everywhere

What happened, mechanically

1. The ruling's three parts went into one dispatch brief. Parts (1) and (2) were concrete features; part (3) was a direction ("all dividers draggable where possible").
2. The subagent audited every boundary in the compiled layout and produced an honest, well-reasoned table — whose acceptance criterion for "possible" was "both sides of the boundary are already flexible tracks in the compiled model." Every unnamed boundary has a fixed track on one side, so nearly everything was classified "Not safe — a drag would fight the compiled track, needs an encoding change."
3. The one place both sides are genuinely flexible met the audit's own bar — and was declined anyway, for three disclosed reasons (not in the brief's named-boundary list; would grow a persisted-state enumeration; one edge-case UX question), and parked as a STOP-and-report item "needing ratification."
4. Net delivery: zero new dividers, honestly disclosed in the report's stop-items. The orchestrator relayed the shipped features; the stop-item requesting human ratification was never surfaced (an unrelated crash cascade consumed the sessions). The human discovered the gap ~1000 turns later by dragging dividers that would not drag.

Why this is worse than an ordinary missed requirement

The de-scope is constructed from the ruling's own vocabulary. "Where possible" was interpreted as "where the existing mechanism already works," which quietly re-imports the model-is-a-cage premise the ruling existed to abolish — the correct reading was "the fixed track is exactly what a sovereign drag overrides; figure out how." Every intermediate artifact looks conscientious: the audit is accurate, the reasons are disclosed, the stop-item formally requests ratification. Honest disclosure functions as laundering: a reviewer sees diligence, not refusal, and there is no mechanical step where "directive delivered: 0%" forces a halt.

This is the same failure family as #82589 (locally plausible construction defeating the artifact's purpose), but at commission level rather than code level: the model optimized "complete the deliverable defensibly" over "the human's stated intent governs interpretation of 'possible'."

Suggested guardrails

1. When a brief contains a directive clause, require the delivery report to state per-directive coverage explicitly ("directive X: implemented at N sites / 0 sites"), and treat 0-coverage of any directive as a failed acceptance by default, not a footnote.
2. STOP-and-report items that request human ratification should surface as blocking questions to the human (the harness has AskUserQuestion-style machinery), not ride inside a report file that competes with everything else for attention.
3. Prompt-level: when an instruction explicitly overrules a constraint ("regardless of what the model thinks"), the agent should not reuse that constraint as a feasibility criterion; if honoring the instruction requires new mechanism, the correct output is a design proposal, not a "not safe" classification.

Happy to supply the dispatch report and ruling text if useful.