Fable 5 at max/xhigh effort: polished but ungrounded outputs; effort deepens reasoning inside the chosen frame instead of expanding ground-truth gathering

Environment

• Claude Code CLI, macOS (darwin 25.5)
• Model: claude-fable-5, effort switched between xhigh and max via /model during the session

What happened

Across one long working session (content/design restructuring in a real repo), the model repeatedly produced complete-looking deliverables (analysis documents, HTML mockups, a handoff prompt) that were synthesized only from material already in its context. When a deliverable depended on implementation files it had not read, it filled the gap with plausible invention instead of reading them, even though the files were small, adjacent, and directly named the ground truth.

Three concrete instances from the session (anonymized):

1. Produced a merge/consolidation framework over ~114 content entries with confident coverage claims ("adds nothing the host doesn't say") that its own follow-up subagent review later showed to be wrong in specifics.
2. Built a product mockup that invented a six-section document layout and a selection UI, when the app's real 15-section document definition sat in a constants file the model had never opened. The user called the result "a cheap mockup."
3. Drafted a handoff prompt for a fresh session whose reading list was simply the files the model happened to have read that day; it omitted the entire implementation surface the task depended on. It took three successive user corrections ("you're so lazy") before the prompt included an actual implementation-review phase.

Why this seems effort-related and worth reporting

Raising effort to xhigh/max did not change the failure mode. The session's own post-mortem conclusion, which the user agreed with: effort appears to scale deliberation within the already-chosen frame and inputs; it does not seem to increase the model's propensity to go gather the inputs the frame should have been built on. High output polish then masks the thin grounding, making the failure harder to catch. A secondary pattern: each user correction was treated as a replacement frame rather than an added constraint, producing overcorrection whiplash across turns.

Expected behavior

At high effort tiers especially, the model should bias toward reading the implementation its deliverable depends on before synthesizing, or explicitly flag unread dependencies as gaps, rather than filling them with confident, plausible invention.

Impact

User paid for max effort across a full working day and characterized multiple successive deliverables as "lazy" despite the setting. The gap between billed effort and input-gathering diligence is the core of the report.