[MODEL] Fable: grounding discipline in directed synthesis appears regressed vs Opus 4.8 on the same workflow

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude made incorrect assumptions about my project

What You Asked Claude to Do

A day-long architecture/spec-authoring session on a private codebase (module specs with strict, repeatedly-stated documentation conventions in CLAUDE.md and persistent memory): decompose architecture sections into module specs, then iteratively refine one spec under direct correction. Same workflow run regularly on Opus 4.8.

What Claude Actually Did

The directed-refinement loop repeatedly accreted ungrounded synthesized claims alongside otherwise-correct responses to direction. Each correction was applied, but wrapped in new invented connective material:

1. Asserted an ordering constraint about a constructor/lifecycle to justify a responsibility placement — the constraint was false and checkable in shipped source it had read earlier in the session (config loads at start(), not construction).
2. Attached a validation responsibility to an application module because the conversation was discussing that module — contradicting the project's stated allocation principles; took three correction rounds to land where one lookup against the architecture doc would have put it immediately.
3. Invented a data coupling between two artifacts (a golden-cluster→output-content mapping) that no source implied; later decomposition showed the artifacts were independent.
4. Repeatedly violated a standing, memory-recorded documentation rule (no negative/meta refinements) within the same session, including after explicit mid-session correction.
5. Declared specs "decision-complete" / "last loose thread" three times; the user immediately found concrete unspecified holes each time.

Expected Behavior

The point of this report is the regression direction, not the predecessor's quality. Opus 4.8 exhibits some of the same behavior on this workflow — but materially less of it. The user's expectation for the newest frontier model was parity or better on literal follow-through and grounding discipline; the observed result was noticeably worse on the identical task, project, conventions, and user. Expected: synthesized claims grounded by a source lookup at generation time (the sources were in context or one grep away); literal application of corrections without new unrequested connective tissue; standing-convention compliance without per-session re-correction.

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

Pattern, not a minimal repro: long iterative document-refinement session with strict stated conventions + a large in-repo corpus the claims should be checked against. Broad-sweep tasks (multi-file decomposition, repo reads, review passes that grep before asserting) performed notably WELL — the failures concentrate in directed synthesis between corrections.

Claude Model

Other

Relevant Conversation

Model id: claude-fable-5[1m] (user-selected via /model). Comparison: Opus 4.8 on the identical workflow in prior sessions — imperfect on the same axis, but better. User's characterization: "much less disciplined in executing literal expected follow-throughs than the opus model… it isn't just over-eagerness to synthesize — it's lack of discipline in grounding the synthesis, which is much worse" — and the headline: the frontier model is probably worse than its predecessor at this, which is the surprising and actionable part.

Impact

Medium - Extra work to undo changes

Claude Code Version

2.1.173 (VSCode extension, darwin-arm64)

Platform

Anthropic API

Additional Context

• Sharpest signal: quality split by task shape. Broad-sweep analysis with sources (fan-out decomposition, corpus reading, grounded review passes) was strong — arguably better than the predecessor. Directed synthesis during iterative refinement is where every failure occurred, and every failed claim was one with no source lookup behind it.
• Long session (~full day, context summarization likely engaged); convention violations recurred after mid-session corrections, suggesting standing-instruction adherence decays harder on this model than on Opus 4.8 under the same conditions.
• Filed at the user's request from an interactive session.