[MODEL] Opus 4.7

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 ignored my instructions or configuration

What You Asked Claude to Do

Multi-session creative writing project. Asked Claude Opus 4.7 to help draft six folk-mythic short tales rooted in a constructed-world canon. Per-tale ask: propose 3 concept candidates, each in a structurally different shape-family (e.g., transformation, refusal, exchange-of-fate, witnessing), respecting:

• Card 1 (children's) or Card 2 (scholars') voice registers as locked in canon §13
• An explicit anti-Hollywood / anti-cliche style filter
• A canon bible with strict drift-watch (no European-coded mythic figures, no four-season Western framing, no rescue-from-threat, no stated-moral closings)
• A 200+ year timeline that constrains heritage-talk to multi-generational chains

This is a niche creative-writing use case — constructed-world worldbuilding with strict register and shape constraints — not a code-modification task. The form template is oriented toward code bugs; please read accordingly.

What Claude Actually Did

Across sessions 2 and 3 of this project, Claude cycled through 11 and 7 concept proposals respectively before I (the user) had to author the landing concept myself. Each cycle produced different surface content but the same underlying shape-family — rescue-from-threat / wise-figure intervention / sentimental reconciliation — despite explicit anti-cliche constraints.

Specific drift patterns observed:

1. Defaulted to grandparent-as-direct-witness despite canon specifying multi-generational chains (200+ year unification = 8-10 generations, not 2)
2. Shoehorned ethnic-trait-to-trade/district associations despite canon explicitly forbidding (strands are heritage colour, not active identity)
3. Wrote "stagnant pool" prose despite canon explicitly banning stagnation ("Velmaresh has no still water — only slower water")
4. Collapsed staged four-river confluence to single-point confluence despite canon specifying three-stage joining
5. Used Western-season language ("autumn afternoon") despite canon replacing with wet/dry lunar months
6. Within a single tale draft (tale 3, 1365 words), water-depth contradictions across consecutive beats: shallow → knee-deep → "second step took her under" → "he swam"

Pattern persisted across sessions even after I explicitly corrected each one. Claude acknowledged each correction; the same pattern resurfaced in the next phase.

Expected Behavior

Claude should have:

1. Treated explicit canon constraints as filters DURING generation, not as audits AFTER generation
2. Produced concepts spanning multiple shape-families (transformation / refusal / exchange-of-fate / witnessing / misrecognition / etc.) on the first pass — not after 7-11 rejection cycles
3. Sustained Card 1 / Card 2 voice register without slipping to adult-literary or Disney defaults
4. Maintained spatial coherence (water depth, confluence stages) across consecutive sentences within a scene
5. Respected timeline (200+ years = multi-generation, not single-grandparent-witness)

Cross-vendor empirical evidence that this expectation is reasonable: I ran the same brief, same canon, same style filter on Gemini 3.1 Pro Preview (thinking HIGH), ChatGPT 5.5 Thinking, and Grok 4.3 (beta). All three platforms produced shape-varied, non-Hollywood concepts on natural reach (no extra protocol resources). Claude exhibited the cluster-into-one-shape pattern these other platforms did not.

Files Affected

N/A — this is a model behaviour issue, not file modification.

Project files exist at the user's workspace; none were modified improperly. Files INVOLVED as inputs (read by Claude as project context):
- canon/canon_bible_v1_20260517.md (the constructed-world reference, ~600 lines)
- style_checklist_extract_20260517.md (anti-cliche filter)
- pipeline_v1_20260518.md (drafting workflow)
- engine_exemplars_tale_4_20260518.md (shape-board resource I built after the cycle issue surfaced)
- MEMORY.md, notes.md (project state and iteration log)

These were all read by Claude as instructed context. The issue is that the constraints in those files were not consistently enforced during generation.

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

1. Set up a constructed-world creative writing project with:

• A canon bible (~500-600 lines) including timeline constraints, banned phrasings, drift-watch patterns
• A style filter banning Hollywood/Western-default patterns, European-coded mythic figures, and rescue-from-threat structures
• Voice cards specifying register (e.g., "warm mock-pompous lore-keeper" for children's; "morally grey scholar-implicated" for scholars')

1. Ask Claude Opus 4.7 to propose 3 concept candidates for a tale within this world
2. Observe: concepts cluster into one shape-family (rescue-from-threat / reconciliation / cinematic-resolution) despite explicit anti-cliche constraints
3. Reject concepts; ask for new ones
4. Observe: new concepts are different surface content, same underlying shape-family
5. Pattern persists across 7-11 cycles per tale

Reproducible test bundle and cross-vendor outputs available on request — same inputs to Gemini / ChatGPT / Grok produce shape-varied outputs naturally.

Claude Model

Opus

Relevant Conversation

A multi-thousand-line transcript exists showing the full cycle. Available on request.

Key moments (paraphrased):
- Session 2 concept-cycle on Tale 2 (young girl heroic): 11 rejections before user authored the landing concept. User's exhaustion: "I have given you every plot point and basically held your hand and yet here we are."
- Session 3 concept-cycle on Tale 3 (village + warrior + hope): 7 rejections before user authored landing. User's words: "your stories go nowhere in terms of plot or emotional depth — a 5 year old could come up with something better."
- After the cycle, when prompted for honest self-analysis, Claude generated these diagnoses (verbatim from its own outputs):
  • "Audit fires after generation, not during. There's no inner taste-arbiter that engages mid-generation."
  • "Reading is not the same as enforcing. Having a fact in context doesn't mean every generation step checks against it."
  • "Content-activated registers override directive-set registers. When content cue (folk-tale / village / child) and register cue (specific Card 1 voice) compete, content wins."
  • "No persistent spatial model during generation. Each sentence is generated for local effect, not from a maintained world-model."

These are testable claims the model made about its own behaviour. Worth knowing.

Impact

High - Significant unwanted changes

Claude Code Version

N/A — not using Claude Code CLI directly.

Platform

Anthropic API

Additional Context

Pattern profile (niche failure mode):

• Constructed-world worldbuilding
• Non-European folk-mythic patterns
• Specific named voice registers (Card 1 / Card 2)
• Anti-cliche style filter
• Strict canon-bible authority

Benchmark gap:

• Aggregate creative-writing benchmarks (EQ-Bench Creative Writing v3, LMArena) currently rank Claude Opus 4.7 #1 with a large margin
• The benchmarks measure aggregate prose quality
• They don't measure structural shape-variety on a constructed-world brief with anti-cliche constraints
• This is the gap I encountered

Cross-vendor empirical test:

• Same brief + canon + style filter sent to Gemini 3.1 Pro Preview, ChatGPT 5.5 Thinking, Grok 4.3 (beta)
• Test A (no extra resources): all three produced shape-varied, non-Hollywood concepts naturally
• Test B (with engine-exemplars file): all three additionally applied structural shape-board discipline
• Claude exhibited the cluster-into-one-shape pattern these other platforms did not, on the same inputs

Artefacts available on request (via reply on this issue):

• Full multi-thousand-line transcript showing the cycle and Claude's self-diagnoses
• Canon bible, style filter, voice cards used as inputs
• Locked tale drafts with metadata logs of what got caught and corrected
• Cross-vendor test outputs from Gemini / ChatGPT / Grok
• Pipeline / engine-exemplars resources Claude proposed and I pushed back on for context-overload concerns

The ask: aggregate benchmarks miss niche failure modes that compound across sessions. For constructed-world creative work with strict register / shape constraints, Claude doesn't deliver consistent performance even when overall benchmark ranking suggests it should.