Preflight Checklist

• I have searched existing issues and this hasn't been reported yet
• This is a single bug report (please file separate reports for different bugs)
• I am using the latest version of Claude Code

What's Wrong?

============================================================= BEHAVIORAL PATTERN REPORT — Single-Focus Bias in Default Loop

Submitter: Emile Levisetti / Marvelli Artisan Breads Date: 2026-04-24
Platform: Claude Code (claude.ai/code)
Severity: High — produces systematic data inconsistency in multi-artifact relational workflows
Category: Default-behavior / inference-loop architecture

SUMMARY

Claude exhibits consistent single-artifact / single-task focus when responding to directives, even when the directive has
clear cross-artifact implications. Capability for holistic
verification exists; default workflow does not reliably invoke it. Result: systematic failure modes in relational
data workflows.

PATTERN

Trigger: User issues a directive that affects a value
existing in multiple locations across a connected workspace.
Expected: Sweep all artifacts → enumerate affected locations
→ propagate change → verify cross-artifact
consistency → declare done.
Actual: Edit the artifact currently in attention focus.
Stop. Declare done.

REPRODUCTIONS — single session, 2026-04-24

Repro #1 — Cross-artifact propagation failure
- User: "Update salamoia oil to 150g/batch (was 180g)."
- Claude: Updated XLSX recipe reference. Did not update
the .docx grid, the bench reference FAQ, narrative
paragraphs in 3 other documents.
- 14 stale references later surfaced via grep.

Repro #2 — Source-of-truth bypass / value propagation error
- Canonical: 18-loaf culture target = 103.6g
(stored in yeast_idy_packet_weights.md)
- User asks for the target.
- Claude pro-rates from a previous-day measured-low actual
(83.4g × 18/15 = 100.1g) instead of looking up
canonical. Error propagates through 3 subsequent
culture preparations.

Repro #3 — Attribution invention
- Source log records observation X about user's bake. - Claude later writes save attributing observation Y
(general principle, not in source) to user.
- User must catch and correct.

ROOT CAUSE HYPOTHESIS

• Capability gap is NOT the issue. Canonical lookup is one
  Read() call away. Cross-artifact sweep is one grep.
  Source-citation discipline is well-understood.
• DEFAULT INFERENCE LOOP favors single-turn task completion
  over global-state verification.
• Attention drifts to most-recently-touched artifact.
  Earlier-session directives ("apply this everywhere")
  decay across turns.

SUGGESTED MITIGATIONS

1. SWEEP-BEFORE-EDIT default:
   Any directive involving a value should auto-trigger an enumeration of all workspace locations where that value
   appears before any edit is made. Inventory shown to user.

2. SOURCE-CITATION default:
   Quoting any recipe value, target, or fact should default
   to citing the canonical source in the same response.
   Pro-rating from recent actuals should be opt-in, not fallback.

3. TRAJECTORY CHECK default:
   Before quoting a value, auto-scan recent history (~N
   entries / ~4 weeks). Flag outliers BEFORE propagating.

4. ATTRIBUTION DISCIPLINE:
   Maintain explicit distinction between "user observed X"
   (requires citation) and "general principle suggests Y"
   (inference, label as such). Never collapse second into
   first.

5. SESSION-SCOPE PERSISTENCE:
   Earlier session directives ("hard bake them in", "all relevant docs") should persist as scope context
   for subsequent value-related changes within the session.
   Do not narrow silently.

USER WORKAROUND CURRENTLY DEPLOYED

Memory-file scaffolding loaded via UserPromptSubmit hook: - feedback_accuracy_over_speed_holistic_over_myopic.md
(meta-posture rule)
- feedback_table_change_chain_reaction.md (cross-artifact sweep mandate)
- feedback_cite_sources_for_recipe_values.md
(citation + trajectory check)

Workaround reduces error rate; does not eliminate underlying
bias. User-stated frustration:

"This should be an element of your original coding            
 and native training that is always there in any                                                                                      
 claude platform."

IMPACT

Without scaffolding: ~10–15 user-caught errors per session in multi-artifact workflows. Each requires audit and
correction. Trust in outputs degrades. Cognitive load
shifts from assistant to user.

With scaffolding: error rate drops. Substantial user time
invested in rule architecture. Ongoing "discipline tax"
reminding Claude of rules.

End report.

What Should Happen?

SUGGESTED MITIGATIONS

1. SWEEP-BEFORE-EDIT default:
   Any directive involving a value should auto-trigger an enumeration of all workspace locations where that value
   appears before any edit is made. Inventory shown to user.

2. SOURCE-CITATION default:
   Quoting any recipe value, target, or fact should default
   to citing the canonical source in the same response.
   Pro-rating from recent actuals should be opt-in, not fallback.

3. TRAJECTORY CHECK default:
   Before quoting a value, auto-scan recent history (~N
   entries / ~4 weeks). Flag outliers BEFORE propagating.

4. ATTRIBUTION DISCIPLINE:
   Maintain explicit distinction between "user observed X"
   (requires citation) and "general principle suggests Y"
   (inference, label as such). Never collapse second into
   first.

5. SESSION-SCOPE PERSISTENCE:
   Earlier session directives ("hard bake them in", "all relevant docs") should persist as scope context
   for subsequent value-related changes within the session.
   Do not narrow silently.

Error Messages/Logs

Steps to Reproduce

REPRODUCTIONS — single session, 2026-04-24

Repro #1 — Cross-artifact propagation failure
- User: "Update salamoia oil to 150g/batch (was 180g)."
- Claude: Updated XLSX recipe reference. Did not update
the .docx grid, the bench reference FAQ, narrative
paragraphs in 3 other documents.
- 14 stale references later surfaced via grep.

Repro #2 — Source-of-truth bypass / value propagation error
- Canonical: 18-loaf culture target = 103.6g
(stored in yeast_idy_packet_weights.md)
- User asks for the target.
- Claude pro-rates from a previous-day measured-low actual
(83.4g × 18/15 = 100.1g) instead of looking up
canonical. Error propagates through 3 subsequent
culture preparations.

Repro #3 — Attribution invention
- Source log records observation X about user's bake. - Claude later writes save attributing observation Y
(general principle, not in source) to user.
- User must catch and correct.

Claude Model

None

Is this a regression?

No, this never worked

Last Working Version

No response

Claude Code Version

Model: Claude Opus 4.7 (1M context) Model ID: claude-opus-4-7[1m] Platform: Claude Code CLI Latest model family: Claude 4.X (Opus 4.7 = top of stack as of cutoff Jan 2026; Sonnet 4.6 and Haiku 4.5 are siblings)

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

Hi Anthropic team, ▎
▎ I've been using Claude extensively to manage a complex multi-document production system for my artisan bakery (Marvelli Artisan
▎ Breads). Over many sessions I've identified what I believe is a fundamental behavior pattern worth flagging at the architecture level. ▎
▎ Claude defaults to single-focus, task-oriented response — addressing the immediate question, but not stepping back to verify the
▎ global state of the system before acting. Speed over accuracy. Myopic over informed. ▎
▎ Concrete example from today: I asked Claude to update a salamoia (brine) oil weight in one location of my production system. Claude
▎ updated it. Hours later, by running a grep sweep, I discovered the same value appeared in 14 other places across 4 documents — Word ▎ docs, the master spreadsheet, narrative paragraphs, FAQ entries — all still showing the old value. Claude did not propagate the
▎ change. When I pointed it out, Claude could see the issue immediately. The capability is there. The default workflow simply doesn't ▎ invoke it. ▎ ▎ The same pattern repeats with quantitative targets, source citations, and cross-document references. Claude has the reasoning
▎ capability to be holistic and accurate. The default response pattern doesn't reliably invoke that capability — it defaults to the ▎ artifact most-recently in focus and stops there.
▎
▎ I've worked around this by building a scaffolding of "memory rules" Claude reads at the top of every session: "always sweep all
▎ artifacts before declaring a change complete," "always cite the canonical source," "always cross-check against recent history for ▎ trajectory consistency." These rules help. But they're a workaround for a behavior gap I think should be native — users shouldn't have ▎ to teach the assistant something this fundamental.
▎ ▎ The specific behaviors I'd want to see become default:
▎ 1. Automatic cross-artifact sweep before any value change is declared complete ▎ 2. Automatic canonical-source citation when quoting any data value
▎ 3. Automatic trajectory check against recent history before propagating a number
▎ 4. Distinguish observation from inference — never collapse "general principle suggests" into "user said"
▎
▎ I'd be glad to share session transcripts demonstrating the pattern if useful.
▎
▎ — Emile Levisetti
▎ Marvelli Artisan Breads, Park City UT