Opus 4.6: fails to synthesize facts established in same conversation

Description

When using Claude Opus 4.6 (1M context) in Claude Code, the model repeatedly fails to combine two or more facts that were established earlier in the same conversation to derive an obvious conclusion. The context window is not near capacity — these are facts from minutes or hours ago in the same session.

Concrete examples from a single session (April 15, 2026)

Example 1: RSIN derivation

• Fact A (established via web research): Dutch VAT number structure is NL + RSIN + B + suffix
• Fact B (established via VIES verification): Fiscal unity VAT number is NL828087970B02
• Expected: Model derives RSIN = 828087970 immediately
• Actual: Model told user to "ask your accountant for the RSIN" — multiple turns later, user pointed out the derivation themselves

Example 2: Payslip already verified

• Fact A (established earlier in session): April payslip verified — gross €4,833.33, net €3,264.47, loonheffing €1,247.92
• Fact B (during operational scan): Gmail shows April payslip email as "UNREAD"
• Expected: Model notes the payslip is already in the data layer, no action needed
• Actual: Model flagged it as "worth checking" — user had to remind the model that we literally just verified these numbers

Example 3: Email thread references

• Model drafted an email referencing the wrong email threads, paraphrasing from memory instead of reading the actual thread content
• When corrected and asked to read the threads, the model found the correct references but had not done so proactively

Pattern

The model treats each sub-question as isolated, answering from the nearest available information rather than cross-referencing everything established in the session. The user described this as "path of least resistance" — locally correct answers that miss the holistic picture.

This is not a context window issue (session was well within limits). It appears to be a synthesis/retrieval issue where the model doesn't proactively scan its own prior outputs when formulating a new response.

Impact

In an AI-assisted workflow where the model is expected to accumulate knowledge across a session and reason across domains, this failure mode forces the user to manually re-state facts and catch missed connections. It undermines trust and significantly reduces productivity.

Environment

• Model: Claude Opus 4.6 (1M context), max effort
• Tool: Claude Code CLI
• Session type: Multi-hour operational session with domain knowledge, email drafting, tax research, and data verification