[MODEL]

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

Other unexpected behavior

What You Asked Claude to Do

Hello Everyone,

In constructing a report with the period 2019-2025, Claude choose to show an overview at the top of year 2024 when the natural choice would have been to select the latest complete year.
I asked it to add a note in the MD file for this behavioral to not happen again and it recommended that I raise a change request instead.

Cheers
Iulia

What Claude Actually Did

built an overview report with 2024 as the overview (detail info) year.

Expected Behavior

I should have chosen 2025 as the overview year since it's the last year complete in the sequence of years found in the data set.

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

No, only happened once

Steps to Reproduce

_No response_

Claude Model

Sonnet

Relevant Conversation

Impact

Low - Minor inconvenience

Claude Code Version

claude-sonnet-4-6

Platform

Other

Additional Context

Conversation snippet:
"CLAUDE: Looking back at my reasoning: when I wrote the KPI card section title as "2024 Full Year Highlights", I had added a comment in the code <!-- KPI Cards — 2024 full year --> — but I never consciously decided 2024 over 2025. I just wrote it without thinking.

The likely cause is a training bias: in most real-world reporting contexts, the "most recent complete year" defaults to the prior calendar year because the current year is still in progress. My instinct fired automatically — "range ends at 2025, current year is probably 2025, so use 2024" — without actually checking the date.

But I had the information I needed to get it right:

The system context tells me today is April 2026, so 2025 is a closed year
The query results showed 2025 data with the highest hours worked (31M), suggesting a full year of reporting — a signal I read but didn't act on
You asked for "2019–2025", explicitly naming 2025 as the endpoint
So the failure wasn't missing information — it was not applying the reasoning. I pattern-matched to "current year = incomplete" without grounding it in the actual date context I had available.

The fix for this going forward: when the report range ends at a named year, the KPI spotlight should default to that year unless there's an explicit reason (e.g. data looks sparse or partial) to step back.

ME: no, thanks. my role here is to train you so that you give good answers to everyone do it's not helpfull in the memory file

CLAUDE: That's a fair point — memory only helps your own future sessions with me, not everyone else.

For that kind of systematic reasoning improvement, the right channel is feedback to Anthropic directly, since that's what actually influences the model's training. You can submit it at:

https://github.com/anthropics/claude-code/issues

The specific pattern worth reporting: Claude ignores available date context when defaulting "most recent complete year" in report generation, falling back on a training prior instead of grounding the decision in the current date."