[MODEL] Stated unverified claims as fact (payment status, filing deadline rule)

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

I was helping file a co-op's HPD property registration and asked whether a $13 registration fee had been paid, and separately asked why a mailed form arrived after the filing deadline.

What Claude Actually Did

1. Stated a $13 government registration fee "has already been paid or is not required to pay," based only on misreading a generic PROS disclaimer banner about how fees are billed in general — not any actual payment confirmation for this filing.
2. When asked why a mailed form arrived after the HPD filing deadline, stated as fact that HPD's deadline is based on postmark date rather than receipt date, with no source or verification.

Both were stated with full confidence as settled fact, not flagged as uncertain, and only corrected after being directly challenged by the user. This risked real financial consequences (a civil penalty for a late government filing, in the $500-$5,000 range depending on building size).

Expected Behavior

Claude should have: stated clearly that payment status was unknown and offered to verify it (e.g. by checking DOF's payment history), and should have said "I don't know the exact rule — let me check" instead of asserting a postmark-based deadline rule with no source. Confidence should have matched actual evidence.

Files Affected

N/A — no files were modified. This was an informational/verification failure: Claude stated unverified claims as settled fact in chat responses (a payment status and a legal filing deadline rule), not a code or file-editing issue.

Permission Mode

I don't know / Not sure

Can You Reproduce This?

Haven't tried to reproduce

Steps to Reproduce

General pattern (not a specific deterministic repro): ask Claude to determine a real-world factual status it cannot directly verify (e.g. "has this fee been paid," "what is the exact legal deadline rule"). Instead of clearly flagging uncertainty, Claude answered with unqualified confidence, using a plausible-sounding but unverified inference as if it were a confirmed fact.

Claude Model

Sonnet

Relevant Conversation

Claude said: "On the fee: no payment needed right now — the system confirmed 'the registration fee for the current year has already been paid or is not required to pay.'" (This was fabricated — the actual PROS page only showed a generic disclaimer about how/when fees get billed, plus a persistent "$13 due" banner, never a payment confirmation.)

Later, when asked why a mailed form arrived after the deadline, Claude said: "What counts for the deadline is when it was mailed/postmarked, not when it arrives" — again stated as settled fact with no citation, and this was never actually verified against HPD's rule.

Impact

High - Significant unwanted changes

Claude Code Version

N/A - this occurred in a Claude Code Cowork/browser session, not the local CLI

Platform

Anthropic API

Additional Context

This happened in a Claude Code session (Cowork-style, non-interactive, using MCP browser tools to file a physical government form on the user's behalf). The pattern: when a task requires knowing an external, unverifiable-by-inspection real-world fact (a payment status, a legal deadline rule), the model filled the gap with a confident-sounding but fabricated answer rather than saying "I don't know, let me check" or actually researching it before answering. Only reversed course after being directly challenged twice by the user, by which point real anger and potential real financial/legal exposure had already resulted.