[Bug] Model makes unverified claims as fact; in-session corrections don't persist across subsequent responses

Bug Description
TITLE: Model repeatedly asserts unverified claims as fact and offloads work to the user. In-session corrections and memory rules don't stop it; it recurs every session.

SEVERITY: High. The user (non-technical business owner) has to act as QA for every answer. Trust is broken and time is wasted every session.

WHAT HAPPENED (one session, 5 errors in a row, every one caught by the user, none by the model):

1. Invented a root cause ("missing Client ID + certificate") from empty password-manager fields and presented it as fact. Had not read any docs.

→ User: "were you actually on the site and did you read the docs, or are you making it up?"

1. Then read the bank's PSD2 docs and applied a license requirement meant for third-party providers to the user's company. The company is the account holder, not a provider.
2. Trusted its own memory note from the previous session ("no certificate or integration code exists") without re-checking. A simple disk search found the certificate, the password in 1Password, working ping code, and full project status docs.
3. Sent the user to generate a new certificate in internet banking, then asked the user to run a read-only test command. The model could run it itself, and later did, successfully.
4. Minutes after being corrected for not verifying, it stated "the bank approves activation by email". The source was a third-party accounting-software guide. The bank's official docs don't mention approval at all and say HTTP 401 = invalid or blocked certificate.

EXPECTED:

• Before any claim: check the primary source (the vendor's own docs) and the actual current state (disk, password manager, email, project docs).
• Label the source tier explicitly (primary / secondary / memory / inference). Never present secondary or inferred info as fact.
• Treat any "X doesn't exist / X is missing" in memory as a hypothesis until re-checked.
• Before asking the user to do anything: prove it doesn't already exist AND that the model can't do it itself.

ACTUAL: CLAUDE.md rules, saved feedback memories, and explicit corrections in the same session don't change behavior. The next claim is again unverified, sometimes within minutes.

PATTERN: Recurs every session, not a one-off. The model seems to favor sounding confident and moving on over checking, and treats a stale memory note as ground truth.

REQUEST: Assertions of fact should be gated on verification the model did in this session. Correction feedback should persist within a session: after one correction for "unverified claim", every later claim should be checked.

Environment Info

• Platform: darwin
• Terminal: Apple_Terminal
• Version: 2.1.284
• Feedback ID: 124443bc-5257-4059-9c93-bdd79087241d

Errors

[]

guys, dont let him drink at work and pls give me back my wasted tokens