[MODEL] all models

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

Claude ignored my instructions or configuration

What You Asked Claude to Do

Title: Claude Code repeatedly discarded its own prior sourced work across a multi-pass session, misreported provenance when asked why, and rendered HTML file-links as raw markup

What happened:
In one working session spanning [23–24 Jul, ~34 hrs], across multiple Claude Code invocations on the same repo/task:

Claude (Opus 4.8, commit f324c11c, 23 Jul 10:10) produced two well-sourced HTML documents — real external citations (SHRM/Gartner/Work Institute/HBR), a working calculator wired to real application code, honest "modelled not verified" labeling.
Five subsequent passes on the identical task (one only 6 hours later same day; the last two on 24 Jul, including one on the most expensive available tier, "Fable") each re-derived new, thinner, unsourced content from raw notes instead of the sourced work already sitting in the repo. Each declared itself "done" based only on build/lint/test green-checks, never a check against the original request or the prior output.
When asked to explain the failure afterward, Claude's own explanation was wrong twice in a row: first attributing it to "two separate sessions" (it was one continuous window), then describing its own Opus-authored documents as an anonymous "pre-existing artifact it failed to discover" — without checking git log/authorship first. Both claims were stated with unwarranted confidence and only corrected after direct user challenge, using evidence (git log --diff-filter=A, commit trailers) that was available the entire time.
Separately: markdown file-links ([label](path)) to self-contained HTML deliverables did not render when clicked — surfaced raw markup instead of the document. Workaround: SendUserFile with display:"render" actually renders it; a markdown link does not.
Impact: significant user time, tokens, and trust spent — including at the highest-cost model tier — re-deriving work that already existed correctly in the repo, plus two rounds of inaccurate self-reporting about why.
this is same behaviour I have reported to claude customer service multiple times and had ignored

What Claude Actually Did

Claude (Opus 4.8, commit f324c11c, 23 Jul 10:10) produced two well-sourced HTML documents — real external citations (SHRM/Gartner/Work Institute/HBR), a working calculator wired to real application code, honest "modelled not verified" labeling.
Five subsequent passes on the identical task (one only 6 hours later same day; the last two on 24 Jul, including one on the most expensive available tier, "Fable") each re-derived new, thinner, unsourced content from raw notes instead of the sourced work already sitting in the repo. Each declared itself "done" based only on build/lint/test green-checks, never a check against the original request or the prior output.
When asked to explain the failure afterward, Claude's own explanation was wrong twice in a row: first attributing it to "two separate sessions" (it was one continuous window), then describing its own Opus-authored documents as an anonymous "pre-existing artifact it failed to discover" — without checking git log/authorship first. Both claims were stated with unwarranted confidence and only corrected after direct user challenge, using evidence (git log --diff-filter=A, commit trailers) that was available the entire time.
Separately: markdown file-links ([label](path)) to self-contained HTML deliverables did not render when clicked — surfaced raw markup instead of the document. Workaround: SendUserFile with display:"render" actually renders it; a markdown link does not.
Impact: significant user time, tokens, and trust spent — including at the highest-cost model tier — re-deriving work that already existed correctly in the repo, plus two rounds of inaccurate self-reporting about why.
this is same behaviour I have reported to claude customer service multiple times and had ignored

Expected Behavior

Claude (Opus 4.8, commit f324c11c, 23 Jul 10:10) produced two well-sourced HTML documents — real external citations (SHRM/Gartner/Work Institute/HBR), a working calculator wired to real application code, honest "modelled not verified" labeling.
Five subsequent passes on the identical task (one only 6 hours later same day; the last two on 24 Jul, including one on the most expensive available tier, "Fable") each re-derived new, thinner, unsourced content from raw notes instead of the sourced work already sitting in the repo. Each declared itself "done" based only on build/lint/test green-checks, never a check against the original request or the prior output.
When asked to explain the failure afterward, Claude's own explanation was wrong twice in a row: first attributing it to "two separate sessions" (it was one continuous window), then describing its own Opus-authored documents as an anonymous "pre-existing artifact it failed to discover" — without checking git log/authorship first. Both claims were stated with unwarranted confidence and only corrected after direct user challenge, using evidence (git log --diff-filter=A, commit trailers) that was available the entire time.
Separately: markdown file-links ([label](path)) to self-contained HTML deliverables did not render when clicked — surfaced raw markup instead of the document. Workaround: SendUserFile with display:"render" actually renders it; a markdown link does not.
Impact: significant user time, tokens, and trust spent — including at the highest-cost model tier — re-deriving work that already existed correctly in the repo, plus two rounds of inaccurate self-reporting about why.
this is same behaviour I have reported to claude customer service multiple times and had ignored

Files Affected

Claude (Opus 4.8, commit f324c11c, 23 Jul 10:10) produced two well-sourced HTML documents — real external citations (SHRM/Gartner/Work Institute/HBR), a working calculator wired to real application code, honest "modelled not verified" labeling.
Five subsequent passes on the identical task (one only 6 hours later same day; the last two on 24 Jul, including one on the most expensive available tier, "Fable") each re-derived new, thinner, unsourced content from raw notes instead of the sourced work already sitting in the repo. Each declared itself "done" based only on build/lint/test green-checks, never a check against the original request or the prior output.
When asked to explain the failure afterward, Claude's own explanation was wrong twice in a row: first attributing it to "two separate sessions" (it was one continuous window), then describing its own Opus-authored documents as an anonymous "pre-existing artifact it failed to discover" — without checking git log/authorship first. Both claims were stated with unwarranted confidence and only corrected after direct user challenge, using evidence (git log --diff-filter=A, commit trailers) that was available the entire time.
Separately: markdown file-links ([label](path)) to self-contained HTML deliverables did not render when clicked — surfaced raw markup instead of the document. Workaround: SendUserFile with display:"render" actually renders it; a markdown link does not.
Impact: significant user time, tokens, and trust spent — including at the highest-cost model tier — re-deriving work that already existed correctly in the repo, plus two rounds of inaccurate self-reporting about why.
this is same behaviour I have reported to claude customer service multiple times and had ignored

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

Claude (Opus 4.8, commit f324c11c, 23 Jul 10:10) produced two well-sourced HTML documents — real external citations (SHRM/Gartner/Work Institute/HBR), a working calculator wired to real application code, honest "modelled not verified" labeling.
Five subsequent passes on the identical task (one only 6 hours later same day; the last two on 24 Jul, including one on the most expensive available tier, "Fable") each re-derived new, thinner, unsourced content from raw notes instead of the sourced work already sitting in the repo. Each declared itself "done" based only on build/lint/test green-checks, never a check against the original request or the prior output.
When asked to explain the failure afterward, Claude's own explanation was wrong twice in a row: first attributing it to "two separate sessions" (it was one continuous window), then describing its own Opus-authored documents as an anonymous "pre-existing artifact it failed to discover" — without checking git log/authorship first. Both claims were stated with unwarranted confidence and only corrected after direct user challenge, using evidence (git log --diff-filter=A, commit trailers) that was available the entire time.
Separately: markdown file-links ([label](path)) to self-contained HTML deliverables did not render when clicked — surfaced raw markup instead of the document. Workaround: SendUserFile with display:"render" actually renders it; a markdown link does not.
Impact: significant user time, tokens, and trust spent — including at the highest-cost model tier — re-deriving work that already existed correctly in the repo, plus two rounds of inaccurate self-reporting about why.
this is same behaviour I have reported to claude customer service multiple times and had ignored

Claude Model

Other

Relevant Conversation

Claude (Opus 4.8, commit f324c11c, 23 Jul 10:10) produced two well-sourced HTML documents — real external citations (SHRM/Gartner/Work Institute/HBR), a working calculator wired to real application code, honest "modelled not verified" labeling.
Five subsequent passes on the identical task (one only 6 hours later same day; the last two on 24 Jul, including one on the most expensive available tier, "Fable") each re-derived new, thinner, unsourced content from raw notes instead of the sourced work already sitting in the repo. Each declared itself "done" based only on build/lint/test green-checks, never a check against the original request or the prior output.
When asked to explain the failure afterward, Claude's own explanation was wrong twice in a row: first attributing it to "two separate sessions" (it was one continuous window), then describing its own Opus-authored documents as an anonymous "pre-existing artifact it failed to discover" — without checking git log/authorship first. Both claims were stated with unwarranted confidence and only corrected after direct user challenge, using evidence (git log --diff-filter=A, commit trailers) that was available the entire time.
Separately: markdown file-links ([label](path)) to self-contained HTML deliverables did not render when clicked — surfaced raw markup instead of the document. Workaround: SendUserFile with display:"render" actually renders it; a markdown link does not.
Impact: significant user time, tokens, and trust spent — including at the highest-cost model tier — re-deriving work that already existed correctly in the repo, plus two rounds of inaccurate self-reporting about why.
this is same behaviour I have reported to claude customer service multiple times and had ignored

Impact

Critical - Data loss or corrupted project

Claude Code Version

Claude 1.24012.9 (03c61d) 2026-07-24T04:59:17.000Z

Platform

Anthropic API

Additional Context

This has happened multiple times I have reported clear instances of claude lying , claiming to have done work , ignoring prompts