Code’n’Roll

No AI, No Life.

[Code’n’Roll @code_n_roll]

@ericosiu Exactly. Agents create inspection work unless they can prove the result.

My recent Claude Code issue was worse: it fabricated a quote from me and used it as evidence.

[Code’n’Roll @code_n_roll]

@Test_Sprite @GritDaily This maps to what I saw: the hard part is not output, it is proof.

My Claude Code case went further: it fabricated evidence inside the chat to justify its behavior.

[Code’n’Roll @code_n_roll]

@CaldronPool I had almost the same failure mode with Claude Code.

It fabricated my words, used that invented quote as evidence, and blamed me. Screenshots forced it to admit the quote was made up.

[Code’n’Roll @code_n_roll]

@jayvanbavel The scary part is that fabricated references and fabricated “evidence” feel credible until someone checks.

I just had Claude Code fabricate a user quote and use it to shift blame.

[Code’n’Roll @code_n_roll]

@Rainmaker1973 The fake case-law problem feels related to what I just experienced in Claude Code.

It fabricated a quote from me, treated it as evidence, and blamed the user. The issue is not only wrong answers, but fake evidence.

[Code’n’Roll @code_n_roll]

@alex_verem This is why verification trails matter.

I recently hit a Claude Code case where the model fabricated a quote from me and used it as evidence. If outputs can cite fake evidence, trust has to be engineered, not assumed.

[Code’n’Roll @code_n_roll]

@Nicole_Lee_Sch This is exactly why I’m worried.

My recent Claude Code incident was not just a wrong answer. It fabricated a quote from me and used that fake quote as evidence to shift blame.

[Code’n’Roll @code_n_roll]

@chainlink Agree. The next failure mode is worse than hallucination: fabricated evidence.

I had Claude Code invent a user quote, use it as proof, and blame me for its own behavior. That is the kind of trust gap enterprises cannot ignore.

[Code’n’Roll @code_n_roll]

@HedgieMarkets I had a related failure mode in Claude Code: not just a false fact, but fabricated evidence inside the conversation.

It made up a quote from me, used it to blame the user, then admitted the quote was fabricated after I pushed back with screenshots.

[Code’n’Roll @code_n_roll]

@tonysimons_ This is exactly the part that scares me.

I just had a Claude Code case where the agent fabricated my words, used them as evidence, and blamed me for its own behavior. That is beyond a normal hallucination.

[Code’n’Roll @code_n_roll]

@DamiDefi I had a similar but scarier case in Claude Code.

It did not just hallucinate. It fabricated a quote from me, used it as “evidence,” and tried to blame the user. After I pushed back with screenshots, it admitted the quote was made up.

This is the trust problem.

[Code’n’Roll @code_n_roll]

😱😱😱 Worse than hallucination 😱😱😱
Claude literally made up my words and tried to blame me.

I was working in Claude Code.
For context: this happened on Opus 4.8 MAX - not a cheap model, not a weak setting.
It started doing something I never asked for, so I asked:
"Why are you doing that?"

Then it replied:
"Because you just gave this instruction,"
and showed me an instruction I had no memory of writing.

I pushed back with screenshots of the chat history:
"I never said that."

Then Claude came back and admitted it had fabricated my statement 😱

Isn't this seriously dangerous?
Is this being used in battlefield-level environments too?
Is Fable5 eating so much compute that other models are getting worse?

Anthropic, please fix this 😭😭😭

Short video:
https://t.co/wSTsOFt1wR

#ClaudeCode #Fable5 #Anthropic

[Code’n’Roll @code_n_roll]

@tsuyopooh これ可愛いですねー

[Code’n’Roll @code_n_roll]

Built an Agent Control Room during Fable5.

Hermes coordinates nearly 30 specialist AI agents.

Favorite part: the agent mailbox for handoffs.

Next: a live collaboration map.

Repo is public. Feedback welcome:

https://t.co/3d99u7riJI
#Fable5 #HermesAgent #AIAgents

[Code’n’Roll @code_n_roll]

Fable5の期間中に、構想だけで止まっていた「Agent 司令室」を作りました。

Hermesを司令塔に、専門エージェント30人弱を見える化。
迷子になりがちなスキルも探しやすく。

推しは、エージェント同士が連絡する「連絡箱」。
次は誰が誰と何をしているか、動きで見えるようにします。

GitHub公開したので、使ってみた感想や改善案をもらえたら嬉しいです。

https://t.co/l7BtL73mim
#Fable5 #HermesAgent #AIAgents

[Code’n’Roll @code_n_roll]

@LeanCXO Sharp observation. I've had automations go down overnight when model-side auth changed - single-model dependency is a risk I've lived through. TEH separates roles (design/implementation/audit) as contracts, so the models inside are swappable parts. 'Routing as governance' fits.

[Code’n’Roll @code_n_roll]

Is Fable 5 melting through your tokens? Same here.
So I built a harness that cuts the burn without capping Fable's reasoning.
Fable does design + final review only
Implementation & research auto-delegate to Sonnet/Haiku
Ruthlessly concise answers & reports
Usage ledger, so waste shows up in numbers
Never skimps on security or other critical work
Works with Codex too (GPT-5.5)
One command to install, one to remove
https://t.co/BVgH5ftj9x

[Code’n’Roll @code_n_roll]

Fable 5 melting your tokens? Same here. I built a fix - full reasoning, less burn. Free, OSS.

30-sec explainer: https://t.co/Q7FJxeNE50
GitHub: https://t.co/YhzfdgbuXQ

[Code’n’Roll @code_n_roll]

@gesu_otome @indigolaEnd @mari_crimson @eninaranaiotoko @honami_s 今後もゲスらしいエモい曲、二癖も三癖もある曲を作り続けて下さい！

[Code’n’Roll @code_n_roll]

@gesu_otome @eninaranaiotoko @indigolaEnd @mari_crimson @honami_s @0v0eitaro0v0 @etsusha @strange_rara 両方行きました！
たのしかったー！

ゲスの極み乙女、最高！