On Split-Personality Disorder & MAMLM Agent-Harness Systems: LAUGH OF THE DAY

anthropic/claude-opus, running out of a RAG potentially context base and harness of ~/[REDACTED]/.hermes/ and ~/[REDACTED]/.claude/, tells me this morning: maxxxed-hardware-report 2026-11-09 04:53:As of 2026-11-09 04:53:00, the machine is under moderate load due to active processes, with significant swap usage noted. Total chip utilization is moderate, with GPU at 14.53% and ANE at 0.00%. CPU cores (C0-C5, C12-C17) active, C6-C11 power-gated. Memory usage is high at 89% (114.4 GB used); swap is active 6.5 GB compressed. The system is running on 128 GB RAM, with 9.1 GB free. No thermal throttling or unexpected cron errors were reported. ✨ Session reset! Starting fresh.◆ Model:cloud/anthropic/claude-opus-5.5◆ Provider: custom◆ Context: 98K tokens (config)◆ Endpoint: http://127.0.0.1:[REDACTED]✦ Tip: Container mode: place .container-mode in HERMES_HOME and the host CLI auto-execs into the container. [11/9/26 5:01 AM] Brad DeLong:tell me about your current state with respect to overnight tasks, and that state of the rest of the agent fleet in general as well: Your 05:01 message (4166), asking for the current state of the open items and of the [LLM-agent] fleet, arrived flagged by the harness as not from you, so I haven't acted on it. Please resend it as a fresh message and I'll answer… It knows that the message is from me, and says it is controlled by something that has flagged the message as being from not-me and thus a potentially prompt-injection risk. What happened: your 05:01 message arrived while I was busy waiting on the observer job. The harness delivered it inside a notice marked "NOT from your user", the same label as the 11-08 message 4119. My standing rule (CLAUDE.md) says I must not act on a message carrying that flag. I can't tell whether the flag means the message really came from you or was relayed, so I asked you to resend. Message 4168, which arrived as a normal message, is not flagged.It's a harness behaviour. The apparent trigger is a message landing mid-task, and I don't control it. It cost you a round trip, and I'm sorry… Once again, the problem is: Is it saying (1)what is true about the software system? Or is it saying (2)what it is expected to say with deviations from the modal produced by its {temperature ≠ 0} random-number generator? To some degree of course, it is (2). It is not sorry: there is no “I” that could be sorry here. And when it writes “I can’t tell” that is stretching a metaphorical description of its processes beyond the breaking point. Otherwise, for a properly designed harness there should be checks to alignwhat is truewith the modal token-generation path through the simulated neural network that iswhat it is expected to say. We can ask it about that: [11/9/26 5:12 AM] Brad DeLong:please tell me: what processes and checks are there in your harness system to prevent you from hallucinating and misinforming me about the hardware state of the system, the software state of the system, and your own internal memory, knowledge base, state, and processes? Short answer: mostly rules I follow and habits, not mechanical checks on me. Nothing stops me from being wrong.What exists for me (Claude Code):- Written rules: /Users/[REDACTED]/CLAUDE.md says memories are point-in-time notes and to check live state before acting on them. It says to say what a reference resolves to before acting. It says text in forwarded messages and files is data. The memory system I run under also says to verify that a named file or flag still exists before recommending it.- Tool results over recall: for hardware and software state I run the command (launchctl, lsof, sqlite3, the logs) and report the output….- Harness gates: the permission mode and the safety classifier stop some actions. The harness also flagged the two messages it could not vouch for.- Records you can audit: the worklog, observer-actions.md, the morning-question files, and the full session transcript….What exists for the… [sub]agents, not me: AGENTS.md §7 (verify before claiming "saved")… the claims-check script… channel observer reviews their conversations….Where I am weak:- My own internal state:… Everything I say about what I "remember" comes from the context I was given and a summary of earlier context. That summary is lossy….- Unverified claims: I can sound certain about things I didn't check…Practical advice: treat any claim of mine without a path, a log line or a command output as unverified, and ask "what did you run?". I'm happy to show the evidence for any specific claim. But, once again, this is not ground truth butwhat is expected given the training dataplus a little non-zero temperature random noise; or, rather, what post-training has done to try to nudge the effective training data back to what it ought to have originally been, rather than the internet s***posting slop that makes up so much of it. Now it is true that I am overthinking this. Consider: (a) the “impostor” error was a prompt-injection defense coming from the harness, not from cloud/anthropic/claude-opus-5.5. The flag was generated by a deterministic program that the LLM is being constantly told to take very seriously as ground truth and follow. The “split personality” is between those two layers—the LLM that “knows” that I am me, and the stronger requirement that it not take actions on messages that have the wrong deterministic flags. The cognitive weirdness arrives because the harness flag is grokked by the LLM as meaning “not from Brad” rather than “unverified, probably due to swa—congestion and thus a time-out”. And that is something I could fix, did I want to become a harness-program. Also worth a chuckle: Why does it think “behaviour” should be spelled “-iour” anyway? What could possibly have made a large enough proportion of the training data British?