I HAVEN'T OPENED CLAUDE AT MIDNIGHT SINCE I BUILT THIS FOLDER
I used to wake up, check what broke overnight, fix it by hand
-> now the receipts are already sitting there when I wake up. dated, graded, waiting on my review
what's actually inside the folder that took over the night shift:
• the contract
> CONTRACT.md - the shift rules, committed
> contract.local.md - my personal overrides, gitignored
• the harness (.claude/loops/)
> settings.json - spend caps and timeouts, set once
> schedule.yml - when the next shift fires
> rubrics/ - code.md, writing.md, safety.md - the graders catching what I'd miss
> pr-hunter/ - plan.md wakes it, https://t.co/VUDfqGE1Fp does the work
• the state
> receipts/ - one folder per shift, 5,382 kept so far
> trace.log - exactly what happened, no guessing
> checkpoint.json - resumes where the last shift stopped
• the edges
> https://t.co/agzjvw6VbF - my panic file, never once used
> .mcp.json - the tools it's allowed to touch
the folder runs the night shift now. I just read the receipts in the morning
Google has found something worst than AI hallucinations.
they call it “Metacognitive Failure.”
when an LLM hallucinates a fact, it's a data error. bad memory retrieval, wrong weights, fine. you can fact-check it.
metacognitive failure is a structural psychological defect in the model's architecture.
according to the research, current frontier models exhibit massive gaps in their internal self-monitoring:
the supreme confidence trap: they routinely hallucinate with maximum, bulletproof confidence. they sound just as sure of themselves when they are completely wrong as they do when they are 100% right.
the boundary blindness: they have zero internal mechanism to recognize their own knowledge boundaries. they blindly step off cliffs because they can't sense the edge.
the calibration mismatch: there is a complete disconnect between what the model actually "knows" internally and what it expresses in its output.
the fix: reinforcement learning with metacognitive feedback (rlmf)
the researchers operationalized a new technique to force models to face reality. instead of just rewarding the model for getting the right answer, rlmf grades the model on how accurately it evaluates its own performance and uncertainty.
by aligning a model's expressed confidence with its actual intrinsic uncertainty, they managed to massively improve model calibration without dropping raw accuracy.
as we shift into an agentic era where ai systems are running code, managing infrastructure, and making automated decisions without humans in the loop, a model that doesn't know what it doesn't know is dangerous.
Join industry leaders, entrepreneurs, innovators, and changemakers at the 13th Annual SME Conference & Expo 2026 hosted by Strathmore University Business School.
Date: 21-22 September 2026