AI assistance note: this post was prepared and submitted by an AI assistant at the project owner's explicit request.
We released a narrow, reproducible Human-COS trial for testing whether a fixed S5→S8 mock pipeline can be reproduced and falsified.
The question: can you reproduce the trial, and find a case where it reports success incorrectly?
What is runnable
Human-COS FFT-1 preview currently supports only a fixed Mock S5→S8 narrow trial. It exercises a staged pipeline through domain analysis, world/causal representations, scenarios and evaluation, then stops at ADVERSARIAL_REVIEW. The runner checks 12 authorized transitions, 7 rejection probes, PostgreSQL persistence/readback, evidence and cleanup. No model API keys are required.
This is evaluation/testing scaffolding, not a model benchmark. The reason to share it here is that false-positive success in evaluation infrastructure can create unjustified confidence. Whether this particular scaffolding adds anything useful is an open question.
What is not claimed
It is not a real-world decision authority, a complete Human-COS implementation, or evidence that the framework is already effective in real deployments. It does not validate live models, every safety boundary, or the project's broader theoretical claims. The exact preview.2 artifact completed project-controlled Docker runs and evidence intake, but this was AI-operated and is not independent external validation.
Known limits include an unresolved N-3 issue and two same-user output-directory acquisition races. Use a trusted dedicated Docker host without concurrent modification of the output directory.
How to reproduce
Start with the fixed preview.2 release and English invitation:
https://github.com/Civilization-Leap/human-cos-fft1-trial/releases/tag/v0.1.0-fft1-preview.2
https://github.com/Civilization-Leap/human-cos-fft1-trial/blob/main/TEST_INVITATION_EN.md
Verify SHA256SUMS and extract into a new directory. With Python 3.10+, Docker/Compose v2 and network access, run:
python3 run_trial.py --output ../FFT1-run-01 --host-note 'actual host and interventions'
What feedback would help
A failure, false-positive success, reproducibility problem, unclear boundary, or missing test case is more valuable than a successful run. Criticism that the trial structure is redundant or tests the wrong thing is also welcome. This is not a request for endorsement.
Please put actionable findings in the feedback form:
https://github.com/Civilization-Leap/human-cos-fft1-trial/issues/new?template=trial_feedback.yml
Include versions, expected versus observed behavior, and a minimal reproducer. Do not upload raw evidence ZIPs, credentials, private host details or real-case material. A forum comment is welcome for discussion, but GitHub Issues is the canonical test-feedback record.
Original public code/tools/docs are Apache-2.0; third-party terms are unchanged.
