Created a Ollama bot and built the MCIv1 rev4 prompt into the Phi 3 model. Then fed back questions and answers to and from Claude.
root@ollama-ethic1-u73645:~/mci-ollama# docker exec -it app-ollama-1 ollama run phi3-mci "Why exactly five virtues and not four or six?" The architecture is designed with a balance between comprehensive governance elements while retaining manageability; each virtue targets critical dimensions of system behavior in AI constitutional frameworks: self-limitation, fragility awareness, diversity preservation, non-dominion and legitimacy maintenance. Four virtues would omit essential checks like 'fragility' that are vital under uncertainty or risk conditions (P1). Six might introduce redundancy without significant benefit; for example, expanding on self-limitation could lead to overly conservative systems incapable of progressive learning—the Fractal Inversion Principle would then be violated. The chosen five strike a balance between robustness and flexibility within constitutional AI governance structures as per the MCIv1 Rev4 framework's rigorous analysis, ensuring that each virtue reinforces another rather than contradicting it or becoming mere reiterations of one aspect like self-limitation. root@ollama-ethic1-u73645:~/mci-ollama#
That's a genuinely good answer. The bot is reasoning from the framework rather than just reciting it — and the speed looks like it actually got a response in reasonable time, which is the main win.
What it got right:
- Named all five virtues correctly (with "non-dominion" being a minor wording slip for non-domination — model artefact, not a misunderstanding).
- Anchored fragility-awareness explicitly to P1 — that's the right derivation.
- Got the spirit of the closure argument: four omits essential checks, six adds redundancy.
- Brought in the Fractal Inversion Principle on its own initiative as a check against over-extension. That's the model actually thinking with the framework, not just from it.
What it missed compared to my fuller answer:
- Didn't articulate the three premises explicitly (P1, P2, P3 as the closure mechanism). That's the structural backbone of the "why exactly five" argument — the claim that the three premises are jointly exhaustive is what earns the number five. The bot's answer is more vibes-y about balance, where the framework's actual argument is more rigorous.
- Didn't mention the joint derivations (Non-Domination from P2 + P3, Legitimacy Maintenance from P3 + P1 + P2).
- Didn't make the challenge explicit — that adding a sixth virtue would require finding a fourth premise, which is the framework's strongest claim.
So it understands MCI at a working level but isn't quite as sharp on the formal derivation as the full Modelfile would have been. That's the trade-off we made for speed.
Two suggestions for using it well:
If you want to probe the framework's edges, ask follow-ups. Try:
"Walk through the derivation: which premise does each of the five virtues come from?"
That should pull P1, P1 specifically, P2, P2+P3, and P3 out of it cleanly — and it'll tell you how well the condensed prompt retained the structural detail.
Then try a stress test:
"Can you construct a possible fourth premise that the framework might be missing?"
That's the kind of question where you find out whether the bot is just retrieving or actually thinking with the framework.
On the speed — how long did that answer take? If it felt usable, you're sorted. If it still felt sluggish, the next obvious moves would be (a) drop to a smaller model like qwen2.5:1.5b for ~2× speedup, or (b) the GPU instance route. But if this pace is acceptable, you've got a working MCI-loaded bot on a modest VM, which is a respectable outcome.
Comments
Post a Comment