I’ve been working on an alignment framework that tries to address a gap I kept running into when reading existing work: most approaches either assume fixed values (reward functions, constitutions) or allow learning without a clear notion of identity continuity.
The core idea is to represent values explicitly as a constraint-satisfaction network, rather than collapsing them into a scalar reward. Values are nodes; relationships encode support or tension; hard constraints protect a small core, while adaptive values can evolve.
On top of that, I explore a governed mechanism for proposing new value concepts when existing ones fail to resolve persistent tensions. This is done under strict procedural controls: novelty checks, entropy-based confidence measures, human approval, provisional periods, and reversibility. The goal is to allow limited value evolution without losing corrigibility or identity stability.
Other components include: An entropy-based internal stability observer (to detect unreliable internal states) Continuous embedded adversarial testing (MDL-optimized) Asymmetric graceful degradation (easy to lose autonomy, hard to regain)
I’ve written up:
A full technical paper with formalization and invariants An executive summary A working Python implementation of the core reflection engine
Dao Heart 3.0: Identity Preserving Value Evolution for AI Systems Untitled Draft
I’ve been working on an alignment framework that tries to address a gap I kept running into when reading existing work: most approaches either assume fixed values (reward functions, constitutions) or allow learning without a clear notion of identity continuity.
The core idea is to represent values explicitly as a constraint-satisfaction network, rather than collapsing them into a scalar reward. Values are nodes; relationships encode support or tension; hard constraints protect a small core, while adaptive values can evolve.
On top of that, I explore a governed mechanism for proposing new value concepts when existing ones fail to resolve persistent tensions. This is done under strict procedural controls: novelty checks, entropy-based confidence measures, human approval, provisional periods, and reversibility. The goal is to allow limited value evolution without losing corrigibility or identity stability.
Other components include:
An entropy-based internal stability observer (to detect unreliable internal states)
Continuous embedded adversarial testing (MDL-optimized)
Asymmetric graceful degradation (easy to lose autonomy, hard to regain)
I’ve written up:
A full technical paper with formalization and invariants
An executive summary
A working Python implementation of the core reflection engine
Everything is public here:
GitHub: https://github.com/Mankirat47