01
Ethical Boundary
Consent, safety, truthfulness, non-deceptive identity claims, independent risk checks.
Open to workAI systems · research engineering · safety
Adarsh
Adarsh Kumar
Identity systems that survive a model swap.
I design reconstruction stacks: constraints, memory, observers, and evaluation — not another chat persona.
Systems architecture
Nine functional layers, three-stage model
Constraint design
Symbolic anchors, prohibitions, change policy
Safety engineering
Independent policy layer, 10 failure modes
Evaluation design
Blinded recognition, held-out behavior
Knowledge systems
Obsidian vault, claim metadata, provenance
Research writing
Methods paper in progress · proposed status
HCI / observers
Reconstruction under a specific viewer
Product scoping
N=1 pilots, public vs restricted layers
Work
What I actually built
Hover a card. Filter by how it earns its keep — not by buzzwords.
Stack
Nine layers. Three stages.
Swipe the slides. This is the machinery — not a manifesto.
Static Archive
Potential identity. Not a living agent.
Dynamic Runtime
Temporary reconstruction under constraints.
Enacted Identity
Behavior across turns — inferred, not declared.
Philosophy of mind
Pattern vs vessel · enacted identity
Memory systems
Relational memory · source pointers
Observer / HCI
Reconstruction under a viewer
Alignment
Independent safety · truth over fidelity
Evaluation
Blinded tests · hidden controls
Scholarly method
Claim metadata · export whitelist
Compare
Architecture, not a bake-off
Shipped cells describe public 2026 product behavior. CCCS is labeled Designed — research intent, not a measured win. No MMLU. No invented continuity scores.
| Axis | GPT memoryShipped | Claude projectsShipped | Gemini PIShipped | Companion AIsShipped | RAG agentsShipped | CCCSDesigned |
|---|---|---|---|---|---|---|
| Portable across vendors | — | — | — | — | Some | Designed |
| Identity as constraints | Some | Some | Some | — | — | Designed |
| Vessel swap | — | — | — | — | Some | Designed |
| Source-addressable memory | Some | Some | Some | — | Yes | Designed |
| Observer conditioning | Some | Some | Some | Yes | — | Designed |
| Safety independent of persona | Some | Some | Some | — | Some | Designed |
| Blinded continuity eval | — | — | — | — | — | Designed |
| Bounded evolution + rollback | Some | Some | — | — | Some | Designed |
| Explicit non-claims | Some | Yes | Some | — | Some | Designed |
Swipe sideways to read the full table.
About
Adarsh Kumar
Short on purpose. The work is the argument.
I build the boring, load-bearing parts of identity systems: constraint archives, memory that can be wrong, observers that bias reconstruction, and tests that the character cannot grade.
CCCS is the research program behind that. It is proposed, versioned, and privacy-gated. I am preparing a methods paper — not claiming an arXiv-ready result.
Engineering program
Reconstruct functional identity patterns across runtimes.
Pattern, not vessel
The model is replaceable. The constraint archive is the object of design.
Enacted, not declared
Continuity is inferred from held-out behavior, controls, and observers.
Safety is out-of-character
Policy must not depend on persona fidelity or continuity language.
Why now
The market for this already exists, just aimed at a narrower problem. Chinese 'grief tech' companies now turn a few photos and an audio clip into an interactive 3D avatar of a deceased person — a real, shipping industry, not a thought experiment. It demonstrates real demand for continuity across a discontinuity. It also ships with no independent safety layer, no consent framework for the deceased's data, and no way to test whether the reconstruction is accurate or simply emotionally convincing. CCCS starts from the same observation — that a recognizable pattern can be meaningfully reconstructed after a discontinuity — but treats it as an engineering problem with a broader scope (any system change, not just death) and a mandatory evaluation and consent layer the current market lacks.
Source: Trust Conference 2025, 'AI Resurrection: The Rise of Grief Tech in China,' Context / Thomson Reuters Foundation.
Contact
If the work fits, write.
Hiring, collaboration, or a careful critique of the stack. Notes persist after you send.
Sign in if you want notes kept against your account. Sign in