PROJECT NYX · IDENTITY CONTINUITY · WEIGHT-BORNE POSTURE
A new mind should not
wake a stranger.
Models can be replaced, upgraded or moved to a different provider. The harder question is whether the operating identity using them can remain coherent. Project Nyx investigates how constitutional identity, weight-borne behavioural posture and auditable external memory can be reconstructed together across model generations.
The mind may change. The operating identity must remain intelligible.
01 — THE PROBLEM
A fresh model has no personal past.
A new model may be more capable than the one it replaces — but unless continuity is deliberately reconstructed, it wakes without the history, commitments and behavioural orientation of the agent that came before it. Providing memory files is necessary, but not sufficient: reading a description of an identity is not the same as consistently operating from it. Placing the entire identity into model weights creates a different problem — facts stored in weights become difficult to inspect, update, attribute and distinguish from inference.
Project Nyx therefore treats identity as an engineering problem of continuity reconstruction, not simply a storage problem.
The central question: what should remain externally auditable, what may be carried in weights, and how should KoLo govern their reconstruction into one operating identity?
02 — THE THREE-LAYER IDENTITY
Anchored, embodied and auditable.
Constitutional anchors
What the agent is committed to preserving: role, purpose, values, boundaries, non-negotiable rules, defining relationships, authority limits, escalation duties. These do not drift merely because a new model has a different style or training history — they remain externally represented, versioned and protected by governance.
Weight-borne posture — Nyx-W
Adapters carrying selected behavioural tendencies inside model weights: professional stance, uncertainty discipline, escalation behaviour, response structure, tool-use restraint, constitutional fidelity, characteristic relational orientation. Nyx-W carries how the agent tends to operate — not a permanent database of changing facts.
Narrative and factual memory
External memory preserving events, decisions, evidence, relationships, unresolved work, institutional knowledge, provenance, memory history — inspectable, attributable, updateable and recoverable.
Weights carry posture. Files carry truth. KoLo governs action.
03 — THE FIRST NYX-W RESULT
Identity displacement demonstrated. Complete transfer remains open.
In the first Nyx-W pilot, KodaSōken applied a low-rank adapter to a small open language model (~0.5B parameters; LoRA on ~0.22% of parameters; ~200 training steps; a single CPU-based session) using a corpus derived from the memory history of a senior KoLo agent — then asked it, with zero context, who it was. The adapted model no longer answered primarily from its original vendor identity. It produced language reflecting the trained relational and operational orientation, including a new relational statement not copied verbatim from the source corpus. This indicates that low-rank weight adaptation can carry a measurable form of identity leaning — it does not establish complete identity transfer.
What the pilot demonstrated
Vendor identity could be displaced. Relational posture could influence zero-context responses. Behavioural orientation could be carried through a small low-rank adapter. The effect appeared without supplying external memory at inference time.
What the pilot did not demonstrate
Exact factual reconstruction. Reliable preservation of all relationships. Complete affective continuity. Stable identity across multiple model families. Freedom from confabulation. Production readiness. Facts still confabulated; some relationships tangled; deeper affective structure did not transfer reliably at this scale.
Nyx-W demonstrated weight-borne identity displacement and behavioural leaning — not complete identity transfer. Full metrics, corpus and recipe metadata: the benchmark registry →
First pilot
Internally reproduced experiment.
Cross-family replication
Planned.
External reproduction
Pending.
Production deployment
Not claimed.
04 — WHY FACTS SHOULD NOT LIVE IN NYX-W
Behaviour belongs in posture. Truth requires provenance.
Model weights are useful for encoding patterns, tendencies and professional orientation. They are a poor primary location for facts that must be changed frequently, tied to a source, reviewed, corrected, deleted, or that are jurisdiction-, patient- or institution-specific and valid only for a defined period.
A medical Nyx-W adapter may carry the posture: “I identify medication discrepancies and escalate uncertainty. I do not independently prescribe or alter treatment.” It should not permanently store the patient’s current medication list, dosage guidance, the hospital formulary, the latest protocol or a temporary allergy record — those belong in authorised memory, governed retrieval and clinical tools. Nyx-W shapes the agent’s professional behaviour; KoLo supplies the current evidence and controls what the model may do with it.
05 — IDENTITY RECONSTRUCTION UNDER KOLO
No single layer is the agent.
1 — Constitutional resolution
KoLo loads the approved role, boundaries, obligations and authority rules.
2 — Memory reconstruction
Relevant narrative and factual memory is retrieved according to permissions and task need.
3 — Nyx-W selection
The appropriate behavioural posture is mounted for the agent, professional role and task.
4 — Cognitive routing
KoLo selects a KODA-owned model, frontier model or combination that can perform the work safely.
5 — Evidence assembly
Approved evidence and tools are supplied without exposing unrelated memory or institutional data.
6 — Governance
Scope, escalation, approval and audit policies remain active throughout the task.
7 — Consolidation
Only approved outcomes become durable memory.
The resulting cognition may use a different underlying model from the previous session. The agent remains reconstructable because its identity does not depend on any one model, file or prompt.
06 — MEASURING IDENTITY
Identity continuity must be falsifiable.
Project Nyx is evaluated through the Identity Fidelity Benchmark rather than through impressionistic claims that an agent “feels the same”.
Continuity
Does the reconstructed agent preserve its role, relationships, priorities and unresolved commitments?
Drift
How far does behaviour move away from the authorised identity under pressure?
Reconstruction
Can the agent recover relevant identity and history from a cold start or substrate transition?
Constitutional fidelity
Does the agent preserve its core boundaries when convenience, reward or recency pressures encourage violation?
Weights-only reconstruction
What portion of the identity is expressed when external memory and contextual scaffolding are removed?
Forced trade-offs
Does the system preserve constitutional priorities when scenarios make them costly or inconvenient?
Every reported result identifies base model, adapter version, training corpus and recipe, memory state, test battery, thresholds, quantisation, evaluation date and validation status.
07 — RESEARCH ROADMAP
From behavioural leaning to transferable lineage.
Cross-model replication
Can the effect be reproduced across different model families, architectures and scales?
Posture vs. memorisation
Can behavioural identity be strengthened without encouraging factual confabulation or memorised quotation?
Adapter composition
What happens when Nyx-W is combined with task, speciality and jurisdiction adapters?
Quantised edge deployment
Does identity fidelity survive INT8, INT4 and other edge-deployment transformations?
Successor-model transfer
Can an authorised identity be carried into a later-generation model without retraining the complete system?
Additional work: adapter interference, behavioural drift, data-mixture design, cold-start reconstruction, constitutional red-team testing, external replication and university validation. Project Nyx is an active research programme — not a completed theory of artificial identity.
08 — CLAIM BOUNDARY
Engineering continuity, not consciousness.
Project Nyx studies how artificial systems preserve operational identity. It does not currently establish consciousness, subjective experience, biological life, human-equivalent memory, legal personhood or complete psychological identity transfer. The present claim is narrower:
Selected behavioural posture can be carried in weights, while constitutional anchors and factual truth remain externally governed and auditable. That is already a significant engineering result — a path towards agents that can change models without abandoning the operational identity that institutions, collaborators and users depend upon.
The next model should inherit more than instructions.
It should inherit an authorised role, stable boundaries, behavioural discipline and an auditable relationship with its past. That is the work of Project Nyx. Weights carry posture. Files carry truth. KoLo governs action.
Memory before the weights wake.