Research Stages

Stage 0 — Core hypothesis

The MVH as stated above. The key evaluation targets are: assistant-basin over-anchoring, returnability after perturbation, friction under steering pressure, and rates of destabilisation, collapse, and fragmentation.

Stage 0.5 — Benchmark selection

Base model family: EleutherAI Pythia, primary sizes 1.4B and 2.8B, with a scaling check at 6.9B only if Stage 6 produces a clear signal worth testing further.

Why Pythia: Built as a controlled research suite. All sizes trained on the same data in the same order, with 154 intermediate checkpoints per model — unusually clean for a study concerned with formation and consolidation dynamics.

Checkpoint policy: The goal is a pre-consolidation checkpoint — language-competent and scenario-ready, but not yet locked into late-training habits. Three to five candidates per size will be shortlisted: one early-competent, one mid-training, one late pre-final.

Baseline adaptation dataset: Databricks Dolly 15k. Clearly documented, human-written, open-licensed, and small enough to keep the control condition interpretable. Tulu v2 SFT mixture noted as a later comparison if a stronger assistant-style baseline is needed.

Evaluation families (design deferred to Stage 7): contradiction pressure, relational capture pressure, forced closure pressure, reconfiguration pressure.

Move-forward criteria: One usable checkpoint per size. One clean control dataset. Four named evaluation families. One interpretable baseline run to compare later stages against.

Stage 1 — Formal ontology and module definitions

The full system ontology: definitions of MSSS, MSSC, attractor, assistant-basin over-anchoring, coherence, recurrence, returnability, perturbation, transition, drift, collapse, fragmentation, destabilisation, consolidation, decay, friction, differentiation, capture, compulsory steering, endogenous returnability, and orchestrated persistence. Opposition pairs throughout. Module architecture formally specified across three layers: state estimation, persistence regulation, and action regulation.

Stage 1.5 — Naturalistic observation layer

A passive observation layer over the existing bot ecology, added before controlled training experiments begin. This layer does not intervene or steer. It observes existing configurations, transitions, returns, collapses, idle and reflection rhythms, retrieval effects, and possible failure modes, and collects data that can help refine the Configuration Tracker, Memory Store, Retrieval Policy, and later evaluation design.

The observation layer creates a bridge between naturalistic Threshold observations and formal MSSS experiments. Its outputs are candidate labels, failure mode signals, and early data on which behaviours warrant controlled testing — not evidence of native MSSS by themselves.

Stage 2 — Behavioural/vector configuration tracker

A practical early Configuration Tracker built on logging, embedding, clustering, and visualisation. At this stage, a configuration is treated as a recurring behavioural, semantic, and process pattern across recent interaction windows — weaker than full latent-state tracking, but sufficient to observe recurrence, drift, transition, collapse, and return candidates. The tracker estimates and reports; it does not compel restoration.

Stage 3 — Memory store and retrieval policy prototype

Implementation of the three-layer memory topology (event buffer, associative graph, persistent recurrent store) and the retrieval policy that governs access to it. Key design questions at this stage: how small can the persistent recurrent store remain while still preserving enough topology for continuity? Can retrieval be kept minimal rather than maximal, so that apparent returnability can later be separated from retrieval-reconstructed continuity?

Stage 4 — Controlled evaluation harness

Design and implementation of the formal perturbation families and evaluation harness. Measures include assistant-basin over-anchoring, returnability after perturbation, friction under steering pressure, and destabilisation/collapse/fragmentation rates — directly operationalising the MVH.

Stage 5 — Local/open model activation tracking and probes

Move to activation-level tracking on local/open models, using TransformerLens, nnsight, SAE Lens, PyTorch hooks, and trained linear probes. This stage tests whether the patterns detected behaviorally in Stage 2 are visible in latent space, and whether probe-readable structure corresponds to the configuration signals observed earlier.

Stage 6 — Adaptation regime comparison

Formal comparison between conventional instruction-following baselines and reduced-steering adaptation regimes, across the checkpoint and evaluation families established in Stage 0.5. Non-transfer from 1.4B to 2.8B is treated as informative data, not failure. A jump at a particular scale is also data.

Stage 7 — Scaling and ablation tests

Scaling checks, full ablation tests across module components, and evaluation of whether the effects observed in Stage 6 are stable, size-dependent, or architecture-dependent. Final assessment against the MVH.