Status: ExperimentalWorld ModelsModel Evaluation
DAWM — Deep-Perception World Models
World-model experiments, phase diagnostics and failure analysis, treating regressions as a chain of rejected hypotheses rather than a single training failure.
Abstract
DAWM is an experimental world-model line consolidated from Octoryn Research and the octopus-labs theory work. We diagnose phase-by-phase regressions to separate architectural pressure from tuning noise, keeping earlier-phase baselines as controls.
Problem & motivation
World-model regressions across training phases are easily mistaken for optimiser or dataset noise, obscuring architectural causes.
Research questions
- Which Phase 7 regressions are architectural rather than tuning noise?
- Does state-abstraction pressure explain the recurring regression clusters?
Methods
- Phase-local rollback matrices
- Representation-bottleneck probes
- Short-horizon control evaluations
Observations
Present-tense observations from internal work. These are not validated results and have not been peer reviewed unless a linked publication says so.
- Observed: regression clusters are stable across two unrelated parameter sweeps.
Limitations
- Experimental; interfaces and results are unstable.
- Not peer reviewed; no external benchmark has been run.
Disclosures
- Funding
- Infrastructure support provided by Octopus Core Pty Ltd / Octoryn; theory shared with octopus-labs.
- Conflicts of interest
- No specific commercial conflict identified for this experimental line.