SISM-ECB-01

Document conventions The terms MUST, MUST NOT, SHOULD and MAY are normative only for conformance with this proposed draft specification. Their use does not imply community-wide adoption or formal standards status. 1. Executive summary The Sensor-Integrated Self Model (SISM) is a functional computational framework organized around four explicit domains: Body Representation (B), Agency (A), Monitoring (M) and Continuity (K). The project tests whether explicit integration of these functions improves structural OOD generalization under strict fairness, blinding and reproducibility constraints. 2. Scientific background Biological and artificial systems contain mechanisms related to body-state estimation, action attribution, monitoring and temporal persistence. Existing approaches commonly study these mechanisms separately or under heterogeneous protocols. SISM contributes an operational decomposition, an explicit integrated self-state, matched causal controls and a preregistered cross-task evaluation framework. The four domains are a Minimal Working Taxonomy, not a complete theory of selfhood. 3. SISM framework The integrated state is S*_t = I(B_t, A_t, M_t, K_t). Each module has explicit inputs, outputs, state variables, intervention points and diagnostic logs. The framework is implementation-independent and makes no phenomenal claims. 4. Mathematical framework At discrete time t, observations o_t are encoded as z_t. The modules update to B_{t+1}, A_{t+1}, M_{t+1}, K_{t+1}; the policy selects actions from z_t and S*_t. Causal operators may remove, freeze, delay, mask or perturb any module. 5. Benchmark architecture SBS-1.0 defines platform-independent benchmark requirements. ECB-01 is the first reference implementation and uses MiniGrid-compatible tasks. Future ECB studies may use other platforms without changing the conceptual benchmark specification. 6. Study design The study evaluates four task families: operational configuration representation, action attribution, monitoring, and continuity. Each includes ID, structural OOD, counterfactual and nuisance conditions. No evaluation split is used for optimization. 2 7. Model families The primary model is SISM-Separate. SISM-Shared is the equivalence comparator. Single-Path, Memory- Router and Concat are the matched controls. Recurrent-Global is reported separately as a high-compute control. 8. Fairness The primary arm MUST pass active-parameter and estimated-compute tolerances, identical data exposure, optimizer steps, batch sizes, stopping rules and hyperparameter-selection procedures. Gradient norms, loss curves, convergence speed, effective updates, early stopping and wall-clock time are reported. 9. Primary endpoint OSAS-Structural = (B_struct + A_struct + M_struct + K_struct)/4. Each primitive metric is normalized using frozen benchmark-defined bounds, direction and clipping rules. No confirmatory-data-derived scaling is permitted. 10. Hypotheses PH1-PH3 test superiority of SISM-Separate over Single-Path, Memory-Router and Concat. PH4 tests practical equivalence with SISM-Shared. PH5 requires all fairness, QC, reproducibility and implementation-independence gates to pass. 11. Analysis unit Task x Seed aggregates are the observed units. Episode-level outcomes are aggregated before inference. Seed is the independent randomization cluster; all task differences within a seed share the same sign flip. 12. Two implementations Implementation A and Implementation B MUST be separate codebases built from the same frozen specification, share no model implementation code, be separately tested and hashed, and be completed before unblinding. They are analyzed separately before any pooled analysis. 13. Missingness A documented technical failure permits one exact rerun. If the rerun fails, the entire paired Task x Seed unit is marked missing for all primary models. No replacement seeds and no confirmatory imputation are permitted.