Developmental Architecture in a Small Artificial Learner: An Integration Audit of a Fourteen-Iteration Inheritance Chain Using Heavy Instrumentation at Cross-Layer Coupling Moments

The organising question for this programme is whether the architectural conditions for developmental learning behaviours in artificial agents can be specified, isolated, and tested through deliberate design — drawing on classical principles of experiential learning and human developmental psychology as genuine architectural commitments rather than post-hoc analogies. The present paper is the integration audit of that programme's first arc: not through scaled training regimes but through a framed environment, staged acquisition, and architectural primitives that leave room for biographical individuation to surface within the framework's rules. A preceding programme (Baker, 2026a–s) built a small tabular developmental agent across fourteen iterations, each pre-registered before code, each addressing a specific architectural question raised by the preceding work. By the close of v0.14 the architecture comprises five cell types in a 20×20 structured environment, three-phase developmental staging, a persistent threat layer with category-level signature-matching, a persistent mastery layer with four mastery interventions, late-target activation through random-location end-state cell sampling, and competency-gated content transformation in which hazard cells transition to knowledge cells when the agent reaches per-cell competency thresholds. Five behavioural signatures associated with developmental learning — rule adherence, focused autonomy, feature-aligned preference formation, the specialist-generalist trade-off, and biographical individuation — were established at v0.8 and progressively preserved across the inheritance chain. Cross-layer behavioural coupling effects were documented at three iterations and named as a phenomenon-class the programme could speak about. This paper audits that cumulative inheritance chain. The architectural arc is not extended; v0.14 is the architecture v1.0 audits unchanged. What is contributed is heavy instrumentation that captures per-state snapshots at six classes of cross-layer coupling moment (attractor banking, cell-type transition, knowledge banking, threat-flag clearing, end-state activation, end-state banking) producing data the iteration papers' per-run metrics could not carry. A supplementary batch at extended run length characterises the architecture's pace distribution under conditions where the experimental window does not censor the data. Across a pre-registered batch of 180 runs spanning six cost levels, three run lengths, with seeds matched to the existing inheritance chain (Baker, 2026t), the integration audit produces five findings. The first concerns preservation. All twelve headline figures from the v0.14 batch are reproduced byte-for-byte under matched seeds. Phase 3 cleanness holds at 60 of 60 across every cost-and-run-length condition — exceeding the v0.10 baseline at the shortest run length and demonstrating that the cumulative inheritance chain produces stronger preservation than any single iteration alone. All six Category α preservation sub-checks pass. The second concerns cross-layer coupling characterisation. The instrumented data confirms the three iteration-level cross-layer effects mechanistically. 862 of 862 cell-type transitions show atomic update of the gate exclusion set with the cell-type state change. 160 of 160 knowledge-banking events on flagged cells show atomic threat-flag clearing. 115 of 115 end-state activation events show drive-weighted score change across all four candidate actions. The phenomenon-class observation across three iterations is confirmed at snapshot resolution; the escalation to structural property of the architecture is supported by the empirical evidence. The third concerns biographical individuation. The original Category γ-1 metric (top-attractor diversity) collapses under the cumulative inheritance chain's mastery interventions, returning Python's dict-iteration-first key (3, 3) for 178 of 180 runs because all attractor preferences have been depleted to zero. A pre-registered amendment (v1.0.1) retires the metric and reframes Category γ around three measures that operate on data which genuinely varies under the inheritance regime: mastery-sequence diversity (mean 9.83 of 10 distinct sequences at 160,000 steps), knowledge-banking-order diversity (96.7% across full-banked runs), and discovery-time diversity given completion (standard deviations 3,029 to 23,192 steps). All three measures pass. The substantive finding the amendment surfaces: across the inheritance chain, biographical individuation has moved from terminal-preference (the v0.8 measure) to trajectory-and-completion (the v1.0 measures), consistent with the architecture acquiring more developmental structure across iterations. The fourth concerns the architecture's pace distribution. A supplementary batch at 320,000 steps, pre-registered under the v1.0.2 amendment, brings 30 of 30 runs to architectural completion. Completion times span 17.6× between fastest (9,016 steps) and slowest (158,645 steps), with median 47,748 and standard deviation 42,094. The distribution is right-skewed with a long slow tail. Decomposition by phase shows that time to final mastery dominates total completion; activation latency (final mastery to end-state activation) is short, with median 0; discovery time scales with cost. Threshold-permutation analysis surfaces a counter-intuitive finding: when the near-cluster's lowest hazard threshold is 1, completion times are roughly twice those of runs where it is 2. The mechanism: an early near-cluster transition introduces a feature-rewarding knowledge cell while the agent is still mid-mastery, dividing the agent's attentional budget between knowledge banking and continued attractor traversal. The fifth concerns the audit's surfacing of its own calibration question. The original Category γ-1 metric's collapse, identified during analysis of the v1.0 batch, was resolved by the v1.0.1 amendment before any paper draft was written. The pattern is itself evidence for the architectural-statement claim that pre-registration discipline, matched-seed audit, and prospective threshold calibration constitute methodological prerequisites for governable architectural research. Two amendments were issued during v1.0; the amendment budget stands at two of three at the close of the iteration. The architectural claims this paper makes are bounded by the pre-registered Category Ω structure. The paper does not claim the present architecture establishes any of its philosophical claims; it claims the audit data supports them as evidence. Forward-looking remarks are reserved for the discussion's closing section and are deliberately broad in their treatment of subsequent work. The programme's contribution is not any individual architectural mechanism but the demonstration that developmental learning signatures can be specified, pre-registered, tracked across a cumulative inheritance chain, and shown to survive — and to relocate — as the architecture acquires more developmental structure. That combination, to the authors' knowledge, has no direct precedent in the developmental AI literature.