Implementation Assumptions

Strong best-fit assumptions made during the build (per Hard constraint #9), with rationale. Each is reflected in code and docs.

Stack

• Single Elixir app, zero hex deps instead of an umbrella. Rationale: a deterministic, offline-testable world kernel is the core asset; module namespaces preserve the spec's separation; umbrella split is mechanical later. (See architecture/overview.md.)
• Built-in OTP :json (OTP 27+) for serialization rather than a hex JSON lib, to keep zero deps.
• SP.Determinism SplitMix64 instead of :rand, for cross-process, version-stable reproducibility.
• No Nx — grids are small; pure data structures suffice. Documented as an extension point.

World model

• Region grid default 6×6, 2 regions initial; spectral 3 bands; 8 material classes. Chosen for tractable soak tests while exhibiting all required phenomena.
• Seam open threshold 0.8; readiness relaxes to an equilibrium that cannot cross the threshold without resonators ⇒ seam engineering is late-stage.
• Conservation claimed/tested for field diffusion and transport; reactions and collapse are explicit transformations (not globally conserved) — documented.
• Field caps: nutrient 5, temperature 2, solvent 2, toxin 3, strain 2, band 3 (clamped each microstep).

Body / viability

• Seed body = core + interoception + chemotactile (the spec's "homeostatic + proximal sensing"). All else must develop.
• Homeostatic set-point 0.5 with a ±0.25 comfort dead-zone; stress damages integrity only outside the band. Upkeep 0.025 + 0.005·n_organs (morphology is never free). Energy intake gated to cells with nutrient > 0.1, capped 0.14.
• These constants calibrate difficulty to "hard but possible"; they are tuning knobs, documented in SP.Body and the ablation reports.

Interface

• Observation schema obs-v1, 29 features; per-seed channel permutation + optional per-channel affine value scramble (default on).
• Action params are relative only; absolute coordinates rejected.

Hybrid time

• Defaults: 3 microsteps/decision, development every 5 decision ticks, 400-tick reference horizon. Lineage is cross-episode via SP.Genome + SP.Eval.

Evolution

• Genome growth_plan + maturation_rate + thrift; repair/1 guarantees developability via prerequisite closure + topological order. Selection is through world viability, not a fitness scalar.

Validation baselines

• Six baselines; Random/LeakageProbe are blind (opaque interface only); the four scripted ones use the debug Lens and are validation-only, never learners.