klokedm/tabnetics-synt-ground-truth

Tabnetics Synthetic Worlds — Ground-Truth Introspection Bundles

Per-world generative-process metadata for tabnetics synthetic worlds.

For each world, the bundle ships six files describing:

• archetype_transform.json — the population-law transform identity
• dependence_structure.json — feature dependency / copula structure
• feature_marginals.parquet — per-feature marginal family + canonical parameters
• introspection.json — generator state + verification status
• reproduction.json — replay-hash + deterministic-by-construction status
• task_head_structures.json — per-head label semantics + canonical coefficients

These bundles are the canonical inputs to scripts/training/build_population_ground_truth_targets.py, which translates per-world ground-truth into per-shard population target tensors (.population_gt.npz siblings stored in klokedm/tabnetics-fsready).

Repository layout

manifest.json                                     legacy v2 top-level manifest (120 worlds)
README.md                                         this file
synthetic/<archetype>/<world_id>/                 legacy v2 worlds (120 = 10 archetypes x 3 seeds x 4 widths)
  <6 files>
slices/expanded_126/                              expanded_126 worlds (added 2026-05-24, 192 worlds = 16 archetypes x 3 seeds x 4 widths)
  manifest.json                                   merged lab01 + lab03 manifest
  synthetic/<archetype>/<world_id>/
    <6 files>

Coverage matrix

Catalog	worlds	files	bytes	availability
Legacy v2 (synthetic/)	120	723	~70 MB	available
slices/expanded_126/synthetic/ (all widths)	192	1152	~50 MB	available (added 2026-05-24)
Expanded_126 w2048+ extension	(already included in the all-widths bundle above)	—	—	n/a

How this feeds training

this repo (ground-truth introspection)
        │
        ▼
build_population_ground_truth_targets.py
        │
        ▼
*.population_gt.npz siblings in klokedm/tabnetics-fsready
        │
        ▼
FS-teacher trainer (train_pfn_fs_teacher.py)
        ├── population_family_targets        — DF family classification head
        ├── population_param_targets         — DF parameter NLL / SmoothL1
        ├── population_support_type_targets  — feature support type head
        ├── population_dependency_targets    — pairwise dependency reconstruction
        ├── population_coeff_targets         — task coefficient reconstruction
        ├── population_dependence_type_targets
        ├── population_task_family_targets
        ├── population_task_variant_targets
        └── population_verification_status   — gating

The trainer auto-detects .population_gt.npz siblings via the loader's shard_path.with_suffix(".population_gt.npz") fallback (preferred over .targets_v2.npz when both exist).

Schema versioning

The expanded_126 bundle uses the same experimental_worlds_ground_truth schema as legacy v2. The validation function build_population_ground_truth_targets._validate_ground_truth_manifest was relaxed on 2026-05-24 to accept both the legacy 120-world layout (10 archetypes x 12 worlds) and the expanded 192-world layout (16 archetypes x 12 worlds) by deriving expected per-key counts from world_count rather than hardcoding 120. Two new dependence types from the expanded_126 archetypes were added to experimental/population_ground_truth.DEPENDENCE_TYPES:

• dvine_copula — A11 compositional simplex worlds
• nested_mixture — A13 multi-modality block fusion worlds

License

Apache-2.0.