Shohin ETTR-IL-v3 Initializer

This private dataset initializes Shohin's Endogenous Typed Theory Reactor (ETTR). It contains formally generated, independently replayed WORLD/COMMAND/QUERY episodes with typed packet, transaction-trace, terminal packet, and answer supervision.

Intended use

The dataset is intended only for the frozen 192,779,435-parameter Shohin ETTR architecture and matched scientific controls. It is not a general-purpose instruction dataset.

Data construction

Every semantic target is produced by a formal primary executor and a structurally independent replay executor. A row is rejected if the executors disagree at any intermediate state or if replaying its generic transaction trace does not reproduce its terminal packet.

The semantic families are:

typed Horn closure;
bounded typed local rewriting; and
guarded resource processes.

The training curriculum covers compiler grounding, atomic transactions, dependent composition, query/counterfactual grounding, and closed-loop source-deleted invariance.

Scale target

The selected training population contains:

40,000 unique semantic cores;
four controlled views per core;
2,560,000 expanded rows; and
1,351,680,000 charged full-objective positions.

Repeated optimizer exposure is not counted as new data.

Configurations

training: candidate-visible source segments and architecture targets.
audit: assessor-only latent semantics and independent execution receipts.
development: held-out development rows and receipts.
reserve: frozen unused candidates.
confirmation-commitments: hashes and counts only.

Confirmation payloads are stored separately and are not mounted during fitting.

Quality controls

The release manifest binds:

primary/replay agreement;
exact trace-to-packet replay;
packet and trace capacity;
semantic and surface deduplication;
graph-isomorphism split ownership;
train/development/confirmation decontamination;
operation, law, depth, query, outcome, renderer, topology, and packet-density coverage;
counterfactual sensitivity;
shortcut diagnostic reports; and
every source, environment, tokenizer, shard, and manifest SHA-256.

See R12_ETTR_IL_V3_INITIALIZER_DATA_PROTOCOL.md in the source repository.

Limitations

This dataset does not prove that Shohin reasons. It defines a high-quality initializer and controlled evaluation substrate. Synthetic transfer is bounded by its formal families, renderers, and query grammar. Natural-language surfaces are admitted only when they round-trip to the same formal semantics.

No protected Shohin checkpoint or model weight is included.