Labs V11–V15

Outcome validation, controlled adaptation studies, grounded data, specialized task models, and deterministic agent environments.

Halo Forge’s next five Lab phases extend the managed own-data workflow without adding a project system or an unrestricted experiment matrix:

Own data → proof run → validated outcome → full run
         → controlled adaptation study
         → grounded reviewed data
         → specialized task models
         → replayable local environments and trajectories

Outcome validation

After a proof run, Assess proof outcome checks the optimizer update, artifact and replay integrity, dataset identity, and compatible development evidence. Technical completion and quality change are reported separately. A full run requires the completed assessment or a reasoned operator override.

Adaptation studies

Experiments → Studies supports paired A/B, dose response, and bounded 2×2 designs. The default paired seeds are 17, 42, and 101. Domain uptake and general-capability retention remain separate evidence.

Grounded data

From an immutable corpus Dataset Version, Create grounded data proposes cited training or development-evaluation records. Every citation retains its document, source span, and source hash. Generated records remain suggestions until a person creates and completes a Review Studio queue.

Specialized task models

Guided Own Data supports text and media classification, multi-label classification, embedding pairs, and reranking. Verified task artifacts include their model head, processor, label or retrieval contract, fixed-input verification, and replay identity. Local serving exposes classifications, embeddings, and reranking without claiming generative-model compatibility.

Agent environments

Evaluate → Environments provides deterministic local fixtures, episode suites, step evidence, snapshots, exact trace replay, comparisons, and reviewed trajectory publication. The first release cannot silently write to external systems and does not add online environment reinforcement learning.

All five phases use additive schema v18, replay v9, the durable Activity scheduler, bounded APIs, and existing Dataset, Review, Evaluation, Artifact, and training services.