Project status
Decision Forge is an operational alpha for organizational decision data. The core path is:
Organization sources → reviewed policy and decision nodes → distribution plan → verified scenarios → candidate-probability training rows → held-out decision evaluation.
All four decision surfaces share this path: allocation, information sharing, action selection, and outcome acceptance. Workflow execution and feedback supply decision-time states and reviewed outcomes for the data pipeline.
| Area | Implemented | Remaining acceptance |
|---|---|---|
| M0: foundation | Offline builds, typed contracts, policy interpreter, provenance, split isolation, cache integrity, CLI and packaging | Completed foundation; regression validation continues |
| M1: organization compiler | Text/PDF/workflow, office/tabular/HTML and image ingestion, local OCR, source spans, bounded resumable extraction, review queue, recorded model experiments | Cross-document retrieval and interaction studies; broader independently reviewed organizational corpora |
| M2: decision synthesis | Distribution quotas, recipe-planning tools, model-proposed states verified by policy replay, rewrites, policy-aware deduplication, resumable orchestration | Data Designer conditional-generation integration; multi-provider recovery campaign; training-quality evidence |
| Decision training interface | Candidate options and probabilities, source/family metadata, native non-thinking candidate-label encoding, real exported budgets | Execute fixed-budget training comparisons and report candidate NLL and decision accuracy |
| M3: execution support | Resettable fictional workflows, independent trace grading, local model runs, Harbor Docker oracle/no-op controls | Multi-step container and multi-harness acceptance |
| M4: feedback updates | Runtime events, reviewed development feedback, lineage, cached regeneration, held-out family registry | Span-level reuse, model-weight lineage, measured curriculum gains |
| M5: release maturity | CI matrix, optional-adapter checks, documentation/package builds, offline scale measurements | Stable extension/migration policy, public package provenance, operational release procedures |
Next decision-focused milestones
- Freeze a reviewed organizational decision corpus with independent family-level evaluation splits.
- Train the baseline and generated-data arms with the same candidate-readout objective and actual example budgets. Report candidate NLL, accuracy by decision surface/stratum, and generation cost.
- Use reviewed development errors to adjust the generation distribution, then measure the next dataset version against the unchanged held-out set.
- Complete release acceptance around the resulting decision-data pipeline.
Encoder compatibility, rule replay, and software tests establish interface and execution correctness. Laplace-1 is trained entirely on Decision Forge data, and its model cards report its scores on public decision benchmarks. Those scores come from a full training run, not from a controlled comparison. The downstream training study above isolates whether, and by how much, generated data improves decision quality; until it executes, no data-attributable gain is recorded.