Skip to content

The Soulforge: Extension of Training

Training begins before the first weight moves.

Soulforge binds an admitted corpus, base-model digest, objective, Recipe, Training Run, evaluation, and candidate weights into immutable lineage. Its tools may change; its evidence must survive the strike.

State of Work records Soulforge as Designed: no corpus-admission service, trainer execution plane, candidate registry, promotion path, or serving integration is delivered. ADR 33 owns the law and complete record schemas.

Soulforge may run during an admitted idle resource window, but sleep supplies no training or consolidation contract. The forge instead follows the explicit passage below.

The Four Passages

Route Receives Leaves behind Refuses
Corpus admission Training Intent and nominated records Corpus Admission, exact membership, negative ledger, and sealed holdout Missing provenance, authority, privacy clearance, relevance, or a viable uncontaminated holdout
Dataset compilation One admitted snapshot Immutable Dataset Manifest with preserved lineage and splits Unadmitted material, split drift, or holdout leakage
Training run Manifest, base digest, objective, Recipe, and resource request Training Run receipt and frozen Candidate Bundle Unpinned inputs, a refused resource window, or unknowable retry state
Candidate handoff Frozen bundle and evaluation contract Riddle Outcomes and an externally owned Promotion Decision Trainer self-grading, mutable custody, or unowned serving

Authority crosses none of these boundaries by implication. Nomination grants no training right; Corpus Admission grants no resource window; a Training Run receipt supplies no independent verdict; and passing Riddle establishes eligibility, not promotion. Orchestrator admits the physical transition. Owning policy and the required Magus/HitL decision govern promotion; the serving owner registers and routes the exact digest.

Observation may trigger rollback or lineage invalidation. It cannot rewrite an earlier record, erase completed effects, or remove learned influence from weights already produced.