Date: 2026-09-20 Status: FEATURE PASSED Lineage: EMMA Native Foundation 001 → Foundation 001.x → v1 → v1b → v1c-field-continuation

Disposition

Foundation 001.x generalized symbolic operations passed its defined feature gate after a bounded curriculum intervention. The successful revision preserves the Foundation 001 decoder topology and uses the same byte tokenizer, tied embeddings, optimizer family, initialization family, and answer-only loss. Foundation 001 remains unchanged and immutable.

The successful artifact is foundation-001x-generalized-symbolic-v1c.pt, with SHA-256 [checksum retained in the private evidence record]. Fresh-process restore and qualification passed before promotion.

Control and implementation

The parent control is Foundation 001: 9,508,800 parameters, 8 blocks, hidden width 320, GQA 8 query/2 KV heads, SwiGLU FFN, RMSNorm, split-half RoPE, 256-byte vocabulary, 1,024-token context, and tied input/output embeddings. No routing, experts, adapters, copy path, embedding split, or decoder redesign was added.

Implementation is in:

• [retained internal evidence] — procedural curriculum, sealed suites, leakage checks, telemetry, developmental checkpoints, and Stage 1 training.
• [retained internal evidence] — bounded field/position continuation from v1b.
• [retained internal evidence] — fresh-process restore, qualification, checksum, and promotion.

Curriculum and controls

The generator is deterministic and versioned as procedural-symbolic-v2. It preserves the original eight Foundation contract families by replaying their exact training contracts while procedurally generating copy, field selection, direct binding, dereferencing, position/role selection, transformations, and two-operation composition. The stream uses balanced family sampling and records effective family exposures, seed, batch size, optimizer settings, and checkpoint lineage.

Sealed suites contain 1,000 examples each for identity, structural field combination, binding, dereference, position/role, transformation, and composition. Prompt and prompt-answer leakage checks are recorded. Hidden identities are generated from a separate seed range. Original retention and hidden qualification examples remain separate from the new suites.

The focused continuation started from v1b at 100,000 exposures and continued to the directive’s hard 250,000-exposure ceiling. It increased field and position families to 20% each while retaining 40% combined exposure for the original contracts. This was a bounded curriculum intervention; the model topology was unchanged.

Results

The v1 Stage 1 run at 100,000 exposures reached identity 98.1%, binding 94.3%, dereference 98.1%, transformation 95.4%, composition 97.6%, but field combination 23.8%, position 19.9%, and original qualification 86/160. The v1b curriculum correction restored original contracts and reached identity 94.9%, binding 93.8%, dereference 97.4%, transformation 93.1%, composition 96.5%, field combination 21.5%, position 23.0%, and original qualification 151/160 at 100,000 exposures. The focused v1c continuation resolved the field/position deficit without changing architecture.

Developmental and resource evidence

Checkpoints were persisted at 10k, 25k, 50k, and 100k exposures for v1/v1b, and at the 250k endpoint for v1c. The successful v1c candidate has 9,508,800 parameters and adds no trainable architecture. The final continuation run used AdamW, learning rate 3e-4, weight decay 0.1, gradient clipping at 1.0, batch size 64, float32 CUDA computation, and fresh optimizer state over the v1b parent candidate. Full run metadata and hashes are in fresh-process-qualification.json.

Tied-embedding telemetry recorded embedding norm, change from initialization, gradient norm, and training loss at developmental checkpoints. The implementation does not separate input/output embeddings and does not introduce a copy head.

Prior-art gate

The directive-specified references were used as design guidance only: MLC/MLC-ML for procedural and systematic splits; variable-binding research for changing assignments and dereference depth; OOD composition work for copy/generalization controls; LLM-Interpret principles for optional copy diagnostics; and NNsight/TransformerLens as future bounded inspection options. No external source code or framework was copied, no external repository became a runtime dependency, and no additional web literature search was performed. The reusable mechanism was the procedural distribution and leakage-control principle. The EMMA-specific result is that the unmodified Foundation 001 architecture transferred to the full gate after preserving original contracts and focusing unresolved field/position exposure.

Artifacts and lineage

• Foundation 001 control was never modified.
• v1 and v1b failures remain recorded under their run directories and checkpoints.
• v1c parent, candidate, promoted artifact, metadata, and fresh-process report are preserved.
• Promotion occurred only after fresh-process restore, exact sealed evaluation, checksum verification, and original qualification.

Unresolved limitations

Machine-readable evidence

• v1 run
• v1b run
• v1c continuation run
• fresh-process qualification