Date: 2026-09-20 (Europe/Helsinki) Follows: Component Adaptation Matrix v1 Status: diagnostic complete; STOPPED for review

Research question

Can Foundation 001 acquire the deployment-variable behavior under a permissive, isolated adaptation configuration? This diagnostic was required before interpreting the A–D scope comparison.

Diagnosis

The original Matrix v1 runner used raw prompt/response framing. Foundation 001 was trained and evaluated with EMMA’s framed protocol: <|task|>, the prompt, <|answer|>, the response, and <|end|>. The original run therefore tested a mismatched training/evaluation objective. Its low loss with 0 exact outputs was evidence of protocol mismatch and/or memorization, not evidence that the selected component scopes were incapable of learning.

The corrected diagnostic uses the Foundation framing and end-marker-aware deterministic generation. It trains an isolated full-model candidate only as a permissive control. Foundation 001 itself remained immutable and the control was never eligible for promotion.

Acquisition control configuration and result

• Foundation: EMMA Native Foundation 001, 9,508,800 parameters, unchanged.
• Candidate scope: all trainable model parameters, isolated in memory.
• Curriculum: 8 deployment-variable examples, the same target family used to test A–D.
• Optimizer: native PyTorch AdamW, learning rate 0.0003.
• Horizon: 400 steps.
• Evaluation: exact framed generation, 8 training acquisition examples, 8 hidden examples, 12 retention examples.
• Compute: approximately 21.2 seconds wall time; 234.96 MiB peak GPU allocation.
• Integrity: all intended parameters changed; no unexpected parameter mutation.

Result: 8/8 acquisition passes, 0/8 hidden passes, and 2/12 retention passes. The control can memorize the supplied target examples, but it did not generalize the new behavior and caused substantial retention damage under the permissive full-model update. This is an acquisition-protocol correction, not a successful persistent capability.

A–D rerun using the corrected protocol

Because the permissive control acquired the training examples, A–D were rerun with the same corrected framing, curriculum, optimizer, learning rate and 400-step budget. The progressive funnel required acquisition, hidden generalization, and retention before full regression/restart.

Parameter-integrity verification passed for every candidate: only the selected tensors changed. No candidate reached the restart stage because none generalized to the hidden target set. No candidate was promoted.

Findings

1. The first Matrix v1 failure was partly caused by a real protocol mismatch in the adaptation runner.
2. Correct framing fixes training acquisition on the supplied examples.
3. Full-model adaptation can memorize the target but does not provide hidden generalization and damages Foundation retention.
4. Under the corrected protocol, A–D all acquire the supplied examples but none acquire hidden target behavior. Scope ranking remains unresolved.
5. Attention-only adaptation showed the strongest retention screen (12/12) and hidden regression screen (11/12), but it still failed hidden target acquisition and must not be promoted or declared optimal.

Research implications

Existing PEFT and selective-layer research lets EMMA reuse explicit target selection, frozen-base accounting, compact adaptation ideas, and retention/interference metrics. The newly recorded Layer Card, GAST, FL-Tuning, ALaST and SplitLoRA references are later inputs, not current implementations. None answers whether a 9.5M native, byte-token, agent-oriented Foundation can learn a genuinely new capability while retaining its existing contracts.

The remaining EMMA-specific uncertainty is now narrower: design an acquisition curriculum that is both novel enough to test capability migration and structurally compatible enough for Foundation 001 to generalize, then compare A–D under that fixed protocol. The current deployment-variable curriculum is suitable for pipeline diagnostics but not yet a valid scope-ranking curriculum.

Proposed follow-up

A proposed follow-up is a curriculum-validity control using a held-out task family represented in the Foundation training objective, the same prompt framing, a small full-model diagnostic, and separate acquisition and hidden-generalization sets. This control would precede LoRA, replay, EWC, routing, experts, or RL. Repeating the A–D comparison is contingent on the permissive control demonstrating generalization.

Evidence: acquisition-control JSON and corrected A–D matrix JSON.