Date: 2026-09-20 (Europe/Helsinki) Protocol: component-adaptation-matrix-v1

Scope and control

This report executes the archived Post-Audit Execution Directive after the targeted prior-art gate. EMMA Native Foundation 001 remained the immutable control. Its local checkpoint is 9,508,800 parameters, 8 transformer blocks, hidden size 320, 8 attention heads / 2 KV heads, FFN size 960, 1,024-token context, and SHA-256 [checksum retained in the private evidence record]. No candidate tensor was written to the Foundation checkpoint; the runner persisted compact evidence only.

An immutability guard prevents overwriting or pruning the qualified Foundation checkpoint. Its focused test passed. Each candidate was an in-memory isolated fork with all parameters frozen except the selected scope. Before/after state comparison verified that only selected parameters changed and that no unrelated parameter changed.

Prior-art gate outcome

The targeted gate is recorded in prior-art-gate-component-adaptation-2026-09-20.md. It allows reuse of established mechanisms and metrics, but does not substitute for Foundation 001 evidence.

• Reusable: explicit target-module selection and freezing (Hugging Face PEFT), compact adapter lifecycle ideas (LoRA/PEFT), and retention/interference measurement concepts (EWC, GEM, Avalanche).
• Minimal reproduction later: low-rank adapters, adapter-based continual learning, replay/stability controls, and selective layer freezing beyond the A–D scope.
• EMMA-specific: the smallest useful unit for agent-oriented capability, immutable candidate promotion, exact parameter-integrity proof, hidden generalization, fresh-process persistence, artifact cost, and teacher-dependence reduction.

The relevant prior-art predictions were that selective updates reduce compute/storage and that very small scopes can underfit. On Foundation 001, the isolation and low-resource predictions transferred: all four scopes trained quickly, used at most 125.5 MiB peak GPU allocation, and changed only intended parameters. The capability-acquisition prediction did not transfer under this first curriculum and budget: every candidate failed exact acquisition. This result is not evidence that one scope is universally inferior; the target curriculum/evaluation needs a controlled revision before scope ranking.

Environment verification

The project GPU environment was used for the run: Python 3.11.9, PyTorch 2.7.1+cu128, CUDA 12.8, CUDA available, and huggingface_hub 0.36.2. Backend verification completed with 97 passed, 2 warnings. Frontend npm run typecheck and npm run build completed successfully; all 18 application routes generated.

Progressive-funnel results

The first pilot ([private artifact]) used a deployment-command target and was rejected at the acquisition screen for all scopes. It is retained as negative evidence. The controlled screening run used a related deployment-variable target, 8 training examples, 120 steps, AdamW, and the same Foundation/validation setup for every scope. All candidates were rejected before expensive full regression and restart testing because acquisition did not exceed the Foundation baseline (0/8 exact target cases for every candidate); retention screens also had 0/12 exact passes under this target/evaluation, so the funnel correctly stopped them early.

All candidates reduced training loss substantially, but loss reduction did not produce exact generated responses. Because no candidate passed the exploratory funnel, restart persistence and full regression were not run, as required by the early-stop policy. No candidate was promoted, and no candidate artifact was created.