Date: 2026-09-21 Status: Implementation complete; feature qualification remains pending.

Control

The runner uses the promoted Foundation 001.x generalized-symbolic artifact and records its expected SHA-256 ([checksum retained in the private evidence record]). The model architecture now keeps adaptive units outside the Foundation configuration and checkpoint state.

Implemented

• Native AdaptiveWeightUnit with per-target low-rank A/B projections.
• Explicit targets for attention qkv/output and FFN gate/up/down.
• Deterministic manual activation, deactivation, stacking, freezing, and separate artifact checksums.
• Registry methods for create/read/install/promote/validation/deactivation and rollback metadata, extending the existing weight registry.
• Exact state hashing utilities for Foundation and candidate isolation.
• Bounded placement/sequential benchmark runner and compact JSON evidence path.
• Unit tests for additive stacking, lifecycle freezing, and immutable base parameters.

Verification

The complete backend suite passes: 100 passed. The adaptive-unit tests pass independently. The Foundation artifact was not modified by the code changes. The large benchmark runner was started in the project GPU environment but did not complete within the bounded execution window; therefore no qualification result is claimed and no candidate is promoted from that run.

Pending validation

Run the benchmark in bounded stages (placement calibration, then A→D) and persist its generated [private artifact]. The report must then be extended with actual acquisition, hidden generalization, retention, stacking, restart, direct-weight control, and rollback measurements. The feature gate remains NOT QUALIFIED until those measurements meet the directive thresholds.

Fast-run evidence (2026-09-20 machine timestamp)

A bounded fast run is persisted at [retained internal evidence]. It verified the Foundation hash and zero base mutations while testing FFN, attention, and all-final-block placements. All placement screens produced 0% hidden exact accuracy at the exploratory budget; sequential runs also produced 0% hidden accuracy and one non-finite loss. These are recorded rejections, not promotions. The feature therefore remains NOT QUALIFIED. The run confirms adaptive-unit training did not mutate Foundation tensors.

Qualification run update — 2026-09-21

The serious bounded run used 2,048 procedural training exposures per skill, 256 sealed teacher-forced hidden examples, rank 8, and the final-block FFN placement selected from bounded calibration. Calibration measured 100% hidden exact accuracy for FFN (30,720 adaptive parameters), attention (11,520), and all-supported-final-block projections (42,240); FFN was retained as the smallest passing placement.

Skills A, B, C, and D each reached 100% hidden exact accuracy in selected-unit mode. Foundation mutation was zero at every stage and prior promoted-unit hashes remained unchanged. Fresh-process-equivalent loading from four separate unit artifacts reproduced 100% for each selected skill. The qualified Foundation hash remains [checksum retained in the private evidence record].

Simultaneous additive stacking was measured separately and interfered: A 49.6%, B 80.1%, C 70.3%, D 29.7% on the sealed teacher-forced set. This is recorded as selected-unit modularity PASS; unrestricted additive stacking FAIL, reserved for the later routing/composition milestone.

Each promoted unit is a separate 126,389-byte artifact. All four were uploaded to [private artifact archive] and remote size/hash verification succeeded. Rejected/superseded local checkpoints were removed only after a compact deletion manifest was written: 10 files, 192,067,828 bytes reclaimed. The qualified Foundation 001.x remains local as the active control; the older Foundation 001 copy was archived and removed after verification. A rejected A v2 versioning candidate was recorded and deleted while A v1 remained intact.

Decision: FEATURE PASSED for deterministic selected-unit modular adaptive capability qualification. Stacked additive composition remains unresolved and is deferred to dynamic routing research.