Date: 2026-10-01. Status: bounded isolation diagnostic pass; useful synthesis and composition unqualified. No training, new weights, promotion or active-library changes.

Objective

Test the proposed architecture of separate frozen member passes followed by orchestration/synthesis. Determine whether independent activation lanes can share one Foundation or need separate Transformer instances. Exercise explicitly cataloged output forwarding, preserve provenance and existing weights, and distinguish execution interference from weak member capability.

The runbook contains the research, locked initial protocol, test matrix, limitations and precise next objective. It links the canonical broader fabric documents; the master checklist remains the status inventory.

Inputs and protection

• Pinned 9,570,240-parameter Foundation; V18 skill_a_v2 and skill_b_v2, each 30,720 parameters, attached to block 7 FFN gate/up/down projections.
• Both members are unqualified candidates. Standalone V18 transfer was 16/70 and 9/70; V19 did not reinterpret them as qualified.
• Only the already opened V18 development dataset, 70 A and 70 B tasks. No sealed data or composition-reserve scoring.
• Full 145-artifact inventory persisted before inference. Catalog pins, dataset hash and runbook hash saved before measurements. No optimizer or training path used.
• Distinct branch specifications record module/version, purpose, base/artifact/tensor pins, contracts and research-only status. Configured sequential compositions record members, pins, order, purpose and typed bridges.

Foundation artifact SHA-256: [checksum retained in the private evidence record].

Implemented and measured

1. Selective branch instantiation with separate model instances and request copies. Concurrent dispatch preserves output order and source references; same-branch requests serialize. Active branches cannot unload. Failures release active-request accounting.
2. Independent A/B generation lanes in a doubled batch through one immutable Foundation, using native request-local per-row gates and separate token histories.
3. Controls: unchanged Foundation, other member, additive activation in one pass, probability-average synthesis and lowest-entropy per-token synthesis.
4. Cataloged A→B and B→A output forwarding. Bridges validate predicted three-digit outputs and wrap the next prompt, preserving the producer record rather than mutating it.
5. Unload/reload, fresh-process restore, in-memory tensor integrity and full on-disk artifact protection.
6. After the user requested an explicit corruption check, a separate read-only historical audit compared original V18 pins, actual loaded tensors, finite values, lineage and reproduced standalone prediction hashes. Loader hardening now rejects historical artifact/provenance mismatches. This is a post-diagnostic audit, not a claim that the strengthened gate ran before the initial V19 inference.

Results

Isolation: every standalone prediction matched its concurrent independent-branch prediction. Both shared-base activation lanes matched their respective standalone predictions on both task families, including reversed dispatch order. This is prediction equality on the tested outputs, not a claim of bit-identical intermediate logits.

Synthesis: neither simple synthesis rule improved either member. Choosing confidence/entropy is not a substitute for task-aware composition or independent verification. Averaging incompatible task solutions can reduce accuracy despite intact isolated branches.

Two-output collection: 4/70 paired independent A/B outputs were both correct. This metric pairs independently generated skill outputs; it is not evidence of a solved joint reasoning task or concurrent branch-specific Drone input handling. Concurrent isolation was measured separately on replicated inputs.

Sequential forwarding: A→B 0/16, B→A 0/16. All intermediate strings satisfied the three-digit schema, so valid syntax did not establish semantic correctness. A and B are inverse rotations; identity is a trivial control. These chains establish executable forwarding, not learned compositional success. This run did not retain per-stage neural forwarding outputs or measure an oracle-intermediate arm, so it cannot assign every sequential miss to the first versus second member. The next protocol explicitly requires those boundaries.

Persistence: unload/reload identical; fresh process loaded only A and reproduced the recorded first-16 prediction hash. All instantiated branch histories and the additive/shared-base control preserved their in-memory tensor hashes. All 145/145 protected artifacts unchanged, no new weight artifacts.

Numeric/historical audit: both member artifacts and loaded tensors match their original V18 records; all 12 parameter tensors are finite. Training-data and Foundation lineage match. Standalone output digests match V18 exactly. This rules out detectable alteration since those recorded artifacts and numeric NaN/Inf corruption. It does not prove that the training procedure learned the intended capability. V18's missing full persisted pretraining inventory remains a historical limitation.

Tests: 17/17 focused tests passed across V19 and V17/V18 regression suites. They cover concurrent request isolation, declared ordering with noncommuting programmed operations, pin/contract refusal, unload during active execution, failure cleanup, retained-packet protection, and rejection of damaged artifact bytes or changed historical training lineage before branch creation.

Measured diagnostic elapsed: 15.71 seconds, excluding child restore and the later audit/tests. Peak allocated CUDA memory: 237,213,696 bytes, approximately 226.2 MiB across all loaded controls. This is not an isolated topology memory benchmark. GPU kernel overlap and speedup were not measured. Shared-base row gates compute both adapter deltas before masking, so sparse-FLOP savings are not established.

Meaning for the architecture

We do not need duplicate Transformer weight bodies merely to isolate compatible same-base adapters in this tested setup. Separate activation lanes and adapter selection preserved standalone output. Separate model instances are also viable, with extra base residency. Different model families or useful independent specialists remain a separate whole-stack execution question.

Isolation is working; member transfer and useful synthesis are not. The V18 members are intact but weak, and their errors remain when isolated. It would be incorrect to blame all failures on weight corruption, or to conclude that parallel modules cannot work. It would also be incorrect to train orchestration around weak members and call that a qualification of capable modular weights.

Weights require computation. Passing a result to another module means forwarding a declared token/typed/latent representation into its compatible model execution, not running a bare tensor file. Typed forwarding is implemented here. Heterogeneous latent forwarding still needs a trained/versioned bridge and compatibility experiments. Remote artifact storage and remote execution remain distinct untested paths.

Follow-up criteria

Preserve these candidates. Develop newly named members with a materially changed input-access/copy mechanism, qualify standalone transfer, then test nontrivial communication with strong members. The runbook specifies new pointer-selection versus attention-access candidates, fresh data, integrity-first gates, bounded pilot budgets, oracle/actual intermediate comparison, causal ablations, separately owned orchestration/link weights and real cost controls.

Repeating the unchanged V18 training configuration, V19 isolation test, or untrained synthesis rules is not supported by the evidence. Existing trained members remain protected references. Each proposed stack or Module Set requires its own objective, pinned artifacts, communication boundaries, validation, and rollback record.