Date: 2026-10-01 Run: isolated-member-composition-v26 Result: bounded overall isolated-composition pass; no active promotion. Novel strict-comparison accuracy remains 93.75%.

Objective

Qualify a complementary independently owned capability, then run it alongside the already-trained V25 equality attachment without modifying either. V26 trains ASCII ordering, freezes it, and synthesizes equality and ordering outputs into LESS/EQUAL/GREATER. The composition phase trains nothing.

The architecture used is:

Implementation code and detailed machine records are retained separately. This public edition presents the study methods, aggregate results, and qualification boundaries.

One immutable Foundation is shared by two request-local passes. No delta addition, no member merging, no common modified activation stream and no extra Transformer are required for this tested arrangement. Host-concurrent execution is not a claim of simultaneous GPU kernels or a speedup.

Research and locked choices

AdapterFusion distinguishes member acquisition from later composition. AdapterHub's composition documentation provides parallel/stacked mechanism references and method-specific constraints. Mixture of Agents studies later synthesis from prior model outputs. V26 uses these as structural references; its deterministic typed synthesizer is not AdapterFusion or a learned MoA aggregator. No outside weights/dependencies/code were imported.

Runbook locked before fitting. Both new ordering seeds must pass 95% development, and the frozen equality member must pass 95% development, before new sealed scoring. Both ordering seeds must pass 95% sealed before composition. The composition member is predeclared seed 2601, not selected using sealed accuracy. Composition and operand-swap gates each 95%. No post-sealed fitting, threshold relaxation or sweep.

Ownership, data and training

Protected Foundation and original head are frozen. Equality uses immutable V25 seed 2502 artifact SHA-256 [checksum retained in the private evidence record]. Each new ordering attachment has 28,800 parameters: rank 10 deltas at blocks 0/3 attention QKV/output. Seeds2601/2602 have separate initializations, minibatch sampling and artifacts. Only the new ordering unit's eight parameter tensors enter its optimizer. Preflight proves zero initial effect, gradients reach the attachment and protected parameters stay gradient-free.

The reader extracts requested first-character operands with original full-value/source coordinates; it never computes equality or order. Ordering asks for Y iff left is strictly earlier in ASCII, otherwiseN. Gold comparison labels are generator/verifier data, not neural input or synthesis input.

All 110 unordered symbol pairs used by V25 data are excluded from V26 strict-pair selection. Fresh remaining pairs:240 training,48 development,48 sealed,16stress. Training contains 512 unique canonical inputs:480 strict directed comparisons plus32 equal-symbol inputs. Development and sealed each 192 source examples but128 unique canonical inputs; each has 48 LESS,48 GREATER and96 EQUAL cases. Stress64 source cases,53 unique canonical inputs, five fields and ten-character values. All 32 equal canonical states recur from training in development/sealed; strict pairs are held out. Repeated equality states are not new semantic transfer.

Each seed's micro-fit uses the first 48 training examples for200 updates and scores 48/48. Full initialization is fresh, AdamW 0.001,no decay,batch 32,clip 1,max1200 updates. Checkpoints200/400/800/1200 use development accuracy and a two-checkpoint plateau rule. Both seeds completed1200 updates. Source/canonical interfaces remain fixed throughout. Current optimizer/best states were saved in two separately labeled resume-only checkpoints; these are not inference modules.

Standalone results

Unchanged parent scores 0/192 on the new task. Nearest-neighbor episodic memory scores 144/192 development (75%). Frozen V25 equality scores 192/192 fresh development; it was not retrained.

Development on novel strict cases is 91/96 (94.79%) for seed 2601 and88/96 (91.67%) for seed 2602, while recurring equality cases are 96/96. Both pass the predeclared overall development gate, but these subtotals show its limitation. Sealed strict ordering improves to95/96 and94/96, with recurring equal cases96/96. The locked overall qualification passed; do not describe development novel transfer as95% for both seeds.

Composition results

Both qualified members are frozen and mounted separately. Every pass activates exactly one module through request-local metadata. Packets carry capability, module ID/version/artifact SHA, request identity, evidence hash, selected module and output. The synthesizer consumes only those packets. It does not receive source bytes or gold labels.

• Both passes followed by synthesis: 186/192 (96.875%).
• Equality-only safe synthesis: 96/192 (50%).
• Ordering-only safe synthesis: 48/192 (25%).
• The best possible single-bit guessing policy under this balanced distribution has a75% ceiling; the safe single-member paths abstain on ambiguous cases.
• A direct deterministic comparator is 100% by construction. There is no learned-necessity or superiority claim.
• Sequential separate member passes and two concurrent host requests yield exactly identical packets/predictions.
• Swapping actual requested source operands: 186/192correct changed targets.
• Changing only an irrelevant source value: 186/192, predictions identical to original.

Novel strict-comparison synthesis is 90/96 (93.75%). Recurring equal cases are 96/96. LESS scores 46/48; GREATER44/48. The overall96.875% should not hide the weaker cases, especially GREATER, which needs both predicates to disambiguate it from equality.

On sealed data the frozen equality member scores 187/192 and ordering191/192. All six synthesis misses are attributable to those member predictions: five false equalityY outputs and one false orderingY output. Three wrong equalities become incorrect EQUAL answers, two conflicting equality/order packets become UNRESOLVED, and the ordering error becomes LESS instead of GREATER. There is no observed new error introduced by scheduling or the correct synthesis truth table. This does not prove isolation eliminates all kinds of interference.

Preservation and restore

All 168 pre-existing binaries—including every V25 weight—remain byte-identical. Base tensors and both active member tensor hashes match before/after composition and causal interventions. Four new inference artifacts and two resume-only checkpoints make174 tracked weight files. A read-only audit preserved all 174.

Inactive ordering requests preserve159/160 original curriculum results and match physical unmounting. Forced-active scores 144/160 and130/160 demonstrate interference with unrelated tasks; these attachments remain operation-scoped. The old Foundation was not corrupted.

Fresh-process restore passed all four inference artifacts with identical48 diagnostic or192 development predictions. The restored V25/V26 composition also reproduced the complete192 sealed predictions. Focused tests: 29 passed, including packet truth table, mismatched evidence rejection, source-span/label separation, interventions, response alignment, activation and native attachment tests.

Conclusions

This is a bounded demonstration of two independently trained native attachments running in separate passes, preserving their weights, and communicating through compatible outputs. It supports the proposed isolated-pass approach as an alternative to blind stacking for compatible modules. It does not prove hidden-state fusion, stacked block correctness, learned synthesis, learned activation/orchestration, sparse loading, remote execution, arbitrary external weight compatibility, whole-Transformer collaboration or a complete agent.

Neither new ordering member nor the composition was promoted into the active runtime/library. Their validated research status is recorded separately in the single master checklist. The exact deterministic comparator remains a stronger minimal solution for this toy task, and novel strict-comparison synthesis scored 90/96 (93.75%), below the 95% threshold. The overall sealed score of 186/192 (96.875%) also includes 96/96 recurring equality cases.