Date: 2026-10-01. Result: bounded positional specialist and typed communication pass; Foundation-attached attention adapters failed. No active promotion or library insertion. Six new, independently named research artifacts; all prior weights preserved.

Objective and preparation

Find a materially better member mechanism than V18's final-block FFN byte generation, then test communication only after standalone transfer. Research supported comparing input-address selection with new attention adapters. The broader target remains modular frozen endpoints, independently trained links/decisions, parallel Context and potentially communicating whole Transformers; this run isolates two smaller boundaries.

Protocols were saved before their respective fits: V20, changed V20.1 revision. Mechanism references and next architectural stages remain in the canonical research plan. No external model download/import occurred.

Reliability fixes:

• AdaptiveWeightUnit.freeze() disables gradients without labeling a candidate promoted. Registry installation takes its lifecycle status from the actual record; explicit validated promotion remains separate.
• Legacy conditional rewrite generation now covers true and false conditions. Historical artifacts/data were not rewritten; its ten-state support remains limited.
• Tests verify a frozen failed candidate cannot gain promoted status from freezing, and registry promotion refuses an unvalidated candidate.

V20: four new candidates, fixed budgets

Data seed 2001: 512 training worlds, 128 development, 128 fresh IID, 128 novel-symbol and 128 unsupported-length worlds. Every world has rotation-by-supplied-amount and swap-by-supplied-indices tasks. Strings use unique characters, so address errors cannot hide behind repeated characters. Supported lengths3–5; held-out length6 is diagnostic. Train/familiar evaluation symbols ABCDEF12; novel symbols XYZ789. Operations are generally noncommuting. Separate64 communication worlds were locked but not executed after the member gate failed.

Two 12,489-parameter pointer members used frozen per-token Foundation normalized states plus explicit positional/task-parameter features. Learned source addresses were rendered through deterministic copy; labels/oracle outputs were not model inputs. Each trained700 updates on its own parameters.

Two 23,040-parameter rank 8 attention adapters targeted QKV/output projections in the final two Foundation blocks, each trained400 updates with response-only autoregressive loss. They had separate weights and IDs; no existing member was retrained.

Pointer development passed, but swap missed the 95% novel-symbol gate. Communication was not run. Attention development failed; its held-out evaluation was correctly skipped. No training budget was extended. Fresh-process pointer IID restore matched exactly.

Zero-Foundation-state IID ablation: rotate128/128; swap121/128. This suggested the positional policy could largely operate without Foundation latents, while content-dependent features could introduce nuisance variation. It did not prove the precise cause of all novel-symbol misses. The attention pilot's very poor training fit also means its failure is not proof that all attention adapters are unsuitable.

V20.1: one justified mechanism change

The revision removed token-content representations from the learned address selector. Two new members v20_1_position_rotate and v20_1_position_swap each contain8,897 parameters. They learn addressing from typed output position, source position, input length and task parameters. Rendering copies selected input symbols. There is no hardcoded transformation inside either learned member and no Foundation forward pass.

The same512 V20 training worlds were explicitly reused as replay at the same700-update budget. Fresh evaluation seed 2011 excluded1,216 prior states, including V20 communication states/intermediates. Development/IID/novel-symbol splits each have 128 new worlds. Newly generated training worlds were unused. A small character universe limits fresh length3 coverage: development contains 10 length3 worlds; IID contains 6. Longer supported lengths have larger coverage. This is a single-seed bounded experiment, not broad generalization evidence.

Address and byte correctness agree because characters are unique. Both members passed the locked supported-transfer gate. Length generalization failed and remains open. Removing content creates symbol invariance by construction; the learned positional function still has a fixed categorical coordinate limit. Neither a Foundation reasoning claim nor AdaptiveWeight qualification follows.

Frozen member communication

After standalone gates passed, both V20.1 members were frozen.64 fresh length4 worlds were selected where A→B and B→A genuinely differ. Each composition is separately named, pinned, ordered and cataloged in retained traces.

The bridge validates typed state and forwards actual predicted bytes plus operation parameters; it does not compute either operation. Producer/intermediate/consumer outputs, oracle diagnostics, member pins, request hashes, source references and composite specifications are retained. Actual and oracle chains both succeeded, eliminating an observed gap in this tested distribution.

The plan selects the operation/order explicitly. It is a deterministic orchestration control, not a learned Activation/Composition/Contribution policy. Collecting two outputs is not a trained synthesis module. Both branches currently receive the same world payload, with independently scoped operations; separate heterogeneous Drone acquisition, G policy, stale/conflicting messages and multi-round collaboration remain untested.

Integrity and provenance

• V20:145/145 existing artifacts unchanged; exactly four new candidate files.
• V20.1:149/149 prior artifacts unchanged, including all four V20 failures/candidates; exactly two new files.
• Historical V19 inventory also matches V20's pretraining inventory. All original145 binaries remain intact across both runs.
• Foundation source/attention-model tensors remained unchanged. No Foundation gradients from pointer training; attention optimizers/gradients were scoped to the new adapters. New pointer and attention weights are separately labeled by family, operation, version, data hash and tensor/artifact hash.
• Pointer candidates are finite and their tensor hashes remained unchanged through communication. Both runs restored supported pointer outputs in a genuinely fresh process. No active registry alias was modified.
• Final focused verification: 32/32 tests passed, covering lifecycle, registry refusal/promotion, independent weights, dataset/oracle consistency, audit-field exclusion, symbol independence, concurrent isolation, forwarding and existing weight/composition regressions.

Measured core run time: V20 approximately55.13 seconds; V20.1 approximately20.89 seconds, excluding child restoration and later tests/audit. These are total experiment durations, not topology latency/compute benchmarks.

Conclusions

We now have two bounded independently trained modules that generalize to new input identities/symbols, remain frozen, execute independently in parallel, and communicate through actual outputs in both meaningful orders. Sequential mistakes can be localized against oracle-intermediate controls. The successful mechanism is learned addressing plus exact copying, not additive weight merging.

We have not qualified arbitrary-length reasoning, Foundation-attached AdaptiveWeights, neural latent links, learned task-semantic orchestration, a learned synthesizer, multiple communicating whole Transformers, remote execution, app integration, multi-seed robustness or retention across later learning stages. No deterministic-control advantage is demonstrated: explicit operations are straightforward software baselines.

Next move: preserve these members as a working bounded communication substrate. Test separately owned task-semantic orchestration/link candidates on outcomes that distinguish one/many/order and real costs. Treat Foundation-attached adaptation as a separate failed mechanism needing a new representation/optimization diagnosis; do not call the pointer result its repair. Do not repeat these operations just to chase another100% score. Length6 failures require an independently declared relative/variable-length addressing experiment if that capability is pursued.