Preserved interpretation: four-owner final passes
The distributed composition answered 949/1,024 newly worded knowledge questions correctly (92.68%), retained every token of all 768 earlier skill answers and reproduced all 256 earlier conversation losses exactly. The parent continued serving after the controller observed the expert process exit with code zero. All eight frozen final checks passed.
| Measurement | Development | Independent final |
|---|---|---|
| Knowledge answers correct | 982/1,024 | 949/1,024 |
| Knowledge accuracy | 95.90% | 92.68% |
| Entity-cluster gain lower bound | 0.95020 | 0.91309 |
| Earlier skill outputs reproduced exactly | 192/192 | 768/768 |
| Earlier correct answers lost | 0/151 | 0/650 |
| Earlier conversation losses reproduced exactly | 128/128 | 256/256 |
| Parent serves after observed expert exit | Pass | Pass |
The complete result contains the decisions, all-owner agreement, process-exit observations, archival readback and resource accounting. The parent answered none of these synthetic knowledge questions correctly before using the expert. The final entity-cluster gain interval was [0.91309, 0.94043], over 128 entities. Facts were training material; newly committed question wording was held out.
All four GPU instances, their disks and the study security group were removed; the protected hosts remained unchanged. Compute was bounded at $9.80, with $0.16 estimated disk cost and an $11.23 conservative transfer allowance. These are planning estimates rather than an AWS invoice. The complete archive was read back successfully with SHA-256 59679091f8a3c15afbb697d7e78afabeb785fda0579ea6eea7301345f43abe36.
What made the difference
The learned expert already stored much of the knowledge, but did not answer varied wording reliably. The preserved original 1.7B model now interprets a question into a name and requested field. Strictly validated arguments produce a canonical question that the learned expert answers through ordinary greedy generation. Invalid interpretation falls back to the original question. The interpreter receives no factual values, and serving uses no factual lookup table. Questions outside the explicit directory domain use the prior parent.
Three machines each own disjoint portions of the trained parent and preserved interpreter. A fourth owns the 134,225,920-parameter learned expert. No owner holds either complete 1.7B model. Total stored parameters are 3,556,978,688; this is additional capacity, with additional inference work and storage.
All weights stayed unchanged during composition and evaluation. The parent retains its actual Adam state at update 448; the expert retains its separate 1,024 tail updates. The preceding training result records how those expert weights were learned. The method and frozen plan describe the interpretation rule, runtime, ownership limits and final gate.
Selection and reproduction
All 32 outputs of the earlier diagnostic reproduced exactly across the four owners. Development then evaluated the already exposed branch-final questions. The full development archive was read back and selection was committed as 0f6334462733d6fd11e6471cfeeb2bdf818a391e before the new final started. Numerical source was frozen at 7029fb4e455d9babc86405b50668d83066f9981d. All four owners agreed on the final output identity. No tuning followed exposure of the new final.
The research model release provides the exact model and optimizer parts, tokenizer, numerical inputs, all four owners' output records, notices and reproduction instructions. Its manifest lets each owner download only its assigned portions and verify their full SHA-256 hashes. Reproducing these published questions checks execution; it is not another independent quality final.
What this establishes
This is a passed, bounded example of accessing learned neural knowledge through a larger distributed model while preserving earlier outputs. It supports the next experiment: another owner learns a separate expert while this established graph keeps its state and serves earlier requests.
The experiment used one operator and a fixed explicit domain rule. Earlier skills are exposed retention probes. General assistant quality, automatic routing, independent ownership, arbitrary membership and economical permissionless verification remain unmeasured here. The serving graph was not activated by native consensus; this composition issued no tokens. Earlier failed experiments remain failed.