Learn a second expert while the established model serves

This experiment adds a fifth neural worker to the preserved-interpreter graph. The new worker learns a separately owned two-layer documentation expert. The three parent owners and the existing expert continue answering earlier questions, including the original model's question interpretation.

Status: the first graph passed its final; the second-cohort GPU run is being retried after an isolated controller I/O failure. The earlier branch-cohort proposal remains deferred because its different prerequisite graph failed. This proposal has a new plan and identities; it does not change that failed result.

The plan requires the actual published final for graph 2a81df24652ac4a434ad9ffb3eec09e6038801104645869d33ee35c6a4a2826d. Preparation binds the complete result bytes, verified archive, numerical sources, original interpreter assets, and every earlier answer to retain. A development-only pass, another graph's pass, missing archive verification or an unfinished allocation cannot authorize training.

The first expert keeps its four-owner process group. The second uses a separate group containing the same three parent owners and the fifth worker. The new owner receives only its tail and a frozen output-head replica. After the parent owners produce its 28 fixed training batches, the new owner performs 560 local updates while the established paths serve requests. No participant holds a whole 1.7B model. The graph has 3,691,204,608 stored parameters; this measures added capability with additional capacity, not an equal-budget advantage.

The interpreter, parent and first expert remain fixed, with their real optimizer ages preserved. The new tail starts from the parent weights with fresh Adam state. Its loss trains only the new task. Earlier answers are protected by independent fixed paths and checked directly. An added rule follows existing rules, preserving their precedence. These are explicit domain choices, not a general learned router.

The new training material contains 64 source-anchored facts about the historical NeuroShard 0.4.0 interface. The question specification separates training wording, development wording, final wording and two-fact combinations. Facts and answers are supervision; they never enter inference as a lookup table. Responses use ordinary greedy generation over the unchanged tokenizer vocabulary.

Only terminal update 560 may qualify. It must answer at least 75% of new single-fact questions and 50% of composed questions, with a fact-cluster gain lower bound above 0.10. Every retained answer and conversation loss must remain exact. Both established paths must answer after the first new update and before the last. After the controller observes the new expert process exit, both must answer again. Complete development evidence is archived and selection committed before the independent final opens.

The allocation is bounded to five GPUs, six hours and a $100 planning cap, and starts only after the preceding allocation has been retired. The trial issues no tokens or serving promotion. Native graph settlement, arbitrary peer admission and general assistant quality require their own execution evidence.

First GPU attempt and control-file correction

The first attempt completed 43 new-tail updates before a worker observed an empty control file. The controller used cat > destination, which exposes a file before its full JSON content arrives. The worker's JSON reader failed and its peers then lost their connection. Development quality and the independent final were unopened. The step-zero checkpoint, feature bank, complete source and observations were archived with full readback. All five instances, disks and the study security group were removed; compute was bounded at $1.48.

The corrected controller validates each signal and publishes it with an atomic temporary-file replacement. It reads live service logs only through complete newline-terminated records. A regression check reproduced the original race, verified the actual replacement command, interrupted writes, duplicate and conflicting signals, and partial service records. Its decision checker still accepts the actual five-process CPU evidence and rejects altered answer tokens. The retry retains the exact prepared inputs, neural source, training recipe, unopened final questions, original overall deadline and $100 planning cap.

Every operator must publish complete control JSON atomically. With the pinned repository on PYTHONPATH, use its existing durable helper:

from neuroshard.evolution.reference_data import save
save(home / 'new-learner-started.json', receipt)

The same requirement applies to the later service and process-exit receipts. Only write a receipt after observing the event it describes.

Completed training attempt and failed development

The retry finished all 560 prescribed updates while the parent and first expert served earlier requests. Single-fact development answers improved from 7/64 to 49/64, but paired answers reached only 2/16 and failed the unchanged 50% gate. Retention then stopped at question 113: configuration serialization had reordered the interpreter examples from name, field to field, name, changing the prompt. The first 112 answers matched; the remaining retention and departure checks were not completed. No final was opened.

The 0/280/560 checkpoints and feature bank were preserved with full readbacks, and all five temporary hosts, disks and their security group were retired. The terminal checkpoint is 698d9ae2ba3a89de5b1bf19ec42e6680fe2d408a96f92efa84b383ac5d95b530. See the complete failed result. A separate composition experiment must bind prompt token identity and execute both constituent neural calls; it cannot change this outcome.