Compose the learned second expert

The closed five-owner experiment passed development and its previously unexecuted final. New single answers rose from 7/64 to 57/64, and paired answers from 0/32 to 23/32. All 1,024 earlier knowledge answers, 768 skill answers and 256 conversation losses remained exact. All five owners agreed on answer identity 47a9803d48f4cd2418268f5af69750803169437dc477450f3d2d1a027e9fb09d. The fifth process actually exited successfully; the other four then served the earlier paths. No owner held a complete 1.7B model.

This establishes a bounded second-domain learning result through an added 134M expert and explicit request composition. It uses one operator and known training facts with held-out wording and combinations. Retention preserves previously incorrect answers too. It does not establish general assistant quality, arbitrary routing, independent ownership or an equal-budget scaling advantage. The expert's original 560 updates were retained without retraining; this read-only evaluation issued no tokens and did not activate native serving.

The full evidence archive has a verified storage readback. All five temporary GPUs, disks and their security group were retired. The retirement helper initially required a missing administrative no-replacement record; cleanup was recovered using the actual unchanged allocation, and that failure and recovery are recorded in the result. The numerical result and frozen helpers were unchanged.

See the five-owner reproduction procedure. The public model release contains the new expert, training bank and numerical bundle. All 24 new assets passed complete anonymous SHA-256 readbacks; publication records bind the result, public files and each owner's download requests. The experiment's method and original failure remain documented below.

The fifth peer completed 560 updates while earlier paths served requests. Its single-fact development answers improved from 7/64 to 49/64. Direct generation of two answers reached 2/16 and failed. Retention also caught a prompt serialization bug: saving the configuration reordered the interpreter's example keys, changing the actual model input despite an unchanged semantic configuration identity. The failed trial remains failed; it never opened the second cohort's final.

This experiment restores that trial's terminal checkpoint. It performs no further training. Its five owners retain the same parent, first expert, original neural interpreter and second expert; no owner holds a complete 1.7B model.

Two changes address the observed failures:

1. Serialize interpreter example arguments explicitly as name, then field, with the exact spacing used by the successful first graph. Bind both the rendered messages and actual prefix token IDs into the new serving graph.
2. For an explicit First: … Second: … request asking for two semicolon-separated answers, derive two questions from the raw text and execute the learned expert twice. Join the unchanged neural outputs. Record both generated token sequences, both input questions, and the separately labeled rendered response.

The composer receives no source document, expected answer, topic identifier or evaluation role. It is a bounded typed request primitive, not a general language planner. Single questions and unrecognized request formats execute the existing greedy expert directly. Each constituent call has a 64-token limit; a composed request may therefore use twice the neural generation budget. No comparison at equal inference cost is claimed.

Offline reuse of the already observed single-question development calls first suggested 10/16 correct pairs. The frozen operated development then executed both calls on five separate owners and confirmed that score, with every retention and exit check passing. Its selection was committed before opening the final, which reached 23/32 correct pairs from 64 actual child calls.

The frozen plan preserves the earlier quality gates: at least 75% single-fact accuracy, 50% paired accuracy, paired gain lower bound above 0.10, and zero changes to retained answers or losses. Development has 64 singles and 16 pairs. The previously unexecuted final has 64 newly worded singles and 32 other fact combinations. Facts were training material; wording and final combinations were held out. Previously exposed first-expert questions serve only as retention probes.

Preparation must commit source, input hashes, the existing training identity, checkpoint and exact interpreter prompt tokens. Development evidence and an eligible selection must be archived and committed before final execution. A failed gate closes the trial. Both phases observe the fifth process exit before the earlier expert and parent serve again. Model objects already have complete storage readbacks; all temporary resources retire after evidence preservation.

The CPU integration test performs real training, restores its saved tail into five fresh processes, executes both composed calls, compares them with a complete-model oracle, checks old answers and losses, and observes the fifth process exit. This does not substitute for the frozen GPU quality experiment.

This is operated model research. It adds no native issuance, consensus change, public serving promotion or claim of independent ownership. The original training job identity remains attached to the checkpoint; composing its calls creates no new payable training work.