Learning to query owned experts

The jointly trained token-level interface reached low training loss but still failed combined held-out questions. The next method preserves each expert's successful question interface and trains the assistant to request the relevant facts. This follows the general learned-call direction of Toolformer, with separately owned neural experts supplying the results. That paper does not establish this implementation's accuracy or decentralization.

PlannerTraining backpropagates through the preserved assistant's final owned partition. Its first two owners execute their frozen layers; the output owner computes response-only cross-entropy and returns hidden-state gradients. Only a separate adapter on the last owner enters Adam. Planning temporarily installs that adapter. Ordinary answering and specialist execution retain their loaded weights. The service binds the complete adapter checkpoint and rejects a changed adapter before generation. Every planning, argument, expert and optional answer composition call remains in the replayable, metered response.

Supervision consists of input questions and resolved arguments, without factual reference answers. The grouped corpus supplies existing single and combined questions, ordinary assistant/structured requests and explicit pronoun examples. Original subject/topic groups remain disjoint between fitting, development and final roles. Initial upstream system instructions are preserved in visible user input because this conversation interface starts with a user; both measured arms receive the same normalized conversation. Preparation templates construct labels only and are not inference-time routing rules.

The numerical check trains through actual process boundaries, restores a saved Adam state and reproduces the next update exactly. It checks that the adapter is used only for planning, that ordinary answering remains unchanged after removal, and that another adapter cannot serve under the old service identity. These are execution checks; useful held-out decomposition and answers require the frozen GPU comparison. This fits an interface to existing experts and does not count as a new admitted expert cohort or a resource-matched growth comparison.

The frozen trial fits 1,152 question-only records for 288 updates, with an exact replay of the final 16 optimizer steps. It measures 56 development conversations before and after training; the 144 final conversations remain unopened unless every development gate passes. Single-expert answers, structured tasks and unchanged ordinary answers each require at least 87.5% accuracy; combined and pronoun requests require at least 75%. Combined accuracy must improve by at least 12.5 percentage points over the untrained-planner baseline, with no category regression. Separate question-plan thresholds also apply. Both arms use the same frozen experts, router and optional neural answer composer. All composer calls are recorded.

The budget is five A10G owners for at most two hours, with a $20 planning cap and automatic termination. The run stops on infrastructure or numerical failure and never adds epochs or selects another checkpoint after scoring. Five focused data, distributed training and conversation checks passed in 28.60 seconds before the prescription was frozen. This is an interface-learning experiment; the six-item live-LLM checklist remains unchanged.

The initial allocation was stopped during bootstrap, before numerical execution: the shared conversation encoder labeled earlier assistant turns as well as the final plan. The corrected prescription masks all preceding turns. A full check of all 1,352 prepared records reproduced the exact inference prefix and decoded only the intended final JSON plan from the supervised tokens. Four focused data checks passed in 0.05 seconds. The retry retains the original records, optimizer schedule, quality gates, combined $20 cap and absolute deadline of 2026-09-16 23:24:03 UTC. Its earlier frozen source and stopped allocation remain recorded.

The corrected run completed. Correct development answers improved from 13/56 to 48/56, but the frozen gates failed and the 144 final cases remained unopened. All 48 original question plans were exact. Eight pronoun plans used valid paraphrases instead of the required literal wording; all eight actual answers were correct. The experts returned both correct values for all 16 mixed requests, and correct values for all 24 single factual requests. The final composer mishandled six mixed responses and one single response. Ordinary answers were unchanged in 8/8 cases; structured answers reached 7/8. The final 16 optimizer updates replayed exactly. Complete checkpoints, outputs, timings and source identities are in the published result. Both allocations are retired; their combined compute upper bound was $2.70, with storage and transfer separate.

The next prescription starts from that retained planner, with fresh Adam and one fixed 144-update epoch. It teaches a bounded output program: the neural planner chooses the questions and either short, semicolon or assistant rendering. Short forms extract actual neural source values using each expert's declared response schema and preserve their order. They cannot supply an answer from labels or a lookup table. Open-ended composition still uses the assistant. Invalid source values, ambiguous JSON and output overruns fail explicitly.

The comparison reuses the complete published 56-case development result from the trained planner. It verifies the original record root and every actual conversation before reuse. Final baseline inference executes afresh only after all development gates pass. Correct-answer thresholds and the prohibition on category regression remain unchanged; mixed accuracy must improve by another 12.5 percentage points. Prospectively, pronoun planning is judged by the actual resolved name and field, together with the selected output operation, instead of exact synonymous question wording. Other question-plan gates remain exact. This change does not turn the previous failed trial into a pass.

This bounded trial allows five A10G hosts for at most one hour and a $10 planning cap. It tests composition of existing learned facts. Three admitted cohorts, broader assistant quality and matched-resource growth remain separate unmet criteria in the fixed checklist.

The answer-plan run completed its development and previously unopened final comparison. Development improved from 48/56 to 53/56 correct answers. Final answers improved from 129/144 to 136/144: mixed requests rose from 23/32 to 29/32, pronoun requests from 15/16 to 16/16, directory remained 32/32, protocol 16/16, structured 29/32 and ordinary-answer retention 14/16. All final answer and category-retention thresholds passed. The complete frozen gate failed: only 25/32 mixed plans were exact, below 28/32. Four further correct replies omitted a terminal full stop in a directory question; three mixed plans were invalid. These observations do not retroactively change the failed gate.

The final 16 optimizer updates replayed exactly. Median final response time was 3.04 seconds versus 2.89 seconds for the baseline; this is a sequential finite workload, not a public load result. Inference bandwidth is unavailable: the driver observed the all-owner transport counter while cached inference used separate branch transports. Its recorded zeros must not be presented as zero traffic. Complete evidence and limitations are in the result. All five GPU hosts, disks and the temporary security group are retired. The compute upper bound was $3.26, with storage and transfer separate.

This establishes an improvement in using existing distributed expert knowledge. It does not establish repeated new-data learning, resource-matched growth or authorization to promote this service on the native chain. Future question-plan criteria should assess resolved requests instead of incidental punctuation, with fresh acceptance data and rules committed before evaluation.

The operator service also accepts an optional planned_tariff file containing integer prompt_atom_price, output_atom_price and the installed tokenizer's context limit. quote_planned reserves the worst case across planning, two argument calls, two answers and optional composition. Generation returns a metering receipt; successful complete replay reproduces that receipt. Actual prompt processing and neural output are charged separately. Planning includes the additional adapter parameters in its owner shares; deterministic rendering adds no neural token charge. Every receipt conserves its integer total and fits the quoted bound. These receipts do not themselves authorize a ledger payment, establish economically sufficient prices or replace funded neural verification. The quote explicitly covers provider execution; verification and retained evidence require separate funded obligations before native acceptance.

Bounded planner work now has an optional native settlement profile. A planner_work genesis entry commits the initial adapter and Adam state, the frozen assistant partitions, token batches, full optimizer schedule and numerical execution profile. Each reserve_planner names the three assistant owners and locks an already accepted native audit budget. A claim_planner contains one to sixteen consecutive updates, every intermediate checkpoint, and receipts from the reserved owners. The existing weighted audit machinery must accept complete replay before issuance; its honest-quorum assumption still applies. Equal execution-role shares sum to the fixed update reward. Numerical work identities exclude job titles and document annotations, while binding the consumed weights, Adam tensors, ordered token targets and effective optimizer parameters. Training never changes serving_root.

The existing local graph operator accepts planner_work, planner_rows and planner_checkpoints configuration paths. These refer to the prescription, tokenized JSONL training rows and content-addressed complete checkpoint files. train_planner requests bind prescription, input_checkpoint and stop; planner_audit carries the native claim. Both use a fresh optimizer restored from the window's input checkpoint, and both persist every produced state. Only the final assistant owner stores the trainable adapter. Missing or corrupt input produces no positive audit report. An auditor's configured backend must submit a fresh execution request and return its report, without reusing another auditor's numerical result. Queue identifiers must be fresh for execution retries; native transaction retries retain the original signed transaction.

The five-process CPU integration trains actual updates, performs three fresh complete replays, and submits their verdicts to the native state machine. It checks exact issuance, preserved serving state, forged and unavailable work, duplicate rejection, persistence and continued-training replay without earlier checkpoint files. This is integration validation using a tiny model and one administrator. The planner profile has not been operated on a new GPU-backed chain, does not migrate an existing genesis, and does not yet provide repeated data admission or a planner quality-promotion transaction. The failed frozen answer-plan result remains ineligible for promotion.