Native expert graph serving
The first full GPU integration attempt stopped during startup; its failure record is public. The corrected run retains the original model, job, numerical profile, native genesis and acceptance rules. Its result is pending.
The runtime fingerprint includes the CUDA libraries actually loaded by the process. The tested AMI's login shell selects a different CuDNN library from a background service with no LD_LIBRARY_PATH. Deployment must preserve the original library selection and pass the fingerprint check before allocating model weights. Installing the Python requirements alone does not establish that the loaded runtime is identical. The original profile remains the acceptance criterion; a launcher mismatch must not be accepted by changing that profile.
The candidate implementation connects the measured parent, preserved interpreter and learned experts to a separate native quality decision and escrowed inference. It supports the published composed graph, including its explicit routing and two-question grammar. General automatic expert selection remains item 1 of the checklist. This implementation has local numerical and ledger checks; it is not yet an operated native deployment of the complete 560-update expert.
Commitments and promotion
serving_graph wraps the original research descriptor with the complete parent and compact expert metadata, interpreter tensor inventory, tokenizer file hashes, EOS/context bounds, exact interpretation prompt and executor commitment. It reconstructs the published graph without rewriting parent or expert optimizer ages. Ownership and parameter counts come from the actual tensor shapes.
The opt-in expert_lifecycle profile freezes the earlier graph, the already validated terminal candidate, and the quality policy before the native replay job. This is integration of an existing validated trajectory, not admission of arbitrary future cohorts. Continuous admission still requires item 2.
claim_expert continues to pay accepted, nonduplicate prescribed updates without changing serving. quality_expert is possible only after the complete expert and its feature production have settled. It requires a separate funded native audit quorum covering the bound quality report. An accepted passing report changes serving_root; an accepted failing report keeps the previous graph and closes that candidate's quality decision. Neither outcome mints additional tokens. Unavailable or forged reports preserve serving and permit a retry of the same frozen candidate and policy.
Raw questions and complete billing
infer_expert commits the accepted graph, raw question, token limit and every participating provider. It locks a price covering the maximum of every neural call selected by the committed execution rules:
| Path | Neural calls charged |
|---|---|
| Ordinary question | Parent generation |
| Directory question | Preserved interpretation, then directory expert generation |
| Protocol question | Protocol expert generation |
| Explicit composed protocol question | Both protocol expert generations |
respond_expert binds actual greedy token sequences and prompt roots for all those calls, the rendered text and every participating owner's signed receipt. The numerical auditors reconstruct prompts, execute the graph, check the actual outputs and reproduce the transcript commitment. A matching signature or a well-formed token list alone does not settle payment.
The charged amount is actual generated tokens, including EOS and interpretation, times the request's unit price. Each call's fee is split by participating parameter count, with deterministic integer remainders. Unused experts receive no inference payment. The payer receives unused escrow; an expired unfulfilled request receives the full escrow. Network transaction fees and separately funded audit fees remain distinct. In-flight requests retain their original graph after promotion.
Questions, outputs and provider receipts are public ledger data in this profile. The current five-owner control group receives the question, including owners unused by that particular neural path; configured auditors receive their replay inputs. This is not a private-prompt service or the final streaming chat API.
Numerical executor and retention
sharded.graph_execution.GraphNetwork loads only each owner's parent/expert and interpreter portions. It reuses the evaluated generation, interpretation and composition kernels, verifies installed source and runtime, and records every actual neural call. Owners agree on the request before selecting process groups and compare complete outputs afterward. An oversized context is rejected without leaving unused owners waiting at request completion. No silent truncation occurs.
sharded.graph_service re-executes inference claims. sharded.graph_quality re-executes the new questions against both baseline and candidate, using the original response-quality gates. It checks retained inputs by comparing the committed deterministic computations: selected models, input construction, tokenizer, interpretation, executor and numerical profile must remain identical. Conversation checks additionally bind the original causal input and target mask. This preserves old errors as well as correct answers. It does not establish hardware equivalence, artifact availability or improved answers on new inputs.
The published retained cohort contains 1,024 knowledge questions, 768 skill questions and 256 conversation records. A metadata preflight found all their computations unchanged in 1.87 seconds. This avoids repeating thousands of unchanged neural calls during every quality audit. Newly routed inputs still require actual generation. This is a new audit method, separately checked from the earlier numerical retention run.
scripts/run_native_expert_service.py is a bounded operator queue for the five owners, not a customer-facing daemon. It supports generation and separate fresh inference/quality audits, sends idle heartbeats, records results, and stops on an unavailable operation. Deployment must impose an external deadline and stage the exact committed inputs. Its initial transport still uses a fixed process group; dynamic provider replacement and coordinator recovery remain item 4.
Evidence boundaries
The development branch also supports prospectively committed jobs. Set expert work to neuroshard-prospective-expert-work-v1, replacing feature_root and batch_roots with batch_count. The prepared records, ordered schedule, initial checkpoint, optimizer recipe and numerical profile remain fixed before execution. The existing executor's --produce-features mode computes and durably reads back the new feature bank. claim_expert_prefix binds its root and every batch in the worker receipts and native replay obligation. Only an accepted prefix populates expert_work.execution_profile(state) for training; a missing or rejected audit leaves training disabled. The producer and auditor share numerical kernels but execute in separate processes from available bytes.
neuroshard-prospective-expert-lifecycle-v1 replaces the known candidate_graph with a candidate_template containing the exact initial expert. A matching neuroshard-prospective-expert-graph-quality-v1 policy commits that template and the evaluation inputs before training. The terminal graph is materialized from the settled checkpoint of the complete prescribed job. Incomplete work cannot be served, and a failed quality decision preserves the accepted graph. Neither prefix acceptance nor quality promotion issues training rewards. These formats remove precomputed-output dependencies; repeated cohort activation and general growth admission remain separate unfinished work. Existing operated profiles continue to use their frozen source and formats.
Validation of this development path includes new prefix production, independent execution replay, forged batch rejection, then training from the retained actual features. The producer discovers the same numerical result without receiving its expected root. Separate lifecycle checks bind the newly derived terminal graph and preserve issuance, refunds and serving through rejected audits. These small-model checks establish execution behavior, not a fresh large-model quality gain or completion of the live-LLM checklist.
The extensible graph descriptor retains the original experts and supports additional separately owned tails. The serving executor derives its process count from the graph; its shared parent partitions participate in every expert path, while each additional owner loads only its tail. The prospective template identifies exactly one untrained expert, so materialization can bind a third or later learner rather than assuming the protocol expert. Explicit fallback rules preserve their order; the separately committed learned service selects among the declared experts using user embeddings. A six-process CPU check executes three learned expert paths, independently replays them, preserves the four earlier serving cases, and excludes unused tails from payments. This extends bounded execution and admission metadata. Automatic composition, fresh large- model quality, and dynamic process replacement still need their own evidence.
Repeated cohorts can opt into neuroshard-expert-admission-v1 at genesis. propose_expert_job commits the next prospective work, graph template and data inventory. Native validator owners vote with a retained review commitment; activation requires the configured delay and both the original and current electorate quorums. Activation preserves balances, issuance, paid-work history and the accepted serving graph. Expired jobs and failed quality decisions leave that graph available. No per-cohort genesis edit is involved.
The inventory binds immutable publisher revisions, source cursors, original conversation objects, token commitments and exact feature-batch membership. Previously admitted evaluation documents cannot become training data. Replay is limited to documents actually consumed in accepted training windows. expert_data.review checks available original bytes against independently read upstream rows, retokenizes complete assistant targets with the serving tokenizer, and screens current training/evaluation near-duplicates. Native votes attest to this review; signatures do not establish semantic truth or exclude all poisoning. The reviewer explicitly leaves semantic curation pending. Historical exact identity exclusion is enforced; historical approximate-duplicate detection is not implemented.
The general conversation format removes the release-question training format's fixed batch and dataset restrictions. Its separate graph quality policy binds the admitted evaluation conversations, baseline, prospective candidate and preapproved scoring/retention rule. Admission cannot substitute easier test records or relax that rule after training. This scorer still measures short answers and pairs of answers; it is not a general chat-quality benchmark. Checks cover three successive admission transitions, rejected data and quality, restart, expiry, original-byte review, actual fresh prefix/training production, independent numerical replay and a six-owner general quality rejection. The state-transition fixtures do not establish three useful learning cohorts. An unattended cohort operator and a prospective large-model campaign remain necessary for checklist item 2.
The real five-process CPU checks execute all four serving paths, replay their outputs, reject fabricated response text and a falsely reported quality pass, and check that old paths remain identical. Ledger checks cover separate quality approval, failed/unavailable audits, pinned in-flight graphs, composed budgets, refunds and supply conservation. The ledger tests seed a post-training boundary; they do not claim that the real 560-update expert has earned native rewards. The focused graph, expert-work and portable-lifecycle suite passed 43 checks in 155.36 seconds, including the bounded operator queue's idle and shutdown paths.
The GPU training backend separately passed complete prefix production and updates 0–8. The next operated integration must verify this serving wrapper on the pinned GPU profile, settle the complete original expert job, admit its audited graph, and pay for an answer using earned NEURO. Quorum honesty and actual operator independence remain explicit assumptions; adding machines under one administrator does not demonstrate independent operators.