Native jobs and serving for portable model shards

Status: implemented candidate; a new experimental genesis is required. This does not migrate the public 0.4.0 network or import the research checkpoint as paid work. The balanced learning result supplies the numerical recipe for the integration trial.

Admission and computation

The optional portable_lifecycle manifest enables propose_shard_job and vote_shard_job. A proposal commits the currently settled parent, prepared data, output job identity, numerical executor source commitment, reference checkpoint, terminal cursor, window bound and quality policy. The quality policy commits its serving baseline and evaluator/data/generation rules before any paid update. Activation requires strictly more than two thirds of both snapshotted and current native voting weight, aggregated by owner. It also waits for the native activation delay. A timed-out proposal refunds its bond; ordinary transaction fees remain burned.

Activation changes the prescribed computation, not model tensors or the paid cursor. The first child may use the newly authorized job identity while its parent remains the previously settled checkpoint. Architecture, parameter births, optimizer semantics and shard layout cannot change. Each prepared job can activate once; paid-step identities survive subsequent jobs and failed quality. A live job permits one reservation or execution claim at a time.

Existing reserve_shards, worker receipts, funded audit escrow and native commit/reveal verdicts settle one to four updates per window. The dedicated executor checks the genesis and live reservation before allocating work, loads only its owned parameters, preserves Adam state, and records boundary witnesses. Every honest auditor replays every partition sequentially. Only accepted new updates mint the prescribed NEURO reward. This still replicates computation; it does not reduce verification to a signature check.

Separate quality approval

After every activated update settles, quality_shards submits a report binding the already-settled candidate, frozen serving baseline, prepared data, policy, complete results and pass/fail decision. Its funded audit obligation is separate from training. Auditors reproduce both checkpoints' scores and greedy answers, check the prescribed decision, and attest the complete report. A valid report may say fail: correctness of evaluation and model improvement are distinct.

Only a valid, passing report with a positive native audit quorum changes serving_root. Quality approval mints zero tokens. A failed quality decision keeps the serving checkpoint and already-earned training rewards. It closes that job; the learned checkpoint and paid cursor remain available for a later explicitly admitted recipe. A failed or unavailable claim releases the report slot. Another publisher can report on the same fixed checkpoint and policy; retry cannot select a different model, evaluator or dataset. This prevents an invalid first report from permanently blocking a correct quality decision.

Quality witnesses are segmented at each generation and role-level loss scan. Each segment uses the existing closed-graph verifier and event bound. The outer commitment fixes segment order, complete coverage and the quality statement; auditors must finish all segments. This bounds individual witness groups, not total evaluation cost or total artifact storage.

Complete service manifests are exchanged as size-declared, hash-checked chunks. The underlying control messages retain their 2 MiB limit; each decoded manifest is capped at 128 MiB and their combined allocation at 256 MiB. Segmenting the numerical witness alone does not make its complete metadata a small message.

The initial integration repeats the exact 64-update balanced recipe and its already exposed final evaluation. It must reproduce the accepted research learned-state root. It is an integration/reproducibility test, not another independent learning experiment. No new data, optimization search or model growth is authorized by that trial's policy.

Paid sharded inference

infer_shards locks the caller's maximum price and fixes the serving checkpoint, tokenizer, prompt tokens, greedy stopping rule and one provider per partition. The checkpoint stays fixed if serving is promoted while that request is open. respond_shards requires signed receipts from every assigned provider and complete numerical replay before releasing payment. Providers split the fee by owned parameter count, using the training adapter's deterministic rounding. Payment transfers escrow; it does not mint tokens. Unused token budget returns to the caller. Unanswered requests expire with a full escrow refund.

An inference provider must also arrange a funded audit offer. Its audit cost can exceed the requested inference price. This candidate demonstrates correct accounting and execution, not sustainable market pricing. Prompts and responses are public, the current decoder is bounded greedy generation, and model size does not imply general assistant competence.

Operator entry points and artifact contract

scripts/run_native_shards.py executes training, quality and serving, and replays one partition for an auditor. Its immutable execution descriptor commits the original balanced recipe plus the native wrapper's sources. Training and serving require a pinned genesis, a live assignment, and the existing assigned worker key. The command writes checkpoint/witness commitments and signed receipts; it does not submit payments itself.

The existing scripts/replay_sharded_claim.py backend accepts portable training, quality and inference claims. Its operator-controlled catalog adds an execution JSON path for this profile, retains all rank-specific checkpoint and witness paths, and includes the original baseline for quality replay. The backend checks hashes, executes each partition in a fresh subprocess and retains completed partitions across restart. A missing file or crashed process produces no validity vote; explicit numerical mismatches produce rejection.

The chain holds commitments, not multi-gigabyte model tensors. This trial stages verified artifacts on operated hosts and preserves them separately. A public artifact discovery/replication network and independent providers remain work outside this candidate. Native quorum safety still assumes less than one third Byzantine bonded weight and correct verifier software. One infrastructure owner running several keys does not demonstrate independent ownership.

Acceptance checks

• Changed parent, unsupported recipe, missing admission or reused job cannot train.
• Native window boundaries preserve full-model/Adam results against continuous execution.
• Every shard and communication segment must replay; fabricated tokens cannot reuse an honest witness.
• Training acceptance alone cannot promote serving; quality and inference mint zero.
• Failed quality preserves prior serving and earned rewards; absent audits never pay.
• In-flight inference retains its model version, pays assigned providers once, and refunds unused or expired escrow.
• Every transition preserves supply conservation and global work deduplication.

The existing genesis without portable_lifecycle retains the prior adapter's behavior. Operators must explicitly choose this new manifest and its complete consensus source hash; merging source does not upgrade a running chain.