Protocol and development documentation

Preserved interpretation accesses a learned neural expert while retaining original instruction-following weights. The four-owner final passes: 949/1,024 newly worded knowledge answers, with all earlier outputs and losses reproduced exactly.

Frozen feature reuse develops repeated training of an added shard without recomputing the immutable distributed prefix each time.

Knowledge rehearsal isolates repeated exposure while retaining the existing generated-answer and retention requirements.

Start here	Purpose
Live LLM checklist	Six fixed completion goals, the active milestone and evidence of completion.
Join the alpha	Pinned source, public ledger observer, starter credits and paid multi-turn chat during the funded window.
Alpha deployment result	Measured latency, automatic recovery, public evidence, funding and remaining limits.
Operated alpha	Atomic funded admission, provider maintenance, automatic recovery and the distinction between AWS hosts and independent operators.
Provider LLM service	Frozen accepted-model concurrency, streaming, recovery and complete-cost trial, including failed setups.
Provider runtime	Native discovery, assigned-partition restoration, authenticated execution and recovery.
Hosted chat	Lightweight client, complete price caps, provisional streaming, settled conversation history and public-data limits.
Finite sponsorship	Complete learning/serving cost scope, replay-quorum assumptions and bounded research funding.
Continual admission results	Three prospective admitted cohorts, automatic continuation, retained answers, an equal-resource growth comparison and complete native replay; checklist items 1 and 2 complete.
Public testnet	Install the client, join, earn test NEURO and request inference.
LLM protocol	Supported training, payments, consensus and model-serving rules.
Model card	Capabilities, limits, evaluation and provenance.
API	Native ledger and inference interfaces.
Data pipeline	Immutable ingestion, collection and verification.
Node deployment	Pinned runtimes, service templates, recovery and monitoring.
Model evolution	Experimental full-model training, growth, verification and reproduction.
Text protocol	Model/tokenizer identity, response windows, document evaluation and text conformance.
Native lifecycle	Curated fresh data, training, quality decisions and paid generation on isolated networks.
Lifecycle evidence	Real-model integration, failed promotion, recovery and continued training.
Compact optimizer disputes	Bounded SGD refutations, complete-audit requirements and measured costs.
Funded auditing	Prepaid complete replay, reporting windows, collateral, payments and refunds.
Funded audit evidence	Two-host execution, recovery, public peer connectivity, costs and failure analysis.
Audit admission RFC	Proposed atomic reservations and refutation funding; not activated.
Candidate operations	Recovering operators and auditors, full-node joining and bootstrap ownership.
Scaling design	Compute groups, audit funding, capacity-backed growth and release gates.
Inherited ledger rules	Reference consensus, accounting, bonds and dispute assumptions.
LLM experiments	Measured outcomes, including failures, with historical evidence links.
Research requirements	Remaining conditions for a stronger public deployment.
Learning milestone	Frozen useful-learning, continual-learning and two-host scaling contract.
Learning result	Completed 128-step run, rejected quality gate, all generations and downloadable checkpoint.
GPU learning reference	Complete-conversation full-model training, held-out evaluation, checkpoint recovery and a bounded GPU trial.
GPU reference results	Completed 1.7B training, exact same-host recovery, inconclusive test gain, all answer pairs and target audit.
Cooperative experiment	Frozen target-quality, two-GPU training and replicated-serving comparisons.
Cooperative results	Measured task learning, shared parameter agreement, communication overhead and inference failover.
Four-worker methods	Frozen local-window training, complete group checkpoints and replica-serving comparisons.
Four-worker results	Reduced communication, exact process-crash recovery, failed quality preservation and a corrected compression feasibility probe.
Shared-gradient learning	Passed task/retention screen with lower communication; buffer equivalence and stronger batching controls.
Batched comparison	Passed full-length quality/retention screen and 1.45× faster training; 38% more GPU seconds, fixed model size.
Persistent model shards	Train and generate through disjoint model partitions with complete optimizer/RNG checkpoints.
Shard recovery results	Physical-host replacement and exact recovery of a 1.7B model across two GPU shards.
Adaptive shard experiment	Portable Adam state, reference regularization, redistribution and gated model growth.
Adaptive shard results	Measured redistribution, complete GPU replay, native settlement and quality decisions.
Native shard replay	Bond-weighted complete replay, sponsor-funded audits and bounded GPU update issuance.
Portable jobs and serving	Native recipe activation, separately audited quality approval and escrow-paid inference across model shards.
Continued learning / result	Three-shard continuation completed; generated-answer gate failed despite lower loss.
Calculation-step learning / result	Large generated-answer gain, failed sorting/filtering retention; all final answer pairs published.
Weight consolidation / result	Prior answers retained and large new-task gain; two sorting regressions still fail the final gate.
Answer-balanced continuation / result	Complete frozen gate passes: new answers 378→462, all 188 correct prior answers retained, three model shards.
Preserved interpretation	Original-model interpretation combined with a learned expert; distributed quality evaluation.
Second expert with interpretation	Separate learning owner, continued earlier serving and exact retention; execution requires the preceding final to pass.
Native expert lifecycle result	560 distinct audited updates, 560 NEURO, separate graph promotion, earned-token inference and complete public ledger replay; checklist item 3 complete.
Ordinary serving diagnostic	Inference-only planned-path screen on ordinary development questions; gold controls separate knowledge, selection, decomposition and assembly. Not a training, promotion or new-final result.

This directory contains technical documentation and pinned dependency profiles. Manuscripts, publication figures and raw experiment dumps are outside the tracked tree. The published paper remains available. Historical evidence is linked to revision 108b4ba3d6c6fb5760ff211b447ee95a67fa9112, preserving access without mixing generated outputs into the current checkout.

eval/data/input.txt is a deliberate exception: the small licensed Tiny Shakespeare corpus is a test and source-compatibility fixture used by the reference implementation. Its path and bytes are retained; its license and digest are recorded. Network genesis/data manifests under networks are also required protocol inputs, not disposable training output.

Example settings and compact reproduction plans live under config. Run python scripts/check_repository.py after staging moves to check tracked-file boundaries and local Markdown links. CI also checks package contents so local archives, website files and manuscripts cannot enter a distribution.

• Compose the learned second expert: exact prompt preservation and actual two-call answers; frozen read-only experiment.

• Native expert checkpoint representation: exact frozen ages, compact references and numerical work identity.