Live LLM checklist
NeuroShard's goal is one useful, openly available assistant whose learning and serving capacity can grow through contributed model shards, coordinated by its own permissionless blockchain.
This is the fixed six-item completion checklist. 5/6 complete. Tasks 1, 2, 3 and 5 are complete against their bounded demonstration criteria; task 6 is complete for the funded operated alpha. Task 4 remains open. A passing experiment counts as supporting evidence; a top-level box is checked only when all of its completion criteria have a committed implementation and reproducible evidence. Changes to these criteria must be recorded explicitly, before the affected work, rather than changing the target after a result.
Completed: three prospectively admitted cohorts in the accepted model lineage. Conversation reached 15/16 single and 14/16 combined answers; audit reached 15/16 and 14/16; storage reached 16/16 and 14/16. Each preserved every measured previously correct retained answer, with three exact complete quality audits. Ordinary requests select and combine learned shards without evaluator labels; fixed broader skills and conversation retention also pass. Under equal seven-host, 9,000-second budgets, addition preserved all retained answers while replacement lost four; replacement served more requests. The declared growth rule remains quality and preservation within the resource budget.
The three accepted extensions span two research geneses. The latest two ran consecutively and automatically under one unchanged frozen continuation. Rejected trials and development repairs receive no cohort credit. This completes the bounded demonstration criterion; broad assistant capability and independent public operation are not inferred. See complete results, replies and reproducible evidence.
Done when ordinary questions automatically select and combine relevant shards, including questions requiring knowledge from different learned experts; at least three successive admitted learning cohorts pass prospectively frozen response-quality and retention gates. Include a comparison against spending the same total resources without growth, and a declared rule for adding, updating or consolidating capacity. Evaluation must cover broader assistant tasks as well as specialized knowledge, without supplying answer labels to routing. More stored parameters alone do not satisfy this item.
Earlier evidence and failures (historical snapshots)
Evidence so far: two specialized experts, improved held-out answers, exact measured retention, and no complete backbone on any one owner. The admitted graph retains explicit routing and composition. A new jointly trained neural interface improved raw single-fact answers to 13/16, but combined answers reached only 2/16 and failed its gate; general retention passed. See the interface result and the earlier published learning result. Continuing that interface reduced training loss without improving combined answers and failed retention. The next owned planner raised correct development answers from 13/56 to 48/56 while preserving the answering weights. All 16 mixed requests retrieved both correct facts, but the final composer mishandled six responses. The subsequent learned output plan improved previously unopened final answers from 129/144 to 136/144; mixed answers improved from 23/32 to 29/32 without category regression. Its exact-plan gate failed: four correct mixed replies omitted a full stop, and three mixed plans were invalid. The frozen result remains failed. Repeated admitted cohorts and matched-resource growth remain unproved. Continued expert jobs can now start from accepted weights with fresh Adam. Distributed feature production and native replay use that accepted expert as the frozen reference. A new quality profile measures retained answers across changed weights; it no longer requires an updated expert to remain identical. The first continued GPU trial raised new single-fact answers from 2/16 to 7/16, but lost eight previously correct retained answers and failed its gate. Its final set stayed unopened; no serving model was promoted. See the continued-learning result. The next method isolates new expert weights and appends a learned binary routing gate while preserving the accepted router. Six-process numerical validation now executes both earlier and expanded learned services on the same owned shards, without duplicating the backbone. The four-GPU trial completed 512 updates: new single answers improved from 2/16 to 11/16 and composed answers from 0/8 to 4/8, preserving all 80 previously correct retained answers. It failed its accuracy gate; the final stayed unopened. Fresh replay reproduced the final four updates exactly. See the isolated-learning result. A separately frozen replacement control reuses the exact new terminal weights while replacing the earlier expert. Learned route aliases let both routes use that fixed capacity; six-process execution checked actual routing, replay and owner payments. The first GPU comparison failed during setup before generating answers. A separate later semantic training trial completed that comparison. Its development gate passed (13/16 single, 6/8 composed), then the previously unopened final failed (13/16 single, 5/8 composed; 6/8 required). The added expert preserved all 80 previously correct answers. Replacing B with identical trained weights scored 44/96 retained answers: 38 previously correct answers lost and two newly correct. This comparison matches training, not total lifetime costs. No native promotion occurred; three successive admitted cohorts and broader ordinary-question quality remain unproved. The next measurement is an inference-only ordinary serving diagnostic of the complete planned path, with gold standalone controls that separate knowledge, selection, decomposition and assembly. It does not train, open a new final or reuse the exposed semantic final. That screen completed with 1/15 ordinary answers passing: 12 selection failures and two decomposition failures. The ordinary access candidate restores the earlier ordinary-question base and fits the new selector on separate training inputs. Its CPU preflight selects all 21 standalone gold routes (13 distinct questions). The frozen GPU run completed all 15 conversations: strict passes improved from 1 to 7, with zero selection failures, five knowledge failures and three decomposition failures. Forced experts answered 10/13 distinct questions correctly. C's three wrong facts exactly reproduce its earlier development output tokens. All six owner traces agree. CPU recovery repaired a missing failure category without changing pass flags; the new run's replay verdict was not persisted before aggregation failed. All resources were retired. Subsequent request preservation and subject-only reference resolution improve 7→9→10/15 without changing expert weights. The completed engineering screen has zero selection/decomposition/assembly failures, all 21 standalone routes correct, all seven retained cases correct, and exact automatic/forced replay with six identical owner transcripts. The C-only learning repair then improves 10→14→15/15 through crossed training contracts and source-grounded claim/job scope contrasts. All 13 distinct forced controls and all 16 C development questions pass; no previously correct tested answer is lost. Six complete serving transcripts agree, and fresh training/inference replay passes. This repairs the exposed ordinary diagnostic. It does not count as a new admitted cohort or independent holdout; repeated prospective learning and operated native admission remain the next proof.
The first full ordinary native cohort completed 128 verified updates: single answers improved 3→13/16 and combined answers 2→9/16, preserving every previously correct retained answer. It failed the 12/16 combined-answer gate and was rejected. A subsequent inference-only literal-question diagnostic reached 11/16 combined but regressed two diagnostic answers; it also failed. Selection and wording-sensitive knowledge remain unresolved. Neither result counts toward the required three admitted cohorts.
Semantic question access now repairs the opened complete-service diagnostic: 15/16 single, 13/16 combined, zero retained answers lost and exact replay. A small frozen semantic encoder selects a canonical training question; the existing expert generates its answer from weights. The question index contains no answers. This is a development repair, with no new final, neural expert training or native promotion. Three fresh admitted cohorts and the matched-resource comparison remain required.
The prospective semantic campaign completed 128 verified updates, improving singles 1→13/16 and combined answers 0→11/16 with zero retained-correct losses. Three fresh quality audits agreed; native quality rejected the candidate. Subsequent selector repairs use that opened set as development only. No fresh cohort was admitted. The complete repaired service reaches 15/16 single and 13/16 combined answers on those opened cases, with no regressions, no retained-correct losses and exact replay. It learns question distinctions inside the retrieved expert while preserving the earlier fallback. This is a development pass; three newly admitted cohorts and the resource comparison are still required.
Completed: the real sharded LLM automatically rejected its bootstrap and a substituted immutable source, then prepared, funded, trained, audited and promoted both fresh cohorts without a genesis or recipe change between jobs. Native cursors consumed 768 distinct training documents, with replay from actually trained windows. The durable publisher resumed across 3,040 process starts; earlier controller/native-process restart and host-restoration evidence remains linked in the completed result. There were 226 during-work serving probes. Full replay reproduced 28,912 headers, 869 signed transactions and exactly 257 issued trial NEURO; paid inference from the final graph settled without additional issuance.
Scope: automatic admission from a prospectively approved, source-backed feed, with explicit provenance, duplicate/contamination checks and source-substitution rejection. Arbitrary-web truth, independent curation and public service reliability are not established by this bounded operated demonstration.
Done when the network repeatedly admits immutable data with provenance, deduplication and contamination/poisoning checks; funds and assigns a bounded training job; verifies the work; evaluates the candidate on precommitted acceptance rules; and promotes or rejects it without manually editing genesis or orchestrating each cohort. Demonstrate restart, rejected data and a rejected candidate while the accepted serving model remains available.
Earlier evidence and failures (historical snapshots)
Evidence so far: immutable ingestion and experimental lifecycle components exist. The growing expert architecture still uses fixed prepared jobs. Bounded planner windows now connect actual replay to the existing native reservation, funded audit and issuance machinery. Five-process CPU validation covers three fresh audits, exact rewards and unchanged serving state. This optional research profile still lacks automatic repeated planner admission and separate planner promotion; it does not complete this item. The continual expert quality rule also accumulates every prior admitted evaluation conversation while keeping its acceptance thresholds and initial assistant anchors fixed. Admission rejects omitted or rewritten history. Native preparation now turns consecutive pinned source rows into reviewable jobs, preserves exact accepted replay provenance, carries prior evaluations forward and refuses stale state. Its inputs drove actual CPU updates and fresh-process replay. Production source integration and passing cumulative GPU learning results are still required. A durable publisher controller now reconciles repeated native jobs, funds execution, submits measured quality failures for audit, and proceeds after rejection. Recovery checks reproduce committed-before-acknowledgement crashes without duplicate transactions. Configured curator reviews share the auditor's sole signer and journal their own approvals or rejections. These recovery checks use explicit numerical fixtures. A further two-cohort integration now executes real prefix production, training and three neural replays per claim, including six-process quality evaluation. It issued only the four verified updates, rejected both quality failures, advanced source cursors automatically and preserved the serving graph without changing genesis. The operated repetition now adds publisher restart, actual repeated- source rejection and nine unchanged responses from persistent accepted shards. Four CometBFT validators agreed, and full replay reproduced all 74 transactions across 449 headers. These small synthetic models and local source fixtures do not establish operated continuous LLM learning. See the implementation and published evidence. Structured immutable feeds now preserve original messages for that native preparation path. Restart-safe publication, cross-window reads and independent source review passed; all nine objects for a three-window, 192-conversation feed passed full public S3/CDN hash checks. This supplies the transport and preparation bridge; operated GPU integration and useful promotion remain open.
The full ordinary cohort now exercises real immutable data, GPU shards, 96 training audits, three complete quality audits and native rejection while accepted serving remains available. Full ledger replay matched 446 signed transactions and 8,995 headers; 129 verified updates issued 129 NEURO including bootstrap. No useful cohort was promoted. Repeated useful automatic admission remains open; the rejection machinery alone does not complete this task.
The prospective semantic campaign now reproduces bootstrap rejection with three exact audits, refuses a hash-consistent substituted source answer, and automatically admits the next real GPU job. A controlled restart after twelve verified updates preserves checkpoint and issuance; accepted shards return identical responses while the controller and research validator processes are down. An earlier startup outage remains recorded. Useful repeated promotion is still pending.
Done when the already validated second expert completes this entire path on an operated NeuroShard-native candidate chain: available numerical inputs and checkpoints; real configured execution audits; accepted training claims; exactly-once NEURO rewards; a separate quality decision admitting the serving graph; and a raw-question inference request paid from earned NEURO. Bind the actual model, optimizer ages, tokenizer, routing and composed-call billing. Publish the ledger and execution evidence. Restart/replay, a forged claim, missing data and a repeated claim must leave supply and serving state correct.
Completed 2026-09-16: actual prefix production and all 140 four-update windows settled, issuing exactly 560 NEURO. A separate full-audit quality decision promoted the original validated graph. A raw composed request used earned NEURO, paid both neural calls and refunded unused escrow. Forged work, withheld bytes, duplicates, restart and recovery onto replacement hosts were checked. Full application replay verified 1,728 signed transactions, 16,359 headers and all four saved validator states. Numerical and ledger evidence, replay instructions and the public 2,856-tensor catalog are published. See the result and its limits.
This was a finite candidate operated by one administrator. Its GPU resources are retired after preservation; independent operation and unattended public chat remain covered by the other items. The integration reproduces the previously published expert's quality gate and does not count as another successful learning cohort.
Current development: operated alpha admission and recovery replaces unaccepted audit offers with atomic work reservation against funded standing services, charges occupied capacity, and adds provider maintenance and native recovery from advertised spares. Its committed CPU native preflight passes both automatic owner replacements, nine complete numerical replays and exact replay of 620 blocks and 143 transactions. The operated LLM alpha now passes its ordinary and automatic process-recovery gates and is available for a funded public window. The independent availability soak remains open. AWS machines under our account do not satisfy the independent-administration criterion below.
Done when independent providers can discover work, acquire assigned shards, join, leave and replace unavailable providers through the protocol. Demonstrate required-backbone-shard loss, coordinator loss, state recovery and continued service for the growing graph across separate failure domains and supported hardware/network profiles. Complete a declared load/availability soak with independently administered operators; no one operator may control one third or more of consensus voting power in the demonstrated release deployment.
Evidence so far: operated sharding, earlier host-replacement experiments and continued older serving after the newest expert exited. The latest five-owner experiment used one operator and controlled cloud networking.
In development: native provider registration, graph/rank offers, collateral and capacity reservation, complete prepaid audit assignment and timed coordinator replacement now have adversarial state-transition coverage. Receipts bind assignment epochs; failed or expired service refunds its unused budgets. Five separately keyed CPU processes now execute actual model partitions through certificate-pinned HTTPS, matching the fixed-group executor for ordinary and multi-turn fixtures. Providers discover their own native assignments, restore committed partitions without SSH and jointly sign the complete response. Operated LLM recovery has since passed below; the availability soak and independent administration remain required. See the provider-market design.
The subsequent native provider preflight passes actual registration, partition restoration, signed execution and settlement across five provider processes and four CometBFT validators. A killed coordinator and a killed backbone owner are each replaced by a fresh key and endpoint. Each recovered request pays once after three complete numerical replays. Application replay matches 660 headers, 84 transactions and all four stored states. This uses a small synthetic graph on one host; accepted-LLM load/cost evidence and independent operation remain required.
The accepted LLM trial now passes fresh-coordinator and fresh-backbone replacement across seven hosts in five availability zones. Requests recover and settle once in 622.34 and 641.97 seconds after three complete numerical replays each. Six ordinary requests also meet the frozen latency gates. All disposable resources are retired, and complete application replay matches all seven saved states. These are process-loss trials with replacement on surviving hosts, not a declared independent-operator availability soak. Provider selection and audit-offer saturation resistance remain open engineering work. Separate EC2 hosts do not establish separate administrators.
Done when every accepted obligation has available evidence, complete funded verification and a bounded liability; adversarial tests cover forged work, duplicate work, withholding and failed audit obligations. Publish the quorum and collusion assumptions and measure complete costs, including verification, retention, storage and serving. Configured fees or finite sponsorship must cover those measured obligations within the declared issuance limits. Signatures and token issuance alone are not evidence of independent computation or demand.
Evidence so far: supply invariants, work deduplication, funded native replay quorums and real replay measurements. Task 3 published the complete retained window payload catalog and settled all 140 windows with full audits. Complete retention/storage/serving costs and finite funding are now reconciled below. The planned-service operator now quotes and meters all prompt and output work, including planning and composition, with exact owner shares and a bounded reservation. Real five-process replay reproduced the same receipt and rejected a forged response. These execution quotes explicitly require separate verification and retention funding; they are not ledger payments.
The finite sponsorship now declares complete allocation, controller, network and retention accounting under a $1,000 bounded research budget. Actual hosted requests prepay complete audits and settle all unused execution/provider/audit funds without inference issuance. The public complete-cost report reproduces $792.54 against $1,000, including three learning allocations, all seven provider attempts, both controllers and the entire public object pool's 93-day retention. Existing checks cover forged/duplicate work, withholding, missing auditors, rejection payments and exact refunds; quorum/collusion assumptions are published. This meets this item's finite-sponsorship option. It does not establish independent auditors, cheap verification at arbitrary model size, a token price, outside demand or perpetual funding. The demonstrated admission saturation attack remains open under permissionless hosting.
The public alpha now serves the accepted growing graph with two concurrent customers, streaming, bounded complete prices and pinned graph/tokenizer conversation history. All ten frozen deployment requests pass, including automatic coordinator/backbone process recovery. Six ordinary cases show first text in 33.15–42.17 seconds and complete generation in 34.75–49.34 seconds. A fresh public observer and wallet completed two paid turns, preserved context and resumed a deliberate timeout with the same nonce. All four saved ledger states replay exactly, with zero inference issuance. The first memory-instruction reply was poor and remains published; the following turn retained the phrase. This is a service/context check, not broad answer quality.
Results, replies, failures and reproducible evidence are public. Admission is scheduled through September 22, 2026 at 11:09 UTC, GPUs through September 22, 2026 at 12:39 UTC and the separate ledger through September 26, 2026 at 23:49 UTC; funding/request limits may close admission sooner. The service is a source CLI, with public conversations and 64-token outputs, under one administrator. These are a finite alpha's measured guarantees, not perpetual availability, independent ownership or ChatGPT-level capability. The existing done-when criteria below are unchanged.
Done when the accepted growing graph serves ordinary multi-turn requests with streaming, versioned context/tokenizer handling, bounded prices, correct expiry/refunds and concurrent clients. Meet a published, prospectively frozen latency/load target on supported deployments. State and test prompt visibility, storage and privacy behavior; users must understand which providers or auditors receive their data. Publish representative responses and failure behavior.
Earlier chat evidence before the operated alpha
Evidence so far: the public 0.4.0 client supports a smaller experimental model; the larger expert graph has operated generation and public reproducible weights. The new candidate binds and audits all neural calls used by a raw question, including interpretation and composition. Those pieces are not yet the live growing chat service. An experimental local service now pins gate and owned adapter checkpoints, preserves multi-turn context, and delivers only checked token chunks. Five CPU processes checked discarded delivery, context overflow and changed owner weights. Eight GPU responses then passed full checking, including corrected numerical disagreements. One-pass prompt prefill and sixteen-token target blocks subsequently passed the frozen latency and traffic bounds on all eight exposed GPU workloads: 128-token responses took 8.37–9.56 seconds and full replay took 1.63–1.68 seconds. The weights still fail their learning-quality gate. Public access, native billing and load targets remain open. See block streaming.
Work discipline
All six milestones remain in scope. Items 1–3 and 5 are complete against their fixed demonstration criteria, with item 5 using its finite-sponsorship option. Item 6 is complete for the bounded operated alpha. Reliable permissionless hosting remains item 4; all existing completion criteria are unchanged.
Before an expensive run, record its exact decision, reused artifacts, cheapest adequate preflight, success/failure rule and time/spending cap. A rerun needs an identified cause and a concrete change. Do not repeat long quality evaluation when the relevant model, inputs, numerical path and acceptance decision are unchanged. Real execution required by a new audit path remains a separate claim that must be checked.
Every update records delivered behavior, evidence, remaining blockers and resources consumed. Failed attempts stay visible. A release meeting all six items is the live-LLM milestone; future quality and capacity improvements continue under the same admission rules.
Progress log
2026-09-20: Completed item 6 for the funded operated alpha. The accepted graph passes ten deployment requests, ordinary concurrent latency gates and both automatic process-loss recoveries. A new public observer/customer passes two paid conversation turns, actual streaming and timeout/resume without a duplicate payment. Application replay matches 3,369 headers, 697 accepted transactions and four saved states; serving issues no tokens. Evidence and access are published with the finite service window and $800 ceiling. Checklist 5/6 complete; independent hosting and its soak remain item 4.
2026-09-19: Completed item 5 under finite sponsorship; advanced items 4 and 6. Native providers acquire only assigned model portions and serve the accepted graph through paid, streamed, replay-verified requests. The final seven-GPU trial passes all ordinary latency and coordinator/backbone recovery gates; all ten responses preserve the preceding source's neural results exactly. All experimental resources are retired. Complete costs are $792.54 under the frozen $1,000 cap, including failed attempts and full retention. Independent administration, market admission resistance, the declared availability soak and persistent public deployment remain open. Checklist 4/6 complete.
2026-09-19: Completed items 1 and 2. The accepted conversation → audit → storage lineage now has three prospective admissions across two research geneses, with zero measured retained-correct losses. The latest two cohorts ran automatically, passed all three complete numerical audits, and promoted natively. Equal-budget growth preserved answers that replacement lost. Full replay passes 28,912 headers and 869 signed transactions; the final graph served paid inference. The public evidence and complete measured scope are in the completion report. Seven experiment GPUs and their volumes are retired; compute upper estimate $59.59, with storage/transfer separate. Checklist 3/6 complete.
2026-09-17: The complete ordinary interface now passes 25/25 retained knowledge, 10/12 general skills and 8/8 conversations, with no earlier correct answer lost and exact replay across actual owned shards. A separately frozen prospective assistant screen then passes 10/12 skills and 6/8 conversations, including exact replay; its four failed answers remain recorded. The twenty fresh cases join cumulative retention unchanged. The renewed three-cohort native prescription preserves all original neural training/final bytes, domain gates and thresholds. This is a measured answering improvement; repeated useful native cohorts remain unproved and items 1 and 2 stay open. Results and limits.
2026-09-17: Completed the C scope continuation: 15/15 ordinary answers, 13/13 unique forced controls, and 16/16 C preservation questions. All earlier correct tested answers survived, with identical neural outputs for all seven retained A/B/general conversations. The 64 updates took 575.6 seconds. Four training owners and six serving transcripts agree; fresh replay of updates 60–64 and both inference replays pass. A preceding setup attempt failed before training because publication receipts lacked the fetch destination; the corrected handoff passes local regression checks. Evidence and limits retain both attempts. Both allocations and their volumes are deleted; compute estimates are $1.68 for the passing run and $0.67 for setup failure, with storage/transfer separate. No new final, cohort, issuance or native promotion; tasks 1 and 2 remain open.
2026-09-17: Training C on crossed input contracts improved ordinary answers 10/15 → 14/15, forced answers 10/13 → 12/13, and standalone C retention 13/16 → 15/16, losing no previously correct answer in either inventory. Four training owners and six serving transcripts agree; fresh replay of updates 104–108 and both inference replays pass. The full gate still fails on one claim-window versus whole-job scope confusion. The 108-update run took 985 seconds; all four instances and volumes retired, compute at most $2.28 excluding storage/transfer. Result and next bounded scope repair preserve the exposed-development status: no new final, cohort, promotion or issuance. Tasks 1 and 2 remain open.
2026-09-17: Repaired ordinary access using verbatim standalone requests and bounded neural subject resolution with deterministic substitution. The first revision scored 9/15 and failed one reference; the second scored 10/15 with only the five existing knowledge failures remaining. Both retain all earlier correct cases, agree across six owners, replay exactly and meter every neural call. Across the same 28 ordinary/control requests, neural calls fall from 69 to 48 and prompt tokens from 19,207 to 7,866. All ten allocated instances, their volumes and security groups were deleted; combined compute estimate $2.03, storage and transfer separate. Result manifests bind the local evidence archives. No expert weights changed and no new final, issuance or promotion occurred. Tasks 1 and 2 remain open.
2026-09-17: Completed one frozen five-host ordinary-access allocation, with retained neural weights, a restored ordinary router, a C gate fitted on separate training inputs, and forced-expert controls. Recovered all 15 saved transcripts after an aggregate-scorer failure; strict passes improved 1→7, while remaining knowledge/planning failures keep access closed. Forty-four focused checks pass. All GPU instances, volumes and the temporary security group were deleted; compute estimate $1.16, other charges separate. No next learning cohort, native promotion or issuance was performed.
2026-09-17: Froze an inference-only ordinary serving diagnostic. Fifteen development questions cover retained A/B knowledge and planner-cohort development facts, with gold standalone controls and no explicit two-question grammar. CPU tests classify knowledge, selection, decomposition and assembly without neural execution. The GPU driver remains unlaunched: five owners, one hour, $25 cap, no training, no new final. Tasks 1 and 2 remain open.
2026-09-17: Published the completed semantic learning and replacement comparison. All 512 GPU updates completed; a fresh process reproduced the final four updates exactly. Six primary owners and five replacement owners agreed; rescoring with the frozen implementation reproduced every recorded score and gate. Development passed, but final composed accuracy failed. Preserved all 80 previously correct answers with the addition; replacement lost 38 and gained two. Published all result metadata and the two retained numerical boundaries. All four GPUs, disks and the temporary security group were deleted; compute upper bound $7.05, storage/transfer separate. Tasks 1 and 2 remain open; the exposed final cannot serve as a fresh final again.
2026-09-17: Completed and published the continued-expert trial. All 192 sharded updates executed; the last four replayed exactly in a fresh process. New answers improved, but accuracy and retention failed. The final test remained unopened, with no native activation, issuance or serving change. Preserved 3.22 GB of final/input-boundary tensors with full public hash checks. All four GPUs and temporary resources were retired; compute estimate $3.16, storage/transfer separate. Added native cohort preparation and sealing with actual numerical execution/replay checks. Tasks 1 and 2 remain open.
2026-09-17: Prepared a fixed continued-expert learning trial using 16 fresh protocol facts, 96 actually trained replay conversations, accepted-expert distillation and 192 updates. Four-process execution preflight and eight preparation/data-review checks passed. The final questions stay closed after a failed development gate. This records the experiment contract, not a new GPU quality result or another completed milestone.
2026-09-16: Removed three blockers to repeated expert learning: restarting every job from the original parent, distilling old answers from that same parent, and using unchanged weights as the only retention criterion. Accepted expert initialization, actual generated-answer retention and complete prior evaluation coverage now have implementation checks. Seventeen numerical execution checks passed, including four-process production matching native prefix replay byte for byte; eleven data-review and graph-execution checks passed, including comparison of two expert revisions without replacing the serving weights. Thirteen admission/checkpoint regression checks also passed. This is implementation evidence, not a new model-quality gain. No GPU was allocated for these changes and neither active milestone is marked complete.
2026-09-15: Established the six fixed items from the agreed live-LLM gaps. All remain open; task 3 is active. Prior expert learning, retention, replay and public artifact evidence are credited above. No new GPU run was needed to establish this checklist.
2026-09-15: Task 3 implementation now has actual bounded training execution and recoverable window-boundary payloads. Thirteen focused execution and native settlement regression checks passed in 37.55 seconds, including a fresh process restoring Adam and completing the next window, and rejection of corrupted state, storage failure and an expired deadline. The 11 original numerical kernel files remain unchanged. No GPU was launched. Native settlement of this expert, prefix-backend integration, quality promotion and paid graph inference are still required; task 3 remains open.
2026-09-16: Full CI passed for the resumable training backend. Added complete native prefix execution with timing-independent production commitments and retained feature bytes. Sixteen focused checks passed in 56.26 seconds. The next GPU probe is frozen around the existing parent and updates 0–8, with one GPU, a two-hour limit and a $15 planning cap. All six milestones remain in scope; dependencies determine execution order and none is complete.
2026-09-16: The real-job preflight caught an identifier-domain mismatch missed by the small earlier-cohort fixture. Stopped the driver before neural execution, fixed preservation of the original interpreted-cohort job, and verified all three real claim contexts. Twelve executor checks passed in 37.58 seconds. The replacement probe retains the same model, data, success criteria, GPU and spending deadline; the first attempt remains recorded.
2026-09-16: The bounded GPU backend probe passed in 854.73 seconds after setup. Complete prefix production, updates 0–8, forged measurements and unavailable input handling met the frozen criteria. Retained and hash-read-back artifacts total 5.98 GB. The GPU and attached resources were retired; estimated compute was $0.65 plus storage. No NEURO was issued and no graph was activated.
2026-09-16: Added the candidate graph lifecycle, complete neural-call billing, real five-owner serving/audit execution, and a quality auditor that measures new answers while checking unchanged retained computations. The latter checked all 2,048 original retained records in 1.87 seconds without another GPU run. Native full-job settlement and GPU serving integration remain open. See implementation and evidence boundaries.
2026-09-16: Froze the complete native integration in native-expert-live.json. All 140 original window contexts and four published serving traces pass the metadata preflight. The fresh four-validator genesis boots with zero issuance. The bounded run assigns actual training to the expert owner and replay to three auditors, retains every accepted boundary before pruning local copies, and checks separate quality promotion and inference paid from earned NEURO. Allowance: five GPUs, 28 vCPUs, six hours, $75 planning cap, automatic shutdown. Numerical success and completion of task 3 are not yet claimed.
2026-09-16: The complete integration stopped during serving startup, before generation or training. The runtime guard caught a different CuDNN library selected by the background launcher. After correcting that path, the smaller owner exhausted the remaining startup interval during cold file reads. All five GPUs and volumes were retired; estimated compute was $1.85. The recorded failure remains visible. A same-target retry carries the original library path and releases clean unowned cache before serving. Its deadline and combined $75 cap include the first attempt.
2026-09-16 02:54 UTC: The corrected five-GPU serving probe reproduced four published responses and all six neural calls across three availability zones. Actual prefix production received the configured three-auditor replay quorum. The first eight updates earned exactly eight candidate-chain NEURO. Forged measurements earned zero; a withheld auditor checkpoint blocked the next payment; restoration resumed execution. Duplicate rejection and restart kept the same checkpoint and eight-token issuance. The full 560-update run continues toward separate quality approval and paid raw inference under the original deadline. See the dated progress evidence. This is one administrator, a specialized graph and partial lifecycle evidence; all six top-level items remain open.
2026-09-16: Completed item 3 and published its numerical and signed-ledger evidence. All 560 distinct updates earned exactly 560 NEURO; the separately admitted graph answered using earned-token escrow. Full replay checked 1,728 transactions and 16,359 headers against four validator states. All five recovery GPUs, their volumes and their temporary security group are retired. Three integration attempts used a conservative $31.96 compute estimate; storage and previous learning runs are separate. The other five items remain open. Development also adds prospective repeated-cohort admission, original- data/tokenizer review and execution by three learned expert owners. These mechanics do not establish three useful learning cohorts or independent operators. A bounded ordinary-question decomposition screen passed 10/12 development cases; two lost subject context and require correction before the method is used to route expert calls.
2026-09-16: Added a separately committed conversation executor that separates neural question decomposition from learned shard selection, executes the selected owned models with request-local KV caches, and records every neural call for replay. Seven cache checks passed, followed by a five-process failed- planner/replay check after integration. These cover tiny CPU models, not GPU numerical equivalence or assistant quality. The typed planner screen failed 11/16 cases; its complete responses remain committed. The frozen cached conversation trial now tests the actual model partitions, mixed answers and complete replay. No additional top-level milestone is complete.
2026-09-16: Added a rebuildable historical-data index that checks new cohorts against all admitted documents, reuses verified fingerprints across restart, rejects near copies of old evaluation examples and preserves explicit replay. Five source-data checks passed, including actual prefix production and independent numerical replay. Added verified object restoration with replica fallback; four HTTP fault checks passed. The first cached conversation GPU attempt stopped during one artifact download, before neural execution; all resources are retired. The retry retains every neural input and pass rule, and changes only artifact delivery.
2026-09-16: The cached-conversation retry retrieved all assigned tensors but exposed a startup race: small experts opened their subgroup before a parent finished cold weight reads. A new readiness barrier loads both local models before forming inference groups. Five graph checks passed, including a deliberately delayed parent and exact cached replay. Both failed GPU allocations are fully retired. The next run keeps the original deadline, questions, weights and scoring rules under a new executor commitment. A public serving catalog now provides CDN/GitHub replicas for all 472 required tensors.
2026-09-16: The five-owner cached conversation trial completed and failed its quality gate (26/32 plans, 1/8 complete answers). All measured cached tokens matched and tensor traffic fell 93.23%; complete response replay passed. Eight diagnostic pairs reproduced the failures without caching and recovered learned facts when the training prompt contract was restored. The earlier router screen replaced domain phrases instead of removing them. Development now binds expert input contracts, supports a bounded planner reminder and freezes answer-blind raw-question router fitting on the existing entity-disjoint split. These changes require the next regression result; no milestone is checked. Complete failed responses and diagnosis.
2026-09-16: Corrected input-only routing passed the frozen development gate: 161/164 raw protocol questions, 1,920/1,920 raw directory questions, every original route and all 146 retained general questions. Calibration used only fitting inputs; class-balanced integer learning countered paraphrase imbalance. The CPU fit and evaluation took 114.21 seconds. The next five-owner regression combines this router with committed expert prompt contracts and a bounded planner reminder, under the original GPU deadline and combined $25 cap. These are development routing results; broad response learning remains open.
2026-09-16: Added explicit learned-route mappings for all four installed numerical paths. General conversation and the earlier trained structured-data model are distinct choices; an assistant fallback cannot silently replace that learned skill path. Eight focused graph checks passed in 61.71 seconds, including actual five-process execution of both backbones. A four-class router fit is frozen on the same data split. The separate three-class conversation regression's first allocation stopped at an Ubuntu package lock before neural execution; its $0.76 estimated compute and complete retirement are recorded.
2026-09-16: Four-model routing passed on the unchanged split: 163/164 raw protocol questions, 1,920/1,920 raw directory questions, all 91 structured questions and all 55 general conversations. The fit/evaluation took 163.89 CPU seconds. Complete request data and earlier turns now reach the selected general or structured model; a single-request path preserves its raw input rather than substituting the planner's shorter rewrite. The actual five-process context-preservation check passed in 21.90 seconds. GPU conversation evidence for this extended path is still pending.
2026-09-16: Published all five owners' conversation and review outputs. The corrected three-route service reached 6/8 answers and 28/32 plans. Input-scope checks recovered general science answers; the four-model service reached 7/9 complete answers with exact replay. Two remaining planner errors prevent promotion. A review pass worsened 28/32 drafts to 1/32 and is excluded from the serving path. All five GPUs, volumes and the temporary security group were retired; this allocation cost at most $2.34 compute, storage separate. See the complete results.
2026-09-16: Implemented an experimental trainable connection between owned model activations. It generates one token stream through both backbones and every expert tail, sharing the trained prefix once. It starts as the unchanged general model and accepts projected source activations instead of rewritten subquestions. Seven numerical checks passed in 18.16 seconds: actual five-owner generation, exact initial seed tokens, reproducible continuation, two-owner gradients matching joint execution, causal masking and request-local caches. These are CPU numerical checks; this method has not yet passed an LLM learning or quality gate and has no native admission. Task 1 remains open.
2026-09-17: The automatic numerical loop completed two real training cohorts against four operated CometBFT validators: four updates earned four trial NEURO, both measured quality failures were rejected, and serving stayed pinned. A driver waiting for the original audit deadline initially missed the shortened reveal window; that failed attempt is retained. Auditor recovery now follows current deadlines and may retire an unknown signed audit only when committed history proves its exact context closed. Five new recovery checks passed. The extended numerical integration also passed with publisher restart, actual duplicate-source rejection and nine unchanged responses from persistent accepted shard processes while cohorts executed (128.51 seconds). The operated repetition passed in 406.98 seconds. Four validator app hashes agreed and a fresh replay matched all 74 transactions, 449 headers and the exact final state. Source and evidence are public. These are synthetic models and one administrator; useful cumulative LLM learning and automatic production data operation remain open, so neither task 1 nor task 2 is checked.