Continuing an accepted expert
The first continued-expert trial improved some new answers but lost previously correct answers and was rejected. It is part of tasks 1 and 2 in the fixed checklist; neither task is complete.
Continued jobs bind a named accepted expert in seed_expert. Initialization copies its weights and resets Adam. The parent remains immutable. Both distributed feature production and the native prefix referee compute replay references from that accepted expert; a read-only tail replica counts against the final parent's resident parameter limit. Later audited windows need their current complete boundary, rather than the previous cohort's seed files or optimization history.
The new continual quality profile generates old and new answers. Any previously correct retained answer becoming incorrect fails the gate. Its initial assistant anchors and thresholds remain fixed, and admission requires every earlier admitted evaluation conversation in subsequent retention sets. A proposer cannot silently omit a difficult old question, rewrite its reference answer, or put protected retention conversations into fresh training. Mechanical provenance and duplicate checks still require semantic curation; they do not prove that arbitrary data is true or harmless.
The graph executor can hold one alternative expert revision on that expert's owner for paired evaluation. Both sets of weights stay frozen and count toward the resident limit. Comparing a candidate does not promote it or overwrite the accepted serving weights. Native quality funding includes the additional retained question pairs.
Frozen method trial
The prescription continues the accepted 134,225,920-parameter tail of the sharded 1.7B model. Four machines own three parent partitions and the learner. No owner loads the complete backbone. The learner also needs the frozen output head, and reference production needs the explicit tail replica described above.
- First cohort: 16 source-anchored facts about newer NeuroShard protocol behavior. Four training wrappers and separately worded development/final questions are reconstructed from the curation.
- Training: 192 fixed updates, learning rate 0.00005, fresh Adam, and batches of four new plus four replay conversations. Ninety-six replay conversations cover all 64 earlier fact topics and come only from batches used in the completed 560-update expert trajectory. Replay uses accepted-expert KL and margin terms.
- Acceptance: at least 75% single and composed answer accuracy, a positive paired bootstrap lower bound on new single-fact gains, and zero lost correct answers on the 96 earlier evaluation questions. A loss decrease cannot substitute for these checks.
- Selection: only the fixed terminal checkpoint. A failed development gate leaves the new final set unopened. A passing development result requires publication of the terminal decision before the controller releases the final.
- Resources: four temporary GPUs, 200 GiB disks, a 2.5-hour allocation deadline and a $50 planning cap. Preserve evidence and retire the allocation afterward.
This trial measures direct specialist learning on actual distributed neural computation. It does not demonstrate automatic routing, three admitted cohorts, a matched-resource advantage from growth, a running public chat service, or an independent operator. The other two curated cohorts are reserved for subsequent learning; they are not included in the first training job. Numerical work and model-quality approval remain separate obligations before native settlement and serving promotion.
The committed driver was exercised on four CPU processes before GPU allocation: it produced and persisted a new trajectory from accepted weights, generated both sets of development answers, rejected the failed candidate, and left final questions unopened. This preflight is execution evidence, not a quality gain.
Measured result: rejected
The four GPU owners completed all 192 updates under source commit 76aeaf5864be9d757959cc57c2c3cedeaf26d259. Full CI passed for that source.
| Direct expert answers | Before | After | Previously correct answers lost |
|---|---|---|---|
| New single facts | 2/16 | 7/16 | 0 |
| New composed facts | 0/8 | 1/8 | 0 |
| Retained single facts | 57/64 | 55/64 | 4 |
| Retained composed facts | 7/32 | 6/32 | 4 |
The new single-fact paired bootstrap interval was [0.125, 0.5], but both accuracy thresholds and the retention gate failed. Five gains on other retained questions cannot offset the eight lost answers. One lost response changed the number of atomic units per NEURO from 1000000 to 1000000000: this includes substantive forgetting, not only output formatting differences.
The final test stayed unopened. No native job was activated, no NEURO was issued, and the serving model was unchanged. A fresh Python process restored the step-188 input boundary and reproduced all four final updates exactly, including Adam. This establishes execution reproducibility for that window, not useful continual learning or an independently administered audit.
The retained composed baseline is 7/32 on direct generation. The earlier 23/32 serving result used two composed expert calls; these are different execution paths and cannot be compared as the same benchmark. Promotion still requires evaluation through the actual serving graph.
Published result and artifacts include generated development and retention answers, every window's metadata, the replay report, source/input freeze, and public step-188/192 tensor catalogs. All 36 preserved tensor objects passed full public download/hash checks, totaling 3,221,433,520 bytes. The result archive is 4,886,131 bytes with SHA-256 3659b07782602bca462745ec95055f3479d0c741fd1466c791998e4bc2ef273c. The failed trajectory was not settled; intermediate windows retain metadata and can be reconstructed from the frozen inputs rather than claiming complete paid retention of every boundary.
All four temporary instances, disks and the security group were deleted. The conservative compute estimate is $3.16, with storage and transfer additional. Existing services were unchanged.
This result rejects this update prescription. It does not establish that merely adding another expert will fix learning: a preserved old expert still needs correct routing, the new expert must learn its facts, and combined questions must pass through the serving path. Those remain explicit requirements before another learning or growth claim.
Preparing native cohorts
expert_preparation.prepare now reads bounded, consecutive rows from immutable source descriptors and creates the native conversation, provenance, token and batch commitments. It starts from the ledger's cursors, records every selected row, and rejects missing rows, unsupported publishers and repeated conversations. Replay keeps its original source position and token commitment and requires an accepted training-window entry; an imported checkpoint does not fabricate replay history. Preparation can restart with the same snapshot and recover identical content-addressed inputs without advancing a ledger cursor.
The numerical owner must produce the actual zero-update checkpoint of the prepared job. expert_preparation.seal then binds that checkpoint to the candidate graph and frozen quality contract, rereads the upstream data, and runs the native data reviewer before returning a proposal. It refuses changed admission history or serving state. Under the continual quality profile, the previous cohort's admitted evaluation questions automatically join the next retention obligation, including questions from a rejected candidate. The initial anchors and scoring rules remain fixed.
Five preparation/integration checks cover restart, source replacement between preparation and review, stale state, missing data, replay eligibility, accumulated retention and actual CPU prefix production, two updates and fresh-process replay. They prove the prepared proposal drives the numerical executor. They do not measure model-quality improvement. The subsequent operated repeating controller covers admission, funding, execution and rejected quality with small synthetic models. Connecting that loop to passing cumulative LLM learning remains part of task 2.
Isolated learning result
The next trial kept accepted A/B weights and the earlier router intact, trained a new 134,225,920-parameter tail from accepted B, and fitted only an appended routing gate. All 512 sharded updates completed on four GPUs. Evaluation used the actual learned service and explicit two-question calls:
| Answers | Before | After |
|---|---|---|
| New single facts | 2/16 | 11/16 |
| New composed facts | 0/8 | 4/8 |
| Retained single facts | 57/64 | 57/64 |
| Retained composed facts | 23/32 | 23/32 |
All 80 previously correct retained answers survived. Both new-answer accuracy gates failed. The final set remained unopened, with no native activation, issuance or serving promotion. A fresh process given only the step-508 boundary reproduced steps 509–512 exactly, including optimizer state.
Published evidence includes the complete generated answers, source freezes, preserved boundary tensors and full public hash checks. All four instances, disks and the temporary security group were deleted; compute was at most $6.42, storage/transfer additional. The separately frozen replacement control failed during setup and generated no answers; it supplies no measured growth comparison. Its failures are preserved.
Every incorrect new answer selected the intended expert. Failed composed replies contained an incorrect constituent response. The original training variations mostly changed framing around an identical core question. The next intervention therefore adds semantic wording variations for every training topic and converts paired training into the atomic questions the serving parser executes. Targets come exclusively from the original training inventory. This was explicitly a development-informed intervention: it restarted from accepted B, kept the same 512-step recipe and gates, and left the original final unopened until the development gate passed and the terminal decision was published. Its completed result is recorded below; the earlier isolated trial remains failed.
The frozen semantic trial contains 192 atomic training conversations: 12 per topic, interleaved in 24 batches, with targets and provenance checked against all 96 original training records. Actual tokenizer validation confirmed complete response masking and supervised EOS on every row. The run compares appended capacity with replacing B using the very same learned weights; it includes fresh boundary replay, public terminal publication before final evaluation, and automatic resource retirement. The four-host allocation had a three-hour/$50 cap. This comparison matches training work and still does not measure equal lifetime storage and serving costs.
Semantic learning and replacement result
The four GPU owners completed all 512 updates under source commit 9ba3adda2dbcf97abc678d46e3cea073d07f2296; its full CI passed. The development gate passed, after which the controller published the terminal checkpoint and service commitment and opened the fixed final set. No further training occurred.
| Evaluation | Accepted baseline | Added expert | Replacement using identical trained weights |
|---|---|---|---|
| Development, new single answers | 2/16 | 13/16 | 13/16 |
| Development, new composed answers | 0/8 | 6/8 | 6/8 |
| Final, new single answers | 2/16 | 13/16 | 13/16 |
| Final, new composed answers | 0/8 | 5/8 | 5/8 |
| Retained answers, both phases | 80/96 | 80/96 | 44/96 |
The addition lost zero previously correct answers. Replacement lost 38 previously correct answers and gained two others. Thus 44/96 is its total retained set accuracy, not 44 preserved previously correct answers. The terminal weights, training computation and new-answer scores match between the two arms. This is evidence about isolation on this workload; it does not establish that growth beats every alternative use of equal total resources.
The complete final gate failed: composed answers required 6/8 and reached 5/8. Single-answer accuracy and paired gain passed; the paired gain interval was [0.5, 0.875]. The operative retention rule was zero lost previously correct answers, fixed before training. The three incorrect atomic answers confused the whole-job record limit with the batch limit, the batch limit with the microbatch limit, and the training request with the audit request. Each failed composed answer contained one of these incorrect atomic responses.
The published result and raw evidence include complete answer transcripts, the replacement arm, all four owners' metadata archives, frozen-input references and resource retirement. All six primary logical owners and all five replacement owners agreed. Rescoring with the frozen implementation reproduced every answer score, paired interval, gate and execution root. A fresh process restored only the step-508 input boundary and reproduced the final four updates exactly, including optimizer state.
The 36 retained tensor objects contain both step-508 and step-512 boundaries, totaling 3,221,433,520 bytes; their complete public downloads were hash-verified before retirement. Earlier windows retain metadata, not a claim of complete funded historical tensor availability. No native job was activated, no NEURO was issued and no serving graph was promoted. All four temporary instances, disks and the security group were deleted. The compute upper bound was $7.05, with storage and transfer additional.
This is a specialized, explicit two-question experiment under one administrator. It does not complete repeated admitted cohorts, ordinary-question planning, broader assistant quality or a comparison of equal lifetime costs. The final is now exposed and cannot be reused as a fresh final for later method selection.
The next measurement is the inference-only ordinary serving diagnostic. It evaluates the complete planned serving path on ordinary development questions, including gold standalone controls, without training or reopening this final.
The committed result manifest pins the evidence archive. To reproduce the metadata checks, download the original source.tar.gz, semantic-expert-freeze.tar.gz and completed semantic-expert-evidence.tar.gz from that release. Extract the frozen source into result-source/ and the two evidence archives into the working directory. Run verify_results.py, then score_results.py in an environment with the repository's numerical dependencies. These checks score saved real responses; replaying neural work separately requires the published tensors and frozen GPU execution profile.