Prospective ordinary learning campaign

This campaign tests checklist items 1 and 2 together. It starts from the exposed 15/15 ordinary development service, then requires three new learning cohorts to pass complete automatic answering and preservation before native promotion. The development repair is the initial model, not an additional passing cohort. The checklist remains open until the operated evidence meets its criteria.

The source-grounded admission, conversation and feed cohorts each contain 16 facts, 384 training conversations and 32 prospectively frozen ordinary questions. The final contains single questions and combined questions, including questions that need earlier experts. Twelve general assistant questions and eight multi-turn conversations are fixed preservation anchors. These are bounded assistant tasks, not a claim of frontier-model capability. Evaluation labels and answers reach the scorer only; the answering system receives the conversation.

The initial bootstrap exposed a general-answering failure before any full cohort was trained. Its real prefix and one update were independently replayed and accepted; the measured quality failure was audited and rejected, keeping serving unchanged. The repaired answering system now has the following measured results:

Screen	Knowledge	General skills	Conversations	Previously correct lost
Original native bootstrap baseline	24/25	3/12	4/8	—
First general-interface diagnostic, failed	22/25	6/12	5/8	4
Complete worked-answer development diagnostic	25/25	10/12	8/8	0
Separately frozen prospective assistant screen	—	10/12	6/8	Fresh cases

Both final rows used automatic selection and actual owned shards, and replayed their first complete response exactly. The prospective screen passed its precommitted 75% floor in each category. It still failed four questions: one comparison was routed incorrectly, a conversation calculation was wrong, a sort was wrong, and a word-position answer was wrong. Those failures remain in the record; these results establish bounded assistant competence, not broad ChatGPT-level capability or another learned cohort.

The repair preserves mixed-question decomposition and earlier free-form replies, fits a general/specialist guard using separate training inputs, and uses bounded worked reasoning for explicitly constrained general requests. Its six worked examples contain neither the installed expert facts nor evaluation answers. The final-answer boundary only extracts generated text; it cannot certify its truth. All reasoning tokens are included in complete replay and native billing. A temporary full-model oracle outside the stopped network first matched all 218 preserved tensors and all ten control token streams. Two earlier reasoning methods scored 12/20; an explicit final-answer reminder scored 18/20 before integration. The full-model oracle was deleted before distributed evaluation.

The development outputs and prospective outputs bind the measured complete replies. The renewed prescription retains every unopened admission/conversation/feed training and final byte, domain gate, training recipe and quality threshold. All twenty prospective assistant cases join retention, yielding 25 knowledge, 24 skill and 16 conversation anchors. No full cohort has passed or been promoted under this renewed prescription yet. Checklist items 1 and 2 remain open.

Serving allocation now reserves FP32 weights and working memory without reserving gradients or Adam; training retains its original conservative guard. The parameter and context limits are unchanged. Numerical checks preserve exact owned outputs while checking that inference cannot bypass the parameter limit. The renewed native source commitment includes this allocation correction.

Each full job uses 128 updates, batch size 16, microbatch 4, maximum training length 256, learning rate 0.00005, four warmup updates, weight decay 0.01 and gradient clipping at 1. Only its terminal checkpoint is eligible. A new tail starts from the preserved C expert with fresh Adam state. Earlier experts stay available. Ordinary training paraphrases fit only the appended selection gate. This is isolated addition by default; updating or consolidating remains allowed when the complete service demonstrates preservation and benefit.

Admission requires at least 80% new single answers, 75% combined answers, a strictly positive 95% paired-bootstrap lower bound on single-answer gain, no previously correct retained answer lost, and at least 75% accuracy in each retained knowledge, general-skill and conversation category. Every earlier admitted evaluation joins retention. All numbers and decoding rules are frozen before numerical training. Failure is audited, recorded and stops later jobs under this prescription. It cannot authorize tuning on that final.

The first one-update job exercises rejection using separate frozen questions. An unexpected pass must be published and stops this prescription; the controller cannot manufacture a failure. The first full cohort consumes the remaining fresh source rows and replays the 16 rows actually trained in that first job. Later cohorts replay 16 actually trained rows from the preceding cohort.

Four native CometBFT validators run a new experimental genesis. The existing publisher and curator/auditor daemons consume monotonically advancing immutable public feed heads, review source correspondence and contamination, fund jobs, assign owned computation, audit every window, and submit complete quality for promotion or rejection. The publisher restarts from its durable signing journal between ticks. A separately hashed source window with a substituted answer must be quarantined before funding. Serving probes continue on the accepted graph during training, audits, rejection and promotion. Only future verified numerical windows issue rewards; initial model history cannot mint retrospectively.

Seven disposable GPU hosts own the computation. No host receives the full backbone or preserved interpreter. Three owned prefix stages feed a separate tail owner; each of three auditors executes its own fresh stages and updates. The source-backed curator is installed policy for this reviewed corpus. This does not claim automatic semantic truth verification for arbitrary web data, independent operator ownership, cheap verification or cross-hardware equality.

The first full cohort also has a fixed-capacity control replacing C with the same newly trained weights. Each arm receives seven hosts, the same disk budget and a 5,400-second interval including computation, three full audits, complete quality and two serving replicas. The control recomputes work without submitting it for issuance. A fixed queue of training prompts fills the remaining interval. Report quality, retained losses, throughput, transfer/storage and request overruns; completion of the comparison alone does not establish a benefit from growth. This measures one workload, not lifetime cost or the optimal fixed-capacity method.

Before any full-cohort quality scoring, the observed first windows took about 35 seconds to train but 198 seconds between successive produced boundaries, including publication, fresh audits and native settlement. The original 90-minute comparison estimate was therefore too short for 32 windows plus quality. The recorded resource amendment sets 150 minutes for each identical seven-host arm. It retains the original growth start, including the interruption, and changes only that duration. Source, genesis, data, weights, learning rules, final questions and quality gates remain identical. This is an operational budget correction after training began and before full-cohort scoring; the original 90-minute target must still be reported separately. The original prescription remains preserved beside the amendment.

prepare_ordinary_cohorts.py builds the training-only selectors and scoring inputs. freeze_ordinary_campaign.py binds immutable inputs, policies, source revision and numerical profile. run_ordinary_campaign.py separates allocation, setup, operation and retirement. Actual runtime observations must match the committed profile before initialization. The initial L40S allocation could not place the complete pool. Its partial allocation was retired. The replacement prescription uses seven g5.xlarge A10G instances, at most two concurrent fresh evaluations, an absolute twelve-hour termination deadline and the same $150 planning allowance. Actual compute and storage costs are reported separately. The data, learning recipe, gates and matched comparison interval are unchanged. Allocation and retirement explicitly exclude the three protected network hosts.

Completed full cohort: rejected

The admission cohort completed all 128 updates. Ninety-six fresh training audits matched the 32 four-update windows. All three fresh quality audits matched all 105 evaluated conversations. Singles improved 3→13/16; combined answers improved 2→9/16, below the required 12/16. No previously correct retained answer was lost. Skills stayed 20/24, conversations 14/16, and knowledge improved 25→29/33. The knowledge inventory includes eight questions from the rejected bootstrap; they are evaluation history, not accepted learning.

Native quality rejected the candidate at height 8,978. Serving stayed unchanged. The ledger issued 129 NEURO for 129 verified updates, including the bootstrap's one update. Full application replay reproduced 446 signed transactions and 8,995 headers, including the final state at height 8,994. These validators and auditors remain under one administrator. Replay proves neither independent ownership nor useful learning. The matched-resource control did not run; later cohort finals stayed unopened. At the original 90-minute deadline only 100 full-cohort updates were complete: that original operational target failed.

The result, ordinary replies and ledger replay record the failure. The exposed admission final is development data now. Checklist items 1 and 2 remain incomplete.