Continued-learning result: generated answers did not improve

The fixed 1.7B continuation completed all 96 updates on 14 September 2026 and failed its frozen quality gate. Lower response loss did not produce better answers. This candidate is not promoted, earns no research-run NEURO and does not authorize settlement implementation.

Result

The frozen contract required at least eight net new correct answers on the 256-case primary set, exact one-sided McNemar p < 0.05, no per-family accuracy loss on either task set, and a conversation retention upper bound of at most +0.02 nats.

Measurement	Parent	Candidate	Decision
Correct primary answers, 256 cases	192	191	Fail: zero wins, one loss
Correct prior-task answers, 128 cases	96	95	Fail: zero wins, one loss
Primary response loss	0.144141	0.136006	Measurement only
Prior-task response loss	0.145954	0.138349	Measurement only
Conversation retention loss	0.542989	0.543930	Pass: change UCB +0.001764 nats

Primary lookup and filtering stayed at 64/64 each. Sorting fell from 64/64 to 63/64. Invoice totals stayed at 0/64. On the prior-task set, invoice totals fell from 2/32 to 1/32; the other family counts were unchanged. The primary paired test has p = 1.0. These are complete final sets, not selected examples.

All three development retention checks passed: mean changes were +0.002260, +0.001841 and +0.002201 nats at steps 160, 192 and 224. Those checks neither established generated-answer gain nor overrode the final failure.

Executed computation

Three g5.2xlarge A10G workers jointly trained one model with boundaries [0, 6, 15, 24]. Each held its own student partition, Adam state and frozen reference partition. Owned parameter counts were 503,343,104 / 604,016,640 / 604,016,640; peak allocated GPU memory was 12.65 / 12.97 / 12.97 GiB. No worker held the complete training state.

The training phase took 1,991.00 seconds, including its checkpoint and retention checks. Parent and candidate final evaluations took 365.48 and 357.45 seconds. Every worker completed successfully and produced the same common checkpoint and evaluation measurements. These are cooperating workers under one operator, not independently owned replications.

The implementation repairs made before execution enforce trained replay membership, exclusion by rendered content, the committed freeze on resume, the actual final checkpoint lock and the development abort. Validation passed 412 tests, repository checks and wheel/sdist content checks. Both GitHub CI runs passed for the final selection commit.

Frozen artifacts

• Plan and numerical source reference: af16131.
• Prepared inputs committed at d8cf1d9, identity 1599f1ad67e37fb04e2908b32a544529a76d017d9d0f39c2969a9dc8e8ac9251.
• Actual endpoint selected at 887bb059e21d197311705cbdfee75ccc19245bc9, before final evaluation started.
• Parent checkpoint: 094138fb6e3e2a8962f8455b0bf81de3f2fbe82029222df76aa71e2e22e14d49.
• Candidate checkpoint: 0af8dd9e03917a44257cb76282515bcf6ee67082bb7817b77cb0eead3785f523.
• Candidate learned-state root: b1fa7229b23708a8a8e72c08f02fd26be611cae0768858085e8bfe78af04dcd6.

Each of the three new checkpoints, including its complete Adam state, was uploaded directly from its owners to S3 and verified by full readback. The pre-cleanup evidence archive has SHA-256 d7e064e765a88e0f7d96c538c8cea1aafd21721331bee9a3f12f32516da20536. It contains prepared data, logs, evaluations, numerical sources and checkpoint backup receipts. Large tensor objects are stored separately. This is operator-controlled storage, not a demonstration of public artifact availability.

Reproduce the frozen run from its source commits; later numerical changes must not be substituted for those bytes. Keep the plan in selection-committed status for verification of the recorded final reports. This document records the failed outcome without changing the original contract.

Resource cleanup

All three study GPUs, their root disks and their security group were confirmed deleted at 13:57:58 UTC. Both existing CPU network hosts remained running. Compute through confirmed termination is estimated at $3.77; a conservative planning total including a transfer allowance and $10 for other charges is $20.08, below the $100 cap. These are estimates, not an AWS invoice; retained S3 objects incur ongoing storage charges.

The final evidence archive, including cleanup confirmation, has SHA-256 90a2202424a641eace8f0a91019e224a1523c1cf80aeb5db8ed397c1ea49265e. The machine-readable quality decision is also published in the repository.

Implication

The distributed computation works for this numerical profile. This learning recipe does not solve useful continual learning. Its final examples are now exposed and cannot serve as fresh finals for another recipe. Any subsequent method needs its own frozen inputs and acceptance decision. No serving root or public 0.4.0 state changed.